Enterprise AI

ENTERPRISE
AI in the Cloud
ENTERPRISE AI IN THE CLOUD
A Practical Guide to Deploying

End-to-End Machine Learning and
ChatGPT™ Solutions
Rabi Jay
Copyright © 2024 by John Wiley & Sons, Inc. All rights reserved.
Published by John Wiley & Sons, Inc., Hoboken, New Jersey.
Published simultaneously in Canada and the United Kingdom.
ISBNs: 9781394213054 (Hardback), 9781394213078 (epdf), 9781394213061 (epub)
No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means,
electronic, mechanical, photocopying, recording, scanning, or otherwise, except as permitted under Section 107 or 108 of
the 1976 United States Copyright Act, without either the prior written permission of the Publisher, or authorization through
payment of the appropriate per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923,
(978) 750-8400, fax (978) 750-4470, or on the web at www.copyright.com. Requests to the Publisher for permission
should be addressed to the Permissions Department, John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, (201)
748-6011, fax (201) 748-6008, or online at www.wiley.com/go/permission.
Trademarks: Wiley and the Wiley logo are trademarks or registered trademarks of John Wiley & Sons, Inc. and/or its
affiliates in the United States and other countries and may not be used without written permission. All other trademarks
are the property of their respective owners. John Wiley & Sons, Inc. is not associated with any product or vendor mentioned
in this book.
Limit of Liability/Disclaimer of Warranty: While the publisher and author have used their best efforts in preparing this
book, they make no representations or warranties with respect to the accuracy or completeness of the contents of this book
and specifically disclaim any implied warranties of merchantability or fitness for a particular purpose. No warranty may be
created or extended by sales representatives or written sales materials. The advice and strategies contained herein may not
be suitable for your situation. You should consult with a professional where appropriate. Further, readers should be aware
that websites listed in this work may have changed or disappeared between when this work was written and when it is read.
Neither the publisher nor author shall be liable for any loss of profit or any other commercial damages, including but not
limited to special, incidental, consequential, or other damages.
For general information on our other products and services or for technical support, please contact our Customer Care
Department within the United States at (800) 762-2974, outside the United States at (317) 572-3993 or fax (317) 572-4002.
If you believe you’ve found a mistake in this book, please bring it to our attention by emailing our reader support team at
wileysupport@wiley.com with the subject line “Possible Book Errata Submission.”
Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may not be available in
electronic formats. For more information about Wiley products, visit our web site at www.wiley.com.
Library of Congress Control Number: 2023950107
Cover image: © CSA Images/Getty Images

Cover design: Wiley
I dedicate this book in memory of my dad, Jayachandar,
who is my eternal hero and has always been my
inspiration during good and tough times.
I also dedicate this to my mom, Leela Jayachandar, who is

the epitome of kindness.
Finally, I dedicate this to you, dear reader, for picking up

this book and bringing positive change and innovation
into your companies.
ACKNOWLEDGMENTS
THERE HAVE BEEN so many people who helped make this book a reality. Without them you would not be read-
ing this today.
I would like to thank my wife, Suji, my son Rohan, my friends Dominic, Sundi, Karthik, J Siva, Ramkumar,
Aravind, Rajesh, Kumaran, Stanley Mammen, for his tips, CM for helping me promote the book, and many oth-
ers for their ongoing support and for putting up with me through it all.
I had the opportunity to learn from many of the top minds in the enterprise AI space, as well as some brilliant
authors, during the development of this book. I hope some of that has been translated into the knowledge I have
imparted here in this book.
Special thanks to everyone who helped edit this book. Thanks to Kenyon Brown, acquisition editor, who helped
me throughout the writing process and made it a pleasant experience.
Thanks to Kezia Endsley for her tireless editing assistance, for always being quick and positive, and for keeping
me on target through it all. She was very supportive of my efforts, and her kind approach helped me when the
going was tough. Her feedback and input on every chapter of the book helped make it ten times better than it
would have otherwise been. I would also like to thank Evelyn Wellborn for catching many errors as part of the
proofreading process that has certainly elevated the quality of this book.
Thanks to Navin Vijayakumar, managing editor, and Pete Gaughan, senior managing editor, for their insight into
planning the content and marketing tips. Also many thanks to Magesh Elangovan and Vijayprabhakar Settu for
their help with proofreading and images respectively.
A huge thanks to my friend and technical editor, Irwin Castelino, for reading the first drafts and making his
technical edits. His critical evaluation and suggestions helped me make this a valuable product from the reader’s
perspective.
Without these folks and more, this book simply would not have happened.
Thank you,
—Rabi Jay
ABOUT THE AUTHOR
RABI JAY IS a seasoned digital transformation specialist with more than 15 years steering a ship through the
uncharted waters of industries such as retail, aerospace, and software technology. Armed with a plethora of certi-
fications, including AWS Machine Learning, Azure, ITIL, and SAP, he’s your go-to guy for anything cloud and AI.
In various roles, from digital platform lead to VP of architecture, he’s led transformative projects in Martech, AI,
and business process optimization. But Rabi Jay isn’t just about executing projects; he’s about shaping the very
landscape of digital transformation. As a global alliance manager for major cloud corporations, he’s had a bird’s-
eye view of tech upheavals across diverse industries.
The author of SAP NetWeaver Portal Technology, published by McGraw-Hill, Rabi Jay is also the voice behind
the LinkedIn newsletter and podcast Enterprise AI Transformation, where he dishes out cutting-edge insights. In
his role as VP of digital transformation, he has championed the convergence of human-centric design with state-
of-the-art AI platforms and change management methodologies.
Why does all this matter? Because Rabi Jay is not just an expert; he’s a visionary. Passionate about guiding
companies through the evolving maze of AI and digital transformation, he’s got his finger on the pulse of what’s
next. And he’s bundled all this expertise into this latest endeavor: Enterprise AI in the Cloud: A Practical Guide to
Deploying End-to-End Machine Learning and ChatGPT Solutions.
He is excited about what the future holds for us and looks forward to playing a lead role in shaping this AI-fueled
smart world of the future. Connect with him to be part of this revolution.
➤➤ LinkedIn: www.linkedin.com/in/rabijay1
➤➤ Website: rabiml.com
➤➤ Twitter: https://twitter.com/rabijay1
➤➤ YouTube: www.youtube.com/@rabijay1
➤➤ Podcast: https://open.spotify.com/show/7vCeNI8c02pvIgeYDE8xwS
How to Contact the Author

Connecting and community are vital to the growth of artificial intelligence, which I am passionate about. If you
are passionate about AI strategy, machine learning, generative AI, platform operations, and governance, I encour-
age you to connect with me on LinkedIn at www.linkedin.com/in/rabijay1. You can also continue the
discussion about AI with me on Twitter at @rabijay1. I also encourage you to join my Enterprise AI Newsletter
at www.linkedin.com/newsletters/7033665977435222016.
I appreciate your input and questions about this book. If you have feedback or questions, please contact me.
ABOUT THE TECHNICAL EDITOR
IRWIN CASTELINO IS an executive with 15+ years of experience in digital transformation and data analytics, lev-
eraging AI and ML to optimize business processes for organizations. His expertise spans the implementation and
integration of enterprise applications such as SAP and Oracle. He has delivered predictive solutions to optimize
container shipments and improve equipment maintenance to enable better ROI and uptime of resources. The
solutions he has delivered have consistently been a mix of internal and external data flows, with very large data-
sets being an integral part of the delivered solutions. His clients have ranged across the supply chain, beverage,
food distribution, pharmaceuticals, personal products, banking, energy, chemicals, manufacturing, automotive,
aerospace, and defense industries.
CONTENTS
INTRODUCTION xvii
PART I: INTRODUCTION
CHAPTER 1: ENTERPRISE TRANSFORMATION WITH AI IN THE CLOUD 3
CHAPTER 2: CASE STUDIES OF ENTERPRISE AI IN THE CLOUD 19
PART II: STRATEGIZING AND ASSESSING FOR AI
CHAPTER 3: ADDRESSING THE CHALLENGES WITH ENTERPRISE AI 31
CHAPTER 4: DESIGNING AI SYSTEMS RESPONSIBLY 41
CHAPTER 5: ENVISIONING AND ALIGNING YOUR AI STRATEGY 50
CHAPTER 6: DEVELOPING AN AI STRATEGY AND PORTFOLIO 57
CHAPTER 7: MANAGING STRATEGIC CHANGE 66
PART III: PLANNING AND LAUNCHING A PILOT PROJECT
CHAPTER 8: IDENTIFYING USE CASES FOR YOUR AI/ML PROJECT 79
CHAPTER 9: EVALUATING AI/ML PLATFORMS AND SERVICES 106
CHAPTER 10: LAUNCHING YOUR PILOT PROJECT 152

CONTENTS
PART IV: BUILDING AND GOVERNING YOUR TEAM
CHAPTER 11: EMPOWERING YOUR PEOPLE THROUGH ORG CHANGE

MANAGEMENT 163
CHAPTER 12: BUILDING YOUR TEAM 173
PART V: SETTING UP INFRASTRUCTURE AND MANAGING OPERATIONS
CHAPTER 13: SETTING UP AN ENTERPRISE AI CLOUD PLATFORM

INFRASTRUCTURE 187
CHAPTER 14: OPERATING YOUR AI PLATFORM WITH MLOPS

BEST PRACTICES 217
PART VI: PROCESSING DATA AND MODELING
CHAPTER 15: PROCESS DATA AND ENGINEER FEATURES IN

THE CLOUD 243
CHAPTER 16: CHOOSING YOUR AI/ML ALGORITHMS 268
CHAPTER 17: TRAINING, TUNING, AND EVALUATING MODELS 315
PART VII: DEPLOYING AND MONITORING MODELS
CHAPTER 18: DEPLOYING YOUR MODELS INTO PRODUCTION 345
CHAPTER 19: MONITORING MODELS 361
CHAPTER 20: GOVERNING MODELS FOR BIAS AND ETHICS 377
xiv
CONTENTS
PART VIII: SCALING AND TRANSFORMING AI
CHAPTER 21: USING THE AI MATURITY FRAMEWORK TO

TRANSFORM YOUR BUSINESS 391
CHAPTER 22: SETTING UP YOUR AI COE 407
CHAPTER 23: BUILDING YOUR AI OPERATING MODEL AND

TRANSFORMATION PLAN 416
PART IX: EVOLVING AND MATURING AI
CHAPTER 24: IMPLEMENTING GENERATIVE AI USE CASES

WITH CHATGPT FOR THE ENTERPRISE 433
CHAPTER 25: PLANNING FOR THE FUTURE OF AI 465
CHAPTER 26: CONTINUING YOUR AI JOURNEY 479
INDEX 485
xv
Introduction
WELCOME TO Enterprise AI in the Cloud: A Practical Guide to Deploying End-to-End Machine Learning and
ChatGPT Solutions. This book is the definitive guide to equip readers with the methodology and tools necessary
to implement artificial intelligence (AI), machine learning (ML), and generative AI technologies. You have in your
hands a powerful guide to potentially transform your company and your own career.
In this book, you learn how to
➤➤ Develop AI strategy, solve challenges, and drive change
➤➤ Identify and prioritize AI use cases, evaluate AI/ML platforms, and launch a pilot project
➤➤ Build a dream team, empower people, and manage projects effectively
➤➤ Set up an AI infrastructure using the major cloud platforms and scale your operations using MLOps
➤➤ Process and engineer data and deploy, operate, and monitor AI models in production
➤➤ Use govern models and implement AI ethically and responsibly
➤➤ Scale your AI effort by setting up an AI center of excellence (AI COE), an AI operating model, and an
enterprise transformation plan
➤➤ Evolve your company using generative AI such as ChatGPT, plan for the future, and continuously inno-
vate with AI
From real-world AI implementation, AI/ML use cases, and hands-on labs to nontechnical aspects such as team
development and AI-first strategy, this book has it all.
In a nutshell, this book is a comprehensive guide that bridges the gap between theory and real-world AI deploy-
ments. It’s a blend of strategy and tactics, challenges, and solutions that make it an indispensable resource for
those interested in building and operating AI systems for their enterprise.
HOW THIS BOOK IS ORGANIZED

This book is not just a theoretical guide but a practical, hands-on manual to transform your business through
AI in the cloud. My aim is to provide you with the tools and knowledge you need to harness the power of AI for
your enterprise with a methodology that is comprehensive and deep.
➤➤ Part I: Introduction: In Part I, I explain how enterprises are undergoing transformation through the
adoption of AI using cloud technologies. I cover industry use cases for AI in the cloud and its benefits, as
well as the current state of AI transformation. I also discuss various case studies of successful AI imple-
mentations, including the U.S. government, Capital One, and Netflix.
➤➤ Part II: Strategizing and Assessing for AI: In this part, I discuss the nitty-gritty of AI, such as the chal-
lenges you may face during the AI journey, along with the ethical concerns, and the four phases that you
can adopt to build your AI capabilities. I then discuss using a roadmap to develop an AI strategy, finding
the best use cases for your project, and evaluating the AI/ML platforms and services from various cloud
providers. It’s like a step-by-step guide to your AI adventure.
➤➤ Part III: Planning and Launching a Pilot Project: This part covers all the challenges and tasks centered
on planning and launching a pilot project, including identifying use cases for your project, evaluating
appropriate platforms and services, and launching the actual project.
Introduction
➤➤ Part IV: Building and Governing Your Team: People make magic happen! Part IV explores the organi-
zational changesrequired to empower your workforce. I guide you through the steps to launching your
pilot and assembling your dream team. It’s all about nurturing the human side of things.
➤➤ Part V: Setting Up Infrastructure and Managing Operations: In this part, you roll up your sleeves and
get technical. Part V is like your DIY guide to building your own AI/ML platform. Here, I discuss the
technical requirements and the daily operations of the platform with a focus on automation and scale.
This part is a hands-on toolkit for those who are hungry to get geeky.
➤➤ Part VI: Processing Data and Modeling: Data is the lifeblood of AI. Part VI is where you get your hands
dirty with data and modeling. I teach you how to process data in the cloud, choose the right AI/ML
algorithm based on your use case, and get your models trained, tuned, and evaluated. It is where the
science meets the art.
➤➤ Part VII: Deploying and Monitoring Models: Yay! It is launching time. Part VII guides you through the
process of deploying the model into production for consumption. I also discuss the nuances of moni-
toring, securing, and governing models so they are working smoothly, safely, and securely.
➤➤ Part VIII: Scaling and Transforming AI: You have built it, so now you can make it even bigger! In Part
VIII, I present a roadmap to scale your AI transformation. I discuss how to take your game to the next
level by introducing the AI maturity framework and establishing an AI COE. I also guide you through
the process of building an AI operating model and transformation plan. This is where AI transitions
from the project level to an enterprise-level powerhouse.
➤➤ Part IX: Evolving and Maturing AI: This is where you peek into a crystal ball. I delve into the exciting
world of generative AI, discuss where the AI space is headed, and provide guidance on how to continue
your AI journey.
WHO SHOULD READ THIS BOOK?

This book is primarily meant for those serious about implementing AI at the enterprise level. It is for those
focused on understanding and implementing AI within the context of enabling and executing upon an enterprise-
wide AI strategy. It is a substantive, content-rich resource that requires focused reading and note-taking, with
takeaways that you can go back to your company and start implementing. Below are some examples of roles that
will benefit.
Data Scientists and AI Teams

This includes data scientists, ML engineers, data engineers, AI architects, and AI/ML project managers interested
in learning about the technical and operational aspects of implementing AI in the cloud.
Data scientists and AI teams often struggle with scaling AI/ML models. This book provides comprehensive guid-
ance on processing large volumes of data, choosing the right AI/ML algorithms, and building and deploying
models in the cloud.
IT Leaders and Teams

This includes IT managers, cloud architects, IT consultants, system administrators, and solution architects who
are responsible for the deployment and operation of AI systems and workloads using the cloud.
This book can help you build a scalable and robust AI infrastructure using cloud components. After reading this
book, you can confidently choose the right cloud provider, manage the entire machine learning lifecycle, and
integrate with the backend systems.
xviii
Introduction
Students and Academia

This book helps students and people in academia learn about AI and its practical application beyond just theory.
It is of particular relevance to those studying business, data science, computer science, and AI-related subjects. If
you are looking for a comprehensive and up-to-date treatment of AI implementation, this book is for you.
I encourage you to read the entire book and supplement your reading with additional resources when needed.
Consultants and Advisors

This book will be of great help to consultants and professionals who advise executives, business and technology
professionals about implementing AI for their companies. They are looking for best practices and a structured
methodology.
Business Strategists and Leaders

These are people who may not be technically savvy but are interested in improving their business processes and
strategies using AI.
C-Level Executives
This is a great book for C-level executives who want to learn about the strategic impact, business case, and execu-
tion of AI transformation at scale across the enterprise.
One of their major struggles is applying AI practically and profitably into their business processes and strategies.
This book provides a detailed list of AI use cases and guides them to develop an AI strategy and business case,
thus helping them make intelligent and informed decisions.
WHY YOU SHOULD READ THIS BOOK

We are living in a world where technology is changing rapidly and is impacting both our personal and busi-
ness lives. In particular, AI and cloud technologies are moving at the speed of light, and businesses are racing
to keep up.
For companies, AI and cloud together form the secret weapon to staying competitive, driving cost efficiency, inno-
vating, and delivering outstanding customer experiences.
However, implementing these technologies is not easy. These technologies are constantly evolving with countless
tools, platforms, strategies, and the added complexity of ethical considerations. Therefore, you need a comprehen-
sive, practical guide that helps you to adopt these AI and cloud technologies successfully.
You need a guide that is not just theoretical but is also practical so that you can easily understand, access, and
implement these technologies successfully. This book will act as your roadmap and as a co-pilot to drive the
engine, namely, the cloud and AI technologies.
UNIQUE FEATURES
This book is a toolbox with tools to achieve end-to-end AI transformation. This isn’t your regular technical book,
and here’s why.
xix
Introduction
Comprehensive Coverage of All Aspects of Enterprise-wide AI

Transformation
Most books on AI and ML focus on one or two aspects of AI transformation, such as strategy, architecture, or
operations. Most of them are focused on building and deploying machine learning models. However, this book
covers the entire end-to-end process, from defining an AI strategy to identifying use cases, initiating projects, scal-
ing your operations, and deploying and operating AI models in production. It elevates pet, PoC ML projects to
real-world production-grade deployments.
Case Study Approach

I take a case study approach, using real-world examples to illustrate the concepts I discuss. I include a number of
detailed strategies, step-by-step processes, templates, checklists, best practice tips, and hands-on exercises.
Coverage of All Major Cloud Platforms

I cover all the major cloud platforms, meaning AWS, Azure, and Google Cloud Platform. Most books focus on
only one cloud provider or a specific aspect of AI and ML. But I cover them all, thus making this book a compre-
hensive guide to AI transformation.
Discussion of Nontechnical Aspects of AI

In addition to the technical aspects of AI and ML, I cover the nontechnical aspects, such as AI growth potential,
team development, and AI-first strategy. I believe that it is important to understand the business implications of
AI, as well as the technical aspects.
Best Practices for MLOps and AI Governance

I discuss best practices for MLOps and AI governance. MLOps is the practice of bringing together machine learn-
ing and DevOps, and AI governance is the process of ensuring that AI systems are used responsibly and ethically.
Up-to-Date Content
I believe that my book is the most comprehensive and up-to-date guide to AI transformation available. It is a
must-read for anyone who wants to use AI to transform their business.
I wish you the best on your enterprise AI transformation journey. You can reach me at www.linkedin.com/
in/rabijay1. You can also keep in touch with my latest progress via rabiml.com.
Hands-on Approach
The chapters in this book have been aligned to the various steps in an enterprise AI implementation initiative,
starting all the way from Strategy to Execution to Post go-live operations. And each Chapter contains a number
of review questions to cement the understanding of the topics covered in the book. In addition to that, I have
added a number of hands-on exercises, best practice tips, and downloadable templates with examples to guide
you in your enterprise-wide AI implementation.
xx
PART I
Introduction
In this section, we dive into how enterprises are undergoing transformation through the adoption of AI
using cloud technologies. I cover industry use cases for AI in the cloud and its benefits, as well as the
current state of AI transformation. I also discuss various case studies of successful AI implementations,
including the U.S. Government, Capital One, and Netflix.
1
Enterprise Transformation with
AI in the Cloud
The future of computing is at the edge, powered by AI and the cloud.
—Satya Nadella
Welcome to the exciting journey of enterprise transformation with AI in the cloud! This chapter is designed
for anyone eager to understand the power and potential of AI in today’s business landscape. You’re prob-
ably here because you sense that AI isn’t just another buzzword, but a game-changer. But how exactly can
it transform your business? That’s the question explored in this chapter.
UNDERSTANDING ENTERPRISE AI TRANSFORMATION

Enterprise transformation with AI in the cloud is about more than just adopting the latest technology; it’s
a holistic approach that redefines how your business operates, competes, and delivers value in the modern
world. Through the integration of AI and machine learning with cloud computing, your business can revo-
lutionize its processes, creating efficiencies, personalizing customer experiences, and fostering innovation
like never before.
This chapter sets the stage for the entire book, explaining why some companies succeed with AI, and others
fail, and how you can leverage AI to become a world-class, responsive, and innovative enterprise. It’s not
just about the technology; it’s about reimagining what’s possible in your business. Whether you’re looking
to automate processes, enhance customer experiences, or drive innovation, you’ll find insights and practical
exercises here to help you embark on this transformation.
If you’re planning to follow the approach outlined in this book, the deliverables from the hands-on exer-
cises and the downloadable templates in the accompanying website will become crucial building blocks
as you progress through the later stages of implementing AI in your organization. Let’s get started on this
exciting journey!
NOTE ML stands for machine learning and includes generative AI.

4 ❘ CHAPTER 1 Enterprise Transformation with AI in the Cloud
Why Some Companies Succeed at Implementing AI and ML While

Others Fail
Implementing AI and ML solutions has not been easy. A 2022 PWC Research report shows that only about 20 to
25 percent of the companies have been able to implement AI across the enterprise in a widespread manner (see
Figure 1.1). The research also shows that the majority of the companies have launched limited AI use cases or
pilots but have not been able to scale AI enterprise-wide. Implementing AI comes with several challenges around
data preparation, building models, scalability, and ethical concerns. Companies need a systematic methodology
to leverage several foundational capabilities across business and technology to implement AI and ML technolo-
gies across the enterprise. In this book, you learn how to adopt a step-by-step, practical approach to achieving
enterprise-wide AI transformation.
AI Adoption by Leaders (%) AI Adoption by Others (%)
Considering AI
5%
Considering AI
Tested PoCs
12%
with limited Enterprise-wide AI
success adoption
10% 20%
Enterprise-wide AI
adoption
Launched PoCs/Ready Tested PoCs with
35%
to scale limited success
15% 23%
Limited AI adoption of
use cases
25%
Limited AI adoption of
use cases Launched PoCs/Ready
35% to scale
20%
FIGURE 1.1: Comparison of enterprise-wide AI adoption by leaders and others

Source: Adapted from www.pwc.com/us/en/tech-effect/ai-analytics/ai-business-survey
.html
Transform Your Company by Integrating AI, ML and Gen AI into

Your Business Processes
Companies now have in their hands a fantastic opportunity to change the way they operate. It is now possible to
integrate AI, ML, and Generative AI in many of your business processes to attain outstanding results. For exam-
ple, you can detect quality issues in production, translate documents, and even analyze the sentiments of your
clients in a Twitter feed.
NOTE Enterprise AI transformation involves the implementation of end-to-end AI, ML, and
Gen AI systems to drive business outcomes.
Understanding Enterprise AI Transformation ❘ 5
HANDS-ON CONCEPTUAL EXERCISE: UNDERSTANDING ENTERPRISE AI

TRANSFORMATION
Define enterprise AI transformation and the importance of adopting AI and ML, including
generative AI technologies for enterprises.
Adopt AI-First to Become World-Class

This chapter covers several crucial factors to consider for enterprise AI transformation, what it implies to be an
AI-first company, and why cloud computing is a game-changer in implementing robust, scalable, and ethical AI.
In the next chapter, I present three case studies and introduce the path these businesses took to become world-
class AI-first organizations. The book’s focus is to help you leverage these best practices across business, technol-
ogy, process, and people domains so you, too, can excel as a professional and propel your company to become
world-class, no matter what your role is. As you learn in subsequent chapters, everyone has a role to play.
Importance of an AI-First Strategy

The AI-first mindset introduces a new way to do business. When implementing AI transformation across the
enterprise, it is essential to adopt this AI-first strategy and an AI-first approach. See Figure 1.2.
3 4
2 5
1 Identify Use
Cases
Plan to
Integrate AI
6
Build PoC–
AI-First Set Up
Strategy Infrastructure,
Mindset Tools, and Increased
Develop Skills Innovation,
Current State
Growth, and
AI-First Strategy Profitability
Enterprise AI
Business
Objectives Transformation Measure Results
Identify Opportunities Becoming World-Class
FIGURE 1.2: An AI-first strategy leads to an enterprise transformation.
An AI-first strategy is first a change in mindset that seeks to embrace AI to achieve business objectives. It involves
identifying use cases to adopt AI, planning for proof of concepts (PoCs), and coming up with a plan to integrate
AI into various aspects of an organization, such as customer service, product development, and decision-making.
It includes building the necessary infrastructure, tools, and skills; building a commitment to learning; partnering
with experts and vendors to be at the forefront of AI; and embracing AI for business opportunities.
Prioritize AI and Data Initiatives

Having this AI-first strategy helps organizations prioritize AI and data initiatives ahead of other projects, embrace AI
technologies to stay at the forefront of innovation, and identify new business opportunities with the help of data.
It also promotes greater collaboration between IT and business and gives companies a competitive advantage to
drive innovation, growth, and profitability.
If you do not adopt the AI-first strategy, you run the risk of losing your competitive advantage. You may not use
the resources effectively, continue to live in silos between the business and IT, and, more importantly, lose out on
opportunities to drive innovation, growth, and increased profits.
Adopting an AI-first strategy is the scope of this book. By following the steps outlined in this book, you learn to
implement an AI-first strategy in your company.
HANDS-ON EXERCISE: AI-FIRST STRATEGY FROM CONCEPTS TO APPLICATIONS
Explain what AI-first strategy means and mention at least three benefits of an AI-first strategy.
Using Netflix or any other case study, explain how the adoption of cloud and AI technologies
led to process transformation.
Explain the role of cloud computing in implementing robust, scalable, and ethical AI.
LEVERAGING ENTERPRISE AI OPPORTUNITIES

The rise of AI and the recent developments in generative AI have been incredible. Let’s discuss the reasons behind
such a surge.
One of the main reasons it has taken off is the cloud
computing factor. The other two pillars of this phenom-
enal growth in AI are the growth of big data and the
power of computer algorithms to build models (see Fig-
ure 1.3). However, the common denominator contribut- AI Adoption
ing to this growth is cloud computing.
Enable One-to-One, Personalized,

Real-Time Service for Customers at Software Algorithms
Cloud Computing
Scale
We are currently seeing two trends in the modern world.
One is that customers constantly demand immediacy
Big Data
and relevancy in service. Moreover, they want to be able

to communicate with their service providers in real time.
This, combined with the desire of companies to leverage
data to constantly improve their products and services
and renovate to build new business models, has been an Network Effect
exciting trend.
To provide this level of personalized service that scales FIGURE 1.3: The triumvirate of AI: cloud comput-
in real time with such large amounts of data, it is no ing, big data, and software algorithms
longer possible to rely on traditional rule-based pro- Source: Regormark/Adobe Stock
gramming. This is where the increased, scalable power
of cloud computing powers with large amounts of data streaming from sensors to create new possibilities.
ML algorithms can crunch the same datasets in just a few minutes, something that took thousands of hours
before. This capability makes it possible to provide one-to-one personalized service to your customers rather than
segment-based marketing.
Leveraging Enterprise AI Opportunities ❘ 7
Companies are no longer tethered to their physical data centers. Thanks to Amazon, Google, and Microsoft’s
cloud services, companies can now set up a lot of computing and networking power in minutes.
The cloud provides machine learning frameworks (such as MXNet, TensorFlow, and PyTorch) and AI services
(such as transcription, image recognition, and speech recognition) that anyone can leverage to do intense machine
learning computations to predict, classify, and
enable quick actions.
Innovative Products
Recently, generative AI has caught the attention and Services
of many. It was made possible by the underlying
computing power that the cloud technologies
provide, along with the data infrastructure to
process petabytes of data to train these models.
The Network Effect: How AI-Powered

Services Create a Virtuous Cycle of
Value Generation Network
Effect
Another reason for this explosive growth of AI
is the network effect, which comes from the fact
that you are quickly providing value to the cus-
tomers, who, in turn, are generating more data More Data Satisfied
and generating more insights. The generation Insights Customers
of new insights allows you to provide better
services, generating more data and creating a
virtuous cycle (see Figure 1.4).
FIGURE 1.4: The network effect
Leverage the Benefits of Machine

Learning for Business
Companies have realized that machine learning can help their businesses reduce costs, increase revenue, and drive
innovation. It does create demand for large amounts of computing network and storage power that can be pro-
vided only through cloud computing. Moreover, 90 percent of their costs come from the inference stage of their
machine learning lifecycle.
Cloud Computing Meets Unique ML Hardware Demands

Different machine learning use cases, such as forecasting and image recognition, place different demands on the
underlying hardware, which make it challenging to meet the workload demands with just a standard type of
underlying infrastructure.
NOTE The need for different architectures for different use cases has forced companies to
rely increasingly on cloud computing, which can meet these unique demands on infrastruc-
ture.
Cloud Computing: An Accelerator for the Growth of AI

Cloud computing is scalable, easily accessible from anywhere, and cost-effective. Cloud infrastructure can be eas-
ily integrated with backend systems, allowing models to be hosted in on-prem, cloud, or hybrid mode.
All these factors make cloud computing a significant disruptor and a powerful force that all companies
must leverage.
HANDS-ON RESEARCH PRESENTATION EXERCISE: EXPLORING THE CLOUD AND

ITS IMPACT ON AI
Objective: Exploring the cloud and its impact on AI.

Step 1: Divide the students into groups and have them research the following:
➤➤ Relationship between cloud computing and big data on AI
➤➤ Popular cloud-based machine learning frameworks such as PyTorch, MXNet,
TensorFlow
➤➤ Cloud and AI services
➤➤ Network effects
➤➤ Cloud computing and its impact on various use cases
Step 2: Prepare a paper on their findings with real-world examples and future trends. Discuss
how it could impact their careers.
Enterprise-wide AI Opportunities
Speaking of AI opportunities, you can categorize AI opportunities as enterprise-wide and industry-specific oppor-
tunities. Let’s discuss enterprise AI opportunities first.
Automate Processes for Increased Efficiency and Cost Reduction

You can integrate AI into existing processes to increase operational efficiency, reduce costs, and free up human
time through optimization and automation.
PROCESS AUTOMATION USE CASE EXAMPLES
Customer complaint Some examples of automation include sorting out and categorizing incoming
categorization customer complaints using natural language processing to understand the content
and classify it as technical issues, shipping issues, or billing issues.
Computer vision Computer vision systems can be used to analyze videos and images to spot
systems in defects in products in a car assembly or computer manufacturing plant.
quality control
Optimize Business Processes with AI

AI can also be used to optimize business processes, which is much broader than automation. It includes analyzing
current processes to identify improvement opportunities, such as predicting demand, forecasting resource needs,
and improving team communication. See Figure 1.5.
AI Applications for Improving Customer Experience

AI can help improve customer experience through the following:
CUSTOMER SERVICE USE CASE EXAMPLES
Personalization Using predictive analytics to analyze past customer interactions,

AI can better understand customers’ needs and behavior and
recommend personalized offers and product solutions to problems.
CUSTOMER SERVICE USE CASE EXAMPLES
Improved customer service Better customer service through chatbots using NLP to answer
via chatbots queries 24/7, route to the right person, and schedule appointments.
Targeted marketing campaigns Targeted marketing campaigns by analyzing customer data to

generate relevant messages and thus increase customer loyalty.
Customer recognition with Using computer vision and speech recognition to recognize
computer vision and speech customers on the website and contact centers and to tailor the
recognition experience accordingly.
Sentiment analysis Conduct sentiment analysis of reviews, complaints, and surveys to

recommend solutions to problems.
Speech recognition for call and Use speech recognition to analyze phone calls and chat to identify
chat analysis areas of improvement as well as to train agents.
Transform
Automate
Compliance
Processes
Processes
Enhance
Employee Enterprise AI Optimize
Collaboration Opportunities Processes
Develop
Improve
Innovative
Customer
Products/
Service
Services
FIGURE 1.5: Enterprise AI opportunities
Identify New Customer Needs and Develop Innovative Products and Services
AI can identify new customer needs to develop new products and services by analyzing social media, purchase
history, reviews, complaints, and browsing behavior.
Use Case Example
A financial company may find that customers are paying a third-party service higher fees to move
money overseas, in which case the company can create a money transfer service that’s cheaper
and better quality.
INNOVATION USE CASES
Cross-selling and Companies can use AI algorithms such as recommendation systems to

bundling products discover that customers who buy certain products also buy other related
products. Hence, this gives the company the opportunity to cross-sell or
bundle the products for sale.
Competitor analysis Companies can analyze competitors’ products, pricing, and customer
reviews to identify customer pain points to meet a need in the
marketplace.
Network analysis for Companies can use network analysis of internal and external partners
business expansion and individuals to identify influential clusters in gaining access to more
business ideas and new customers and contact them via social media.
Enhance Employee Collaboration and Decision-Making with AI-Powered Tools and

Workflows
Consider the following use cases that promote employee collaboration and decision-making:
EMPLOYEE COLLABORATION AND DECISION-M AKING USE CASES
Enhanced collaboration via Using chatbots, better decision-making, and predictive insights,
chatbots and predictive insights employees can collaborate more effectively to drive better
business outcomes.
AI-powered workflows for Employees can use AI-powered workflows to collaborate more effectively
complex projects on complex projects. This can help automate routine tasks so employees
can work more strategically.
Analysis of project management By analyzing project management tools, collaboration tools, and
and collaboration tools productivity apps, AI can explain why breakdowns happen along with
solutions. AI can recommend which employee is best suited for a job.
Videoconferencing for remote Employees can use AI-powered videoconferencing to work remotely and
collaboration collaborate virtually from anywhere. AI can recommend ideal times to
have calls.
Transform Compliance Processes by Streamlining Documentation, Risk Assessment,

and Audit
Here are some compliance-related use cases:
COMPLIANCE-R ELATED USE CASES
Automating AI can automate compliance processes such as documentation, audits,

compliance processes and risk assessment to reduce human errors and help employees use
time more strategically.
Predictive analytics for proactive You can use predictive analytics to identify issues proactively before they
issue identification happen and deal with individuals and departments accordingly.
Chatbots and virtual assistants You can also use chatbots and virtual assistants to answer compliance
for compliance questions from employees.
Changing compliance AI can help keep pace with changing compliance requirements
requirements by staying up-to-date and proposing improvement areas in
compliance programs.
CASE STUDY ANALYSIS: EVALUATING AI OPPORTUNITIES ACROSS INDUSTRIES
Objective: Evaluate the practical AI opportunities across industries.

Materials: HBR case collection, Stanford Graduate School of Business case collection,
and so on.
Step 1: Assign a student or group of students to an industry.
Step 2: Choose a real-world company from an industry and assess the following:
➤➤ Specific AI opportunities, technologies, and applications used by the company
and the purpose or problem they solve
➤➤ AI’s impact on people, processes, products, services, and customer experiences
➤➤ Challenges faced by the company during the implementation
➤➤ Measurable outcomes or benefits from the AI implementation
Growing Industry Adoption of AI

Different industries are at different stages of adoption of AI depending upon the availability of the data, the com-
plexity of the industry, and the compliance requirements. Table 1.1 shows the leading companies across pure play
AI, leading pioneers, visionaries, industry leads, and enterprise majors, as well as leaders in generative AI.
TABLE 1.1: Companies Leading the Way in Adopting AI
AI GIANTS PIONEERS VISIONARIES GENERATIVE ENTERPRISE INDUSTRY

AI MAJORS
Microsoft OpenAI Adept Rephrase.ai Salesforce Medtronic

Amazon c3.ai Synthesia Midjourney BMC Software GE Healthcare
AWS H2O.ai Cohere Infinity.ai HPE Capital One
Google DataRobot Abacus.ai Podcast.ai Dell Carnegie
IBM Snowflake Runway Hugging Face SAP Learning
Baidu Dataiku Anthropic Stability AI ServiceNow Century Tech
Oracle RapidMiner Jasper Broadcom Duolingo
Alibaba Databricks SAS Crowdstrike
NVIDIA Alteryx Informatica Palo Alto
Cloudera Shelf Engine

McDonald’s
Fraud Detection, Risk Management, and Customer Service in the Finance and
Insurance Industries
Here are some finance and insurance industry–related use cases:
FINANCE AND INSURANCE INDUSTRY–RELATED USE CASES
Fraud detection, risk AI is widely adopted in the finance industry for use cases such as fraud
management, and detection, risk management, and customer service.
customer service
Loan processing by Human agents are now being replaced by smart robots that can process
smart robots loans in just a few seconds.
Robo-advisors for Financial advisors are being replaced by robots that can process large
investment decisions amounts of data to make the right investment decisions for customers.
These robots are smart enough to analyze data from social media, emails,
and other personal data.
Claims processing and product AI is used in the insurance industry to reduce claims processing time and
recommendation in insurance recommend insurance plans and products based on customer data.
Revolutionizing Healthcare with AI: From Diagnoses to Robotic Surgery

The healthcare industry was an early adopter of AI.
➤➤ AI is used for medical image analysis, personalized medicine, and drug discovery.
➤➤ AI is also used for diagnosing diseases and treatment, medication management, and robotic surgeries.
NOTE Adoption of AI can enable better patient outcomes such as better health
due to better diagnosis and treatment, reduced costs due to reduced patient re-
admissions, and increased operational efficiency.
HEALTHCARE USE CASES
Medical image analysis, AI is used for medical image analysis, personalized

personalized medicine, and medicine, and drug discovery.
drug discovery
Disease diagnosis, treatment, AI is also used for diagnosing diseases and

and robotic surgeries treatment, medication management, and robotic
surgeries.
Improving patient outcomes Adoption of AI can enable better patient outcomes

such as better health due to better diagnosis and
treatment, reduced costs due to reduced patient re-
admissions, and increased operational efficiency.
Transforming Manufacturing with Predictive Maintenance, Quality Control, and

Collaborative Robots
In the manufacturing industry, AI is quite widely used.
MANUFACTURING USE CASES
Predictive maintenance, quality AI is used for predictive maintenance, quality control, and supply
control, and supply chain optimization chain optimization.
Improving production operations It is used to improve production operations ranging from

workforce planning to product design to defect management and
employee safety.
Collaborative robots (cobots) This industry is also seeing the advent of cobots, which are robots
that work collaboratively with humans.
Revolutionizing Retail: AI-Powered Customer Service, Personalized Marketing, and

Inventory Control
In the retail industry, AI is used for customer service chatbots, personalized marketing, inventory control, and
forecasting demand.
Use Case Example
Amazon is a classic example of how AI can be used to provide product recommendations based on the
user’s browsing behavior.
Autonomous Vehicles, Predictive Maintenance, and Traffic Management in the

Transportation Industry
In the transportation industry, AI is used for autonomous vehicles, predictive maintenance, and traffic route
management.
TRANSPORTATION USE CASES
Self-driving cars Active research is being carried out by Tesla, Volvo, Uber, and Volkswagen,
and though it’s used in controlled situations, pretty soon we will be seeing
self-driving cars more commonly on the road.
Navigational and driver Some of the practical applications of AI are navigating obstacles and
assistance systems roads, driver assistance systems such as adaptive cruise control, automatic
emergency braking, lane departure warning, radars, and cameras to enable
safer driving.
Warehouse automation and AI-powered robots are sorting out products in warehouses for delivery,
shipment optimization moving cases within warehouses, delivering for pallet building, and so on. AI
algorithms are used for shipment route determination and optimization for
cost minimization.
Public transport Even in public transport, AI is used for traffic light management,
management transportation scheduling, and routing.
From Pizza Quality to Smart Farming in Food and Agriculture

AI is also used in food tech in the following ways:
FOOD AND AGRICULTURE USE CASES
Taking orders and serving food AI is used to take orders and serve food.
Quality control in food It is even used to ensure the quality of pizzas.

preparation
Crop management in agriculture In agriculture, it is used to raise crops by analyzing salinity, heat, UV light,
and water.
Smart machinery in farming Smart tractors and plucking machines are also being used in the
farming sector.
Predicting Prices, Automating Processes, and Enhancing Customer Experience in Real

Estate
Here are some AI examples in real estate:
REAL ESTATE USE CASES
Price prediction, buyer-seller Predict prices and rental income, match buyers with sellers, analyze
matching, and market analysis market conditions and trends, process real estate documents
automatically, and use chatbots to answer customer queries
Workbook Template - Enterprise AI Transformation Checklist ❘ 15
REAL ESTATE USE CASES
Automated document Smart home technology for increased security and energy efficiency
processing and in the home
customer support
Smart home technology Voice assistants like Alexa, Google Assistant, and Siri; smart appliances
like refrigerators, washing machines, thermostats, smart lighting, home
security systems
AI-Powered Personalization in Entertainment: Netflix, Amazon, and the Gaming

Industry
Here are some use cases in the entertainment industry:
ENTERTAINMENT AND GAMING INDUSTRY
Recommendation systems in Netflix and Amazon use AI to recommend movies and songs to
streaming platforms users based on their preferences and viewing behavior.
Storyboarding and music In the movie industry, AI is used to build a script for storyboarding
composition in movies and is even used to compose music.
Storyline advancement in gaming AI is used in the gaming industry to advance the storyline in a
specific direction to keep the player engaged.
Improving Efficiency, Reducing Costs, and Advancing Sustainability in the Energy

Industry
Examples in the energy industry follow. The adoption of AI is contributing to lower costs and improved environ-
mental sustainability.
ENERGY USE CASES
Predictive maintenance Predictive maintenance of equipment such as transmission lines and power
and energy management plants, managing the supply and consumption of energy, and integrating with
renewable energy to reduce costs.
Energy efficiency Improve energy efficiency in consumption in office spaces and homes by
in buildings adjusting cooling and heating systems and lighting
Fraud prevention Prevent fraud such as billing errors and meter tampering
WORKBOOK TEMPLATE - ENTERPRISE AI TRANSFORMATION

CHECKLIST
Download the “Enterprise AI transformation checklist” template from the download section of the book (www
.wiley.com/go/EnterpriseAIintheCloud). Please use this as an initial starting point to assess where your
company is on its journey to implement AI across the enterprise. In areas where you answered “no,” it indicates
that more work is needed. Continue working on each section as you progress through this book. Use this as a
healthy reminder for your team to track their progress throughout the AI transformation journey and celebrate
minor and significant milestones.
SUMMARY
The key takeaways from this chapter underscore the necessity for organizations to take AI seriously to gain a
competitive edge. This means elevating machine learning projects from mere PoC pet projects into large-scale,
enterprise-wide transformative efforts. To achieve this, you need to adopt a more systematic AI-first strategy and
methodology, the primary focus of this book.
This chapter also delved into the transformative power of cloud computing, enabling many use cases previously
unattainable and providing a pathway to thrilling consumer experiences. You reviewed some of the use cases that
AI and ML enabled and realized their broad adoption and impact across various industries.
The next chapter showcases three case studies of companies that successfully implemented AI on an enterprise-
wide level.
REVIEW QUESTIONS
These review questions are included at the end of each chapter to help you test your understanding of the infor-
mation. You’ll find the answers in the following section.
1. AI and ML can be incorporated into business processes to
A. Detect quality issues in production
B. Analyze the user sentiment in a Twitter feed
C. Translate documents
D. All the above
2. An AI-first strategy
A. Prioritizes customer service over AI and ML technologies
B. Focuses on product management over AI technology
C. Focuses on embracing AI to achieve business objectives
D. Focuses on building a solid technology platform for AI
3. Prioritizing data and AI initiatives helps
A. Companies to focus on customer service
B. Companies to identify new business opportunities
C. Companies to enable greater collaboration between IT and business
D. Companies to build stronger business cases
4. Neglecting an AI-first strategy will lead to
A. Losing the competitive edge
B. Not using resources effectively
C. Living in silos between AI and the business
D. Innovation, growth, and increased profits
5. What is the main advantage that digital natives have when it comes to adopting AI?
A. They have a risk-taking culture and view failure as a steppingstone for success.
B. They have a lot more money to invest in AI.
C. They have a better brand image than traditional companies.
Review Questions ❘ 17
6. What are the factors to consider for an organization’s AI journey?

A. Its AI readiness
B. Its risk-taking culture
C. Data and resources availability
D. All the above
7. Which of the following is NOT one of the three pillars of AI growth?
A. Cloud computing
B. Big data
C. Computer algorithms
D. Robotics
8. _____is NOT one of the benefits of cloud computing for businesses using machine learning.
A. Scalability
B. Accessibility
C. Increased costs
D. Cost effectiveness
9. Explain the network effect in the context of AI-powered services.
A. Companies can now set up computing power in minutes.
B. Machine learning helps businesses reduce costs, increase revenue, or drive innovation.
C. Companies can provide better services to customers, generating more data and insights in a vir-
tuous cycle.
D. Different machine learning use cases place different demands on the underlying hardware.
10. Choose an example of how AI can improve customer experience.
A. By analyzing project management tools to recommend solutions
B. By identifying clusters that are influential to gain access to more business ideas
C. By using chatbots to answer queries 24/7
D. By automating compliance processes such as documentation and risk assessment
11. Choose an example of how AI can be used to identify new customer needs.
A. By using computer vision to recognize customers on websites and contact centers
B. By analyzing social media, purchase history, and reviews
C. By using network analysis to gain access to more business ideas
D. By automating compliance processes such as audits and risk assessment
12. Cobots are used in the manufacturing industry for which purpose?
A. Workforce planning
B. Product design
C. Collaborating with humans
D. Defect management
13. AI is used in the entertainment industry for which purposes?
A. Building a script for storyboarding
B. Composing music
C. Advancing the storyline in a certain direction
D. All the above
ANSWER KEY
1. D 6. D 11. B
2. C 7. D 12. C
3. B 8. C 13. D
4. A 9. C
5. A 10. C
2
Case Studies of Enterprise
AI in the Cloud
Every company is a software company. You have to start thinking and operating like a digital company.
It’s no longer just about procuring one solution and deploying one. It’s not about one simple software
solution. It’s really you yourself thinking of your own future as a digital company.
—Satya Nadella, CEO, Microsoft
Now that you have learned about some of the possibilities of AI in the previous chapter, this chapter dives
deeper into the fascinating world of enterprise transformation with AI in the cloud through three real-
world case studies. If you are interested in knowing how AI transforms industries, this chapter is tailor-
made for you.
You not only learn how giants like the U.S. government, Capital One, and Netflix have leveraged AI and
cloud technologies, but their stories also serve as practical examples that inform the methodology that I
present in this book.
The key takeaway is how these organizations built resilient systems by embracing cloud-native principles
and AI technologies to become AI-first companies, which is essential for any enterprise aspiring to thrive in
the digital age.
CASE STUDY 1: THE U.S. GOVERNMENT AND THE POWER

OF HUMANS AND MACHINES WORKING TOGETHER TO SOLVE
PROBLEMS AT SCALE
This is an excellent example of an augmented AI case study. The U.S. government spends nearly $2.5
trillion trying to serve millions of customers with social and safety net programs such as Social Security,
Medicaid, Medicare, and the Supplemental Nutritional Assistance Program (SNAP).
Figure 2.1 shows a high-level view of the challenges, benefits, and solutions for this case study.
20 ❘ CHAPTER 2 Case Studies of Enterprise AI in the Cloud
US Government: Human and Machine Working Together
Effective Customer Communications

Improved Workforce Productivity
Improved Program Costs
Benefits
Improved Beneficiary Experience

High Visibility of Operations
Modern Contact Centers Capabilities
Augmented AI, ML
Cloud Migration
Data Lakes Foundation
FIGURE 2.1: Challenges, benefits, and solutions adopted by the U.S. government
The following table lists the challenges faced by the U.S. government:
CHALLENGES DETAILS
Enrollment process Given the large volume of applications, one can only imagine how complex
complexity this process can be and how long it can take. Millions of customers depend on
this for their healthcare, unemployment needs, and to stay out of poverty. The
challenge with the enrollment process includes complex application processes,
huge backlogs in processing the claims, and lengthy adjudication, evaluation, and
eligibility processes.
Long customer Customers had to wait several weeks to get a response even though a large
response time workforce was trying to serve them.
Inadequacies of Legacy systems provided poor customer experience and inability to scale during
legacy systems times of surge in demand. They also lacked self-service capabilities and did not
have a mobile application submission option. Text communications for case status
and the ability to schedule appointments online needed to be included.
Multisource The applications came from multiple sources, such as the web, mail-in, and contact
application processing centers, and in multiple formats, such as PDFs and images. Reviewing these
applications in such large quantities took a lot of time and effort, was error-prone,
needed domain expertise, and was not an efficient process. Setting up human
review systems to process this many applications not only was costly but also took
a long time. It involved custom software for review tasks, complex workflows, and
many reviewers.
Case Study 2: Capital One and How It Became a Leading Technology Organization ❘ 21
CHALLENGES DETAILS
Security concerns Another challenge was the need to maintain data security as millions of documents
amounting to terabytes of data needed to be stored and sensitive data needed to
be protected and encrypted.
High volume of Another major challenge was the number of customer service calls, amounting to
customer service calls more than 50 million calls just for the Social Security Administration team. These
calls were related to pending application status disputes and other benefits-related
questions.
Lack of Moreover, management lacked insight into the program operations, enrollment
management insight issues, backlogs, budgeting, waste, fraud, and abuse.
Revolutionizing Operations Management with AI/ML

The solution to these problems involved simplifying the application process by automating some of the work-
flows behind the approvals using AI/ML technology, enabling self-service capabilities, streamlining the interview
and appeal process, and adding insights around the operations to help management solve issues and propose
improvement opportunities.
Enabling Solutions for Improved Operations

To resolve these challenges, the solution capabilities outlined in this table were enabled:
SOLUTION CAPABILITY BENEFITS
Chatbots Improved customer experience
AI and A2I Automated some of the workflow steps such as adjudication and approvals
technologies
Data-driven insights Provided leadership with the required information to take remedial actions
Contact center Served customers efficiently with their queries
Cloud migration Leveraged some of the operational efficiencies that come with the cloud such as
scalability, efficiency, and performance
CASE STUDY 2: CAPITAL ONE AND HOW IT BECAME A LEADING

TECHNOLOGY ORGANIZATION IN A HIGHLY REGULATED
ENVIRONMENT
Capital One wanted to be a technology organization, irrespective of the fact that they were in the finance indus-
try. They achieved this through a specific growth and innovation template that many organizations must adopt to
become a leading technology organization, regardless of which industry they belong to.
Capital One had to solve a number of challenges, some of which are listed here:
CHALLENGES DETAILS
Data management To begin with, they had the data challenge. They had data spread across multiple
parts of the organization, such as customer relationship management (CRM),
enterprise resource planning (ERP), and streaming data from social media. They had
to bring all that into one place, which they did with Amazon S3.
Infrastructure They wanted to solve their infrastructure problems. They wanted to become an agile
challenges company and reduce their development time from months to days or even minutes.
Cloud They wanted to migrate their 11 data centers to the AWS cloud and leverage the
migration challenge cloud-native development capabilities to develop a resilient system.
Building Amazing Experiences Due to Data Consolidation

To solve their data challenge, they built a data lake. It allowed them to use data as fuel for their present and
future AI models, and they were able to train their models continuously. This helped them to understand human
behavior across multiple channels and time zones in real time and helped them to design amazing human experi-
ences. See Figure 2.2.
Capital One: Technology Leader in a Regulated Industry
Improved Customer Experience

Enhanced Decision Making
Increased Efficiency
Benefits
Cloud Native Apps

Machine Learning
DevOps Solutions
Long Development Cycles

Legacy Infrastructure
Data Silos Challenges
FIGURE 2.2: Challenges, benefits, and solutions adopted by Capital One
Becoming Agile and Scalable by Moving Data Centers Into

the Cloud
How did they do it?
➤➤ They moved their eight data centers to the cloud.
➤➤ They built an 11,000-member technology team over eight years.
➤➤ They instituted training plans and programs for the AWS cloud.
Case Study 2: Capital One and How It Became a Leading Technology Organization ❘ 23
In the process, they were able to reap the following benefits:

➤➤ They reduced the environment build time from three months to three minutes.
➤➤ They had a high-performing agile and scalable infrastructure, which helped them go to market with
their products much earlier, from months to days.
➤➤ They were even able to release applications that took months and weeks to more than once a day.
Moving to the cloud was one of the secrets that many organizations took on this journey to become an AI-
first company.
Building a Resilient System by Embracing Cloud-Native Principles

So far, this section has discussed Capital One’s cloud adoption journey, where they overcame data challenges and
migrated their apps to the cloud across the data centers. However, this transition was a lot of work. Throughout
this process, though, the cloud offered them new capabilities and hence new possibilities.
Adopting cloud-native principles enabled Capital One to build a resilient system by adopting microservices archi-
tecture, CI/CD pipelines for automated workflow deployments, blue-green switching, A/B testing, and failover
techniques that resulted in minimal impact to the end-user experience. We will cover these topics in this book.
Impact of Cloud-First Thinking on DevOps, Agile Development, and

Machine Learning
However, what is even more significant is the mindset shift that the cloud adoption created, making DevOps and
the Agile development methodology possible. The ability to release features quickly allowed product owners to
become more customer-centric, creating more engaging user experiences not only from a functional perspective
but also even a technical perspective through good performance, high availability, and resiliency. As a result, the
teams also became more product-centric, breaking down organizational silos due to the adoption of DevOps.
NOTE Cloud-first thinking lays the foundation for a machine learning platform by adopting
cloud principles when developing applications, deploying them, and serving your customers.
Becoming an AI-First Company: From Cloud Adoption to Thrilling

Customer Experiences
By becoming a cloud-first company, it became easier for Capital One to become an AI-first company. With a data
lake and an agile and scalable infrastructure, it became possible to process large amounts of data in real-time
from multiple sources and then feed them into AI/ML models, which in turn could process them using ample
computing resources and generate insights in real time. This ability to generate insights in real time opened several
opportunities for Capital One.
Moreover, by combining AWS’s high-level AI services and machine learning services like natural language process-
ing (NLP), Capital One could provide thrilling customer experiences through their Eno chatbot and detect and
prevent fraud in real time. They also leveraged machine learning for speech recognition technology in their call
centers to automate training and validation systems. Other applications included credit risk modeling, investment
portfolio analysis, risk management, and image recognition to open bank accounts; they even used ML to
ensure image recognition to open bank accounts. As shown in Figure 2.3, Capital One transformed itself into a
pioneering AI-first company and unleashed the power of its 50,000 associates.
THRILLING
CUSTOMER
EXPERIENCES
AI-First
Company
Mindset
Shift
Cloud-Native Eno chatbot,
Principles fraud detection,
speech
Infrastructure Generating new recognition in call
Transformation insights from centers, etc.
AI/ML models,
Data DevOps, Agile resulting in high
Challenges development, scalability,
and customer- reliability, high
Adopting centric approach availability, and
microservices reduced cost
architecture,
Migration into the CI/CD pipelines,
Cloud: Reduced blue-green
environment build switching,
Scattered data time from months A/B testing,
collected into a to minutes failover
data lake techniques
FIGURE 2.3: Capital One’s transformation journey to AI-first status
CASE STUDY 3: NETFLIX AND THE PATH COMPANIES TAKE

TO BECOME WORLD-CLASS
This section discusses the path organizations take (Netflix, particularly) to become world-class companies (see
Figure 2.4). It can be the path your company may take to become world-class.
Knowing this will help you understand the foundational capabilities you need to adopt and improve as part of
the enterprise AI transformation effort. This section explains what Netflix did to become world-class and the
foundational capabilities they leveraged, which is the core subject of this book—to help companies leverage their
foundational capabilities to become world-class.
NOTE You will find that most companies follow this path—adopting technology results in
process improvement. Process improvement leads to organizational change, which results in
product transformation, which leads to innovation in developing new products and services,
thus generating new business outcomes.
Netflix’s story can easily be applied to any digital-native company such as Uber, Amazon, and Airbnb. They have
all followed this path to become world-class, AI-first companies.
Cloud and AI Technology: A Game-Changer for Netflix’s Business

Model and Success
By adopting cloud and AI technologies, Netflix has been able to drive process transformation. Here are some of
the benefits:
Case Study 3: Netflix and the Path Companies Take to Become World-Class ❘ 25
Technology Adoption
Cloud Computing
Big Data
Machine Learning
Organizational Process Improvement

Excellence
DevOps
Customer-Centric Agile
Collaboration Product Management
Product-Centric
Product Transformation
Product Innovation
Intelligent Products
New Business Models
FIGURE 2.4: The path to becoming world-class
ADOPTION OF CLOUD AND AI IMPACTS AND OUTCOMES

TECHNOLOGIES
Transition to streaming service By adopting the cloud, Netflix could switch from DVDs to a streaming
service. It helped Netflix not only disrupt its competitors but also
become a global company almost overnight, as it could reach people
worldwide.
Implementation of machine Cloud technology also allowed Netflix to use complex,

learning algorithms sophisticated machine learning algorithms to provide personalized
recommendations to customers and improve their experiences.
Adoption of a content Similarly, by adopting a content delivery network, it became possible

delivery network for Netflix to deliver high-quality video services to its customers and
enhance their experiences.
Utilization of data analytics Netflix used customer behavior and profile data from a data analytics
platform to generate new insights and drive their content production,
acquisition, and marketing to focus on their customer’s needs.
Cloud Infrastructure and AI Adoption Drives Process Transformation

From a pure infrastructure perspective, by adopting cloud computing, Netflix has been able to reduce its infra-
structure costs, store and stream data continuously, and scale its computing resources up and down based on the
fluctuating demand on its systems.
Netflix was able to leverage cloud and AI technology to drive process change. Netflix transformed its business
and technology operations by reducing its infrastructure costs, increasing the speed at which it delivers content
to its customers, and enhancing the customer experience through personalized recommendations and marketing
insights through data.
Process Transformation Drives Organizational Change

The next question is how this process transformation caused organizational transformation at Netflix. It is essen-
tial to note that technology adoption led to process transformation, making specific organizational transforma-
tion opportunities possible; this table shows the organizational transformation stages at Netflix:
ORGANIZATIONAL TRANSFORMATION DESCRIPTION

STAGES
Technology adoption By adopting technology, Netflix reduced costs while improving its
operational efficiency and customer experience.
Driving process change It also allowed them to drive process change by changing their
operating model around products and services.
This ability to quickly develop new software helped Netflix focus
more on products and reorganize their business model around
the products and services.
Breaking down of functional silos and It has also helped them break down their functional silos. In the
organizational transformation process, they became an agile, product-centric, and customer-
focused organization.
Thus, the process transformation led them to drive organizational
change as a product-centric, customer-centric, agile organization.
Product innovation driven by Netflix leveraged its new organizational transformation to drive
organizational transformation product innovation.
By being customer-centric, Netflix has been able to come up with
ideas such as new products and services, for instance, streaming
original content. Their customer centricity has also helped them
prioritize their recommendation algorithms as part of their
streaming service.
Continuous innovation and customer Being agile helped Netflix identify the requirement for new
centricity customer experiences on a regular basis. They created local
material to fulfill regional demands, which has helped them
increase globally.
Their customer centricity is also helping them determine the need
for new programs, such as helping customers pick their journey in
their hit program called Bandersnatch.
NOTE The cloud made becoming agile easy. By adopting the cloud, Netflix could develop
software iteratively and quickly. In other words, it helped them adopt an agile development
approach rather than a waterfall methodology. Agile, in turn, helped them become more
product-centric and customer-focused.
By adopting cloud/AI innovation, you can see that companies find it simpler to transform their processes, causing
organizational transformation. Consequently, they become more product-centric and customer-centric.
In turn, this creates the right conditions for innovation, breaking down silos and encouraging experimentation
that eventually results in new products, services, and business models like you just saw in the Netflix
example.
AI helps you respond quickly to business events in an agile manner. Agility is one of the few competitive advan-
tages companies have today.
Today’s customers want to be treated with unique offers based on their specific needs, and in real time. That
means the business must crunch mass amounts of data to generate new insights and act in minutes rather than
what used to be days and weeks. Machine learning has now made this possible. You too can achieve these feats
by adopting the best practices and the comprehensive methodology in this book.
HANDS-ON EXERCISE 1: ANALYZE THE NETFLIX AI TRANSFORMATION
Explain how Netflix enabled process transformation, organizational change, and product
innovation. Research other companies that have achieved similar transformation.
WORKBOOK TEMPLATE - AI CASE STUDY

Download the “AI Case Study” template from the download section of the book (www.wiley.com/go/
EnterpriseAIintheCloud). Capture the case studies from other organizations that you plan to explore and
drive your AI implementation with the benefit of a comprehensive view of the AI implementation process, from
inception to results. Adopt their best practices and avoid their mistakes.
SUMMARY
This chapter delved into case studies of how three organizations transformed into world-class organizations by
adopting technology that spurred process improvement and organizational change. This transformative journey
eventually helped them become customer-centric, leading to product innovation and even pioneering new busi-
ness models.
The next chapter unravels some of the challenges of implementing a machine learning deployment that might
prevent organizations from deploying one into production.
REVIEW QUESTIONS
1. The U.S. government solved its challenges processing multisource applications by
A. Setting up human review systems
B. Using custom software to manage review tasks
C. Automating some of the workflows using AI/ML technology
D. All the above
2. What is the common theme among companies that have become world-class?
A. They focus on reducing business risks.
B. They prioritize enhancing their performance in terms of the environment, society, and governance.
C. They leverage technology to achieve business outcomes.

D. They aim to decrease their revenue.
3. Netflix’s process transformation led to which of the following?
A. Organizational change
B. Decrease in customer experience
C. Increase in infrastructure costs
D. Decrease in operational costs
4. ____ helped Netflix become agile and adopt a product-centric approach.
A. Blockchain
B. Augmented reality
C. Cloud computing
D. Machine learning
5. The path to becoming world-class is
A. Technology ➪ Process ➪ Organization ➪ Product transformation
B. Organization ➪ Technology ➪ Process ➪ Product transformation
C. Process ➪ Organization ➪ Product transformation ➪ Technology
D. Product transformation ➪ Organization ➪ Process ➪ Technology
ANSWER KEY
1. D 3. A 5. A
2. C 4. C
PART II
Strategizing and Assessing for AI
In this part, we discuss the nitty-gritty of AI, such as the challenges you may face during the AI journey
along with the ethical concerns, and the four phases that you can adopt to build your AI capabilities.
I then discuss the roadmap to develop AI strategy, finding the best use cases for your project, and evaluat-
ing the AI/ML platforms and services from various cloud providers. It’s like a step-by-step guide to your AI
adventure.
3
Addressing the Challenges with
Enterprise AI
Our greatest weakness lies in giving up. The most certain way to succeed is always to try just one
more time.
—Thomas A. Edison
This chapter might just be your ally on your path to AI transformation. As someone keen on implementing
AI, whether you are a business strategist, technology leader, data scientist, or curious innovator, you must
be aware that deploying AI, ML, and generative AI solutions is exciting, but it can also be like navigating a
minefield. That’s because AI isn’t just about algorithms and models; it’s also intrinsically connected to data,
infrastructure, change management, regulations, and strategic decision-making. You will soon realize that
an AI transformation is markedly different from a digital transformation.
This chapter is designed to help you understand the technical challenges, ethical dilemmas, and strategic
hurdles. Recognizing these challenges is the first step toward preparing your organization for them. In later
chapters, I guide you with a practical, systematic methodology to address these challenges with clearly
defined steps, best practices, case studies, hands-on exercises, and templates. Each chapter in this book
maps to the tasks in the phases shown in Figure 3.1.
Note that these challenges by no means represent a complete list. Although the challenges are many, they
are not insurmountable. You can convert these challenges into opportunities for innovation and growth
based on the insights and best practices I share throughout this book.
CHALLENGES FACED BY COMPANIES IMPLEMENTING

ENTERPRISE-WIDE AI
Let’s take a quick peek into various kinds of challenges. This list is not exhaustive but should give
you an idea.
32 ❘ CHAPTER 3 Addressing the Challenges with Enterprise AI
Enterprise AI Journey Map

02
01 Strategize and Plan and
Prepare Launch Pilot
03
Continue Your Build and

AI journey Govern Your
Team Strategize and Prepare
09
ENTERPRISE AI • Address the challenges with AI
04 • Design AI systems responsibly
Evolve and JOURNEY MAP Setup
Infrastructure • Envision and align
Mature AI
and Manage • Develop business case and AI strategy
Operations • Strategize
Manage and Prepare
strategic change
08
Scale and Process Data
Transform AI and Modeling Objective
Deploy and
05 Navigating the multifaceted challenges of
Monitor Enterprise AI integration, from data
Models
intricacies to business considerations
07
06
FIGURE 3.1: STRATEGIZE AND PREPARE: Address the challenges with AI
Business-Related Challenges
Here is a sample of business-related challenges:
BUSINESS-R ELATED CHALLENGES
Cost of running The cost of running intensive machine learning workloads may come in the way of
ML workloads deploying models into production.
Keeping up AI and machine learning are constantly evolving, and it is challenging for AI/ML
with change practitioners and cloud engineers to keep up with those changes.
Selecting the Choosing the right use cases can be challenging, as you need to understand the
right use cases business problem, ensure data availability for the use cases, choose the right ML
model, integrate with backend systems, factor in the ethical and legal issues, and
measure ROI for the use cases.
Alignment with Alignment between AI/ML and business strategies can be challenging if not handled
business strategy properly with business prioritization of requirements, understanding AI impact, and
stakeholder communication.
Skills and Challenges include having the right skills and talent due to the shortage of qualified
talent shortages professionals, needing cross-functional expertise, and trying to keep up with new
developments.
Challenges Faced by Companies Implementing Enterprise-wide AI ❘ 33
Data-and Model-Related Challenges

Here is a sample of data-and model-related challenges:
DATA- AND MODEL-R ELATED CHALLENGES
Collecting high- Collecting high-quality data can be challenging and time-consuming. Models
quality data require data in the right format, so data cleaning and labeling data for model
training is required, which can be time-consuming and costly.
Complexity of Machine learning can be complex, with statistics, mathematical concepts, and
machine learning machine learning programming skills that make navigating challenging.
Measuring model Measuring the performance of models, coming up with the proper performance
performance metrics, and validating them during changing business conditions and data can be
challenging.
Monitoring model Monitoring the model performance can be challenging because of changing
performance business conditions, the ability to measure model drift, and the need to ensure the
model performs well over time and aligns with the business needs.
Choosing the Choosing the right model is a challenge as simpler models do not capture all the
right model underlying data patterns; complex models can capture complex patterns, but they
can remain black boxes and are difficult to interpret.
NOTE Poor data quality can impact model performance and sometimes prevent models
from being deployed in production.
Platform-Related Challenges
Here is a sample of platform-related challenges:
PLATFORM-R ELATED CHALLENGES
Integrating with Integrating with backend systems and infrastructure can be challenging due
backend systems to technical, data, security, scalability, human-in-the-loop integration, and
regulatory challenges.
Managing the entire Managing the entire machine learning lifecycle is challenging due to the
machine learning lifecycle complexity around data preparation, feature engineering, model selection, and
tuning, evaluation, deployment, monitoring, model interpretation, and scaling
for large datasets.
Scaling machine Scaling machine learning workloads for large amounts of data in real-time can
learning workloads be challenging, as doing so can require significant computational resources.
Selecting the right Selecting the right cloud provider to ensure proper performance, scalability,
cloud provider reliability, and security, choosing from a wide range of available complex
options, and ensuring alignment with the business goals and culture.
Table 3.1 summarizes all the challenges related to AI transformation covered so far.
TABLE 3.1: AI Transformation Challenges
DOMAIN CHALLENGES
Business The cost of running intensive machine learning workloads.

AI and machine learning are constantly evolving fields.
Choosing the right use cases is tricky.
Data Collecting high-quality data.

Alignment between AI and ML strategy with the business strategy.
Ability to process large amounts of data.
Model Machine learning can be complex.

Measuring the performance of models is tricky.
Monitoring the model performance is a challenge.
Platform Integrating with backend systems.

Managing the entire ML lifecycle.
Scaling ML workloads.
Selecting the right cloud provider.
Operations Model should be monitored for bias, data, and model drift.
Security Managing end-to-end security.

Ensuring compliance with legal regulations.
Governance Having the right skills and talent.

Ensuring responsible and ethical AI.
Here is a sample of security-related challenges:
SECURITY-R ELATED CHALLENGES
Managing end-to- Managing end-to-end security across the entire machine learning lifecycle can be
end security challenging, as security needs to be implemented across the data layer, models,
infrastructure, network, and cloud components.
Ensuring compliance Ensuring compliance, maintaining data privacy, implementing security, and
and avoiding bias avoiding bias and ethical discrimination in models can be challenging.
HANDS-ON BRAINSTORMING EXERCISE: DEVELOP IMPLEMENTATION PLAN TO

OVERCOME AI CHALLENGES
Brainstorm with your team the different ways AI, ML, and Generative AI can integrate with
your company or a company of your choice.
Get Ready: AI Transformation Is More Challenging Than Digital Transformation ❘ 35
HOW DIGITAL NATIVES TACKLE AI ADOPTION

Digital natives, which are companies born in the digital age, are technology savvy and have embraced technol-
ogy as a primary driver for achieving business results, such as cost reduction, efficiency, and innovation. They
are agile, adaptable, and innovative and are quick to adopt new technologies, such as machine learning, artificial
intelligence, and now generative AI. They are better able to implement AI transformation across the enterprise
when compared to other traditional companies. Let’s see why.
They Are Willing to Take Risks

When it comes to adopting new technology, such as AI, digital natives tend to be more willing to take risks and
experiment with new technology. Unlike traditional companies, they have a culture that encourages risk-taking
and views failure as a stepping stone to success.
They Have an Advantage in Data Collection and Curation

Capabilities
When it comes to data, digital-native companies have been adept at gathering large amounts of data from mul-
tiple sources and curating them for business purposes. This is something that traditional companies are not able
to do because of the technical debt they have been accumulating as more and more systems have been developed
over several years. It is a challenge to get the data consolidated into one place to enable consumption.
They Attract Top Talent Through Competitive Compensation and

Perks
Digital-native companies have been able to hire the best talent in the industry by giving them good pay, attractive
career opportunities, and work amenities that attract good talent.
For this reason, I recommend that traditional organizations spend more time and effort on change management.
They should plan for internal employee training and partner with other AI experts to provide guidance and sup-
port during the AI transformation journey.
BEST PRACTICE TIPS TO TACKLING AI ADOPTION
➤➤ Start AI projects with small scope and gradually scale.

➤➤ Partner with other AI experts and companies.
➤➤ Invest in training programs for employees on AI technologies.
GET READY: AI TRANSFORMATION IS MORE CHALLENGING THAN

DIGITAL TRANSFORMATION
AI transformation is more challenging than digital transformation. Understanding these challenges will help you
to plan, execute, and manage your AI implementation. Let’s look at those reasons.
Complexities of Skill Sets, Technology, and Infrastructure Integration

One of the main reasons is that AI transformation is much more complex given the many skills (such as machine
learning, data science, cloud engineering) required, the complexity of the underlying technology, and the need for
business domain knowledge. Bringing the data, the technology, the skill sets, and the infrastructure together takes
a lot of work. A strong mix of technology and business domain knowledge is the foundation for success in this
journey. Figure 3.2 shows the different skills you need. It shows how these skills are a combination of domain,
mathematics, and computer science knowledge areas.
Ml
Finance Engineer DevOps
Data
Scientist
Business Data
Analyst Engineer
Computer Science
Marketing Mathematics
Operations
Domain Expertise
FIGURE 3.2: Skill sets required in a typical AI/ML project
The Importance of Data Infrastructure and Governance

Deploying high-performing models in production requires high-quality data in large quantities. A robust data
infrastructure pipeline, from collection to consumption, is vital (see Figure 3.3). Collecting, ingesting, cleaning,
storing, and consuming data is a challenging task. Companies that do not have this in place already are at a
disadvantage in getting started.
Sources Ingestion and Storage Query and Transformation Analysis and

Transport Processing Output
• ERP • Data • Data Lake • Spark • Data • Dashboards
• CRM Replication • Real-Time Platform Modeling • Embedded
• Apps • Workflow Analytics • SQL Query • Workflow Analytics
• Logs Manager Database Engine Manager • Data
• Third-Party • Event • ML Platform Workspace
APIs Streaming • Custom
• File Storage Application
FIGURE 3.3: A typical data infrastructure pipeline
Change Management to Redefine Work Processes and Employee

Mindsets
AI transformation, unlike digital transformation, demands shifts in work processes and mindsets. Successfully
managing this change is a significant task.
BEST PRACTICE TIP Enterprise-wide AI transformation is complex, so change

management is required to train employees and guide them through the process.
Choosing Between Smaller PoC Point Solutions and Large-Scale AI Initiatives ❘ 37
Regulatory Concerns: Addressing Bias, Ethical, Privacy, and

Accountability Risks
AI introduces regulatory concerns such as bias, ethics, and privacy that must be tackled before deploying models.
HANDS-ON ROLE-PLAY EXERCISE: DECIDING TO PROCEED WITH AN AI-

FIRST STRATEGY
Imagine a scenario where a company is contemplating adopting an AI-first strategy. Form

groups and have each group or an individual take on roles such as CEO, CFO, data scientist,
machine learning engineer, business analyst, product manager, and so on.
Have each person defend their role and brainstorm reasons for and against adopting an AI-
first strategy.
Come up with a presentation capturing your decision in favor or against an AI-first strategy
with justifications.
CHOOSING BETWEEN SMALLER PoC POINT SOLUTIONS AND

LARGE-SCALE AI INITIATIVES
Let’s discuss the pros and cons of choosing between single point and large-scale AI initiatives. This will help you
to manage your resources better, balance risks versus benefits, and, more importantly, ensure alignment of your AI
initiatives with your organizational goals.
The Challenges of Implementing a Large-Scale AI Initiative

Large-scale AI initiatives are much more complex than smaller point-based PoC solutions. Choosing a large-scale
AI initiative or a smaller-point PoC solution depends on your organization’s needs, goals, strategic objectives, and
readiness.
Navigate the Moving Parts, Stakeholders, and Technical

Infrastructure
Large-scale AI initiatives are complex simply because of the many moving parts, such as the data pipeline, the
underlying systems and processes, the models, the business requirements, and the many stakeholders you need to
work with. Getting to know the processes, the data, and the technical infrastructure takes time and effort.
Resource Allocation Challenges in Large-Scale AI Initiatives

Large-scale AI initiatives require a lot of technological, financial, and human resources. It naturally involves coor-
dinating with lots of stakeholders within the organization. Getting buy-in and consensus takes work.
Overcome Resistance to Change

Implementing a large-scale AI initiative across an enterprise requires a change in employee processes, systems,
practices, org structure, and culture. Getting the stakeholder’s trust in the value of AI systems to enable this
change takes work.
Data Security, Privacy, Ethics, Compliance, and Reputation

Because the scope of a large-scale AI initiative is large, the risks associated with the initiative are also large in
terms of costs, data security, privacy, ethics, compliance, and reputation.
Build a Business Case for Large-Scale AI Initiatives

It is more challenging to build a business case for large-scale AI initiatives as compared to a smaller point solu-
tion. However, the benefits of a large-scale AI initiative are also much larger.
Factors to Consider
In a nutshell, given all these considerations, whether an organization should proceed with a small-scale PoC solu-
tion or a more significant AI transformational effort will depend on where they are in their AI journey.
➤➤ If your company is just beginning to test AI and has not identified a particular use case, then a point
PoC solution may be a good fit.
➤➤ However, if your company has already gained some experience through a previous point-based PoC
solution and has identified several use cases, as well as the right resources, skill sets, infrastructure, and
data available for a large-scale initiative, then a large-scale initiative may be the way to go, given the vast
benefits it can have for the organization.
It is recommended that you kick-start the AI transformation effort with a PoC and then follow it up with a larger-
scale AI transformation initiative.
NOTE Deciding between a smaller PoC and large-scale AI transformation depends on the
risk-taking culture and org change capability of the organization as well as the maturity of
their AI processes, practices, data, and org structure.
HANDS-ON CASE STUDY EXERCISE: ENTERPRISE AI TRANSFORMATION

DECISION: PoC VS. LARGE SCALE AI
Objective: Evaluate the challenges of embarking on an enterprise AI transformation effort,

evaluate the cost benefits, and make strategic decisions about the implementation of
AI projects.
Materials needed: Online research, pen, and paper for note-taking, PPT, or Google Slides.
Step 1: Divide the students into groups and share the case study.
Positives: You’re with a midsize retail company with some PoC experience, a risk-taking
culture, and potential transformative initiatives. Yet, you face backend integration challenges,
large dataset management, and employee resistance.
Negatives: The company also has several challenges such as integration with backend systems,
managing large datasets, and having employee resistance to change.
Step 2: The groups should assess the situation and factor in aspects such as skill sets, technol-
ogy, infrastructure integration, data management, governance, security, change management,
regulations, resource needs, and so on.
Step 3: The teams need to decide the next course of action, whether to go forward into a
large-scale AI initiative or continue with their PoCs. The teams must give a short presentation
capturing the situation, their decision, and their reasons.
Step 4: Once all the teams have presented, discuss as a class regarding common viewpoints and
differences in the approach. The individual can capture their own notes and decide how they
would want to apply it in a real-world setting.
WORKBOOK TEMPLATE: AI CHALLENGES ASSESSMENT

Download the “AI challenges assessment” template from the download section of the book (www.wiley.com/
go/EnterpriseAIintheCloud). Gain a good understanding of the potential challenges you may face when
implementing AI enterprise wide. Use it as a proactive tool for planning and to prioritize areas that need special
attention.
SUMMARY
While embarking on the AI transformative journey can be quite exhilarating, the landscape ahead can also be
pretty daunting. Without a doubt, the promises are alluring, but it is essential to approach them with both enthu-
siasm and preparation. This chapter is your first step in the journey ahead.
This chapter considered some challenges in implementing a machine learning deployment that prevents organiza-
tions from deploying them into production. Given the immense power behind an AI solution, it is also vital for
the implementation team to factor in some of the social, ethical, and humanitarian considerations when deploying
an AI solution. The next chapter discusses the requirements to implement responsible AI.
REVIEW QUESTIONS
These review questions are included at the end of each chapter to help you test your understanding of the
information. You’ll find the answers in the following section.
1. What is the main advantage that digital natives have when it comes to adopting AI?
A. They have a risk-taking culture and view failure as a steppingstone for success.
B. They have a lot more money to invest in AI.
C. They have a better brand image than traditional companies.
2. What are the challenges when deploying AI/ML solutions in a business environment? Choose all
that apply.
A. AI and ML technologies are complex to deliver value.
B. The cost of running machine learning workloads can be a barrier to deployment.
C. All business problems are easily solvable using AI/ML.
D. There is a surplus of qualified AI/ML professionals in the market.
3. What is a common data-related challenge in machine learning?
A. There’s usually an excess of high-quality data available.
B. Models don’t need data in a specific format, making the process easier.
C. Measuring the performance of models is straightforward and easy.
D. Collecting high-quality data and preparing it for machine learning can be time-consuming
and costly.
4. What is a significant challenge when managing the entire machine learning lifecycle?
A. It is easier to conduct feature engineering and select models.
B. It is easy because data preparation does not require any significant effort.
C. The complexity around data preparation and feature engineering makes it challenging.
D. None of the machine learning tasks presents any significant challenge.
5. What makes AI transformation more complex than digital transformation?
A. AI transformation requires basic IT skills and infrastructure.
B. AI transformation involves changes in how work is done and the interaction of employees with
technology.
C. AI transformation does not require a substantial amount of high-quality data.
D. AI transformation does not present any challenges related to bias, ethics, privacy, and
accountability.
ANSWER KEY
1. A
2. A, B
3. D
4. C
5. B
4
Designing AI Systems
Responsibly
The real problem is not whether machines think but whether men do.
—B.F. Skinner
This chapter grapples with one of the most pressing issues of our time: designing AI systems responsibly.
It is even more critical if you are a data scientist, project manager, AI developer, or executive involved in
implementing AI. It is not just about developing code or training models but about ensuring that what you
develop is aligned with human values, ethics, and safety.
This chapter explores the key pillars of Responsible AI, from robustness and collaboration to trustworthi-
ness and scalability. This chapter also shares the essential pillars of a Responsible AI framework, the four
key principles of Responsible AI, and some of the nuances of AI design, development, and deployment. As
shown in Figure 4.1, this chapter sets the stage for subsequent chapters for the practical deployment and
scaling of AI. Whether you have just started or are trying to refine an existing implementation, this chapter
will guide you.
THE PILLARS OF RESPONSIBLE AI

Organizations have a social and humanitarian responsibility to ensure that pre-existing prejudices and
biases do not continue to operate in the post AI-deployment world.
Responsible AI is a framework created to protect human principles such as dignity, fairness, and privacy
(see Figure 4.2). It is used to design and build AI systems that are ethical, transparent, fair, and socially
responsible. Responsible AI basically recognizes the impact that AI systems can have on humanity, includ-
ing individuals, communities, and society overall. Its goal is to minimize the negative impacts of AI while
trying to maximize the positive impacts on those involved. The implications to the design of AI systems are
that they must not only be built for robustness and reliability but also should focus on minimizing bias and
discrimination and protecting privacy.
This section covers some of those things you need to consider when building a real-life AI system. A real-
life AI system needs to be robust, collaborative, responsible, trustworthy, scalable, and human-centric.
42 ❘ CHAPTER 4 Designing AI Systems Responsibly
02
03

09
Mature AI
Operations Strategize
• Manage and Prepare
strategic change
08
Deploy and
05 Establish ethical frameworks and practices to
Monitor ensure AI designs, prioritize stakeholder
Models
values, fairness, and trustworthiness
07 throughout the system’s lifecycle
06
FIGURE 4.1: STRATEGIZE AND PREPARE: Design AI systems responsibly
Ethical AI Robust AI
Human- Collaborative
centric AI Responsible AI AI
Trustworthy
Scalable AI
AI
FIGURE 4.2: Key pillars of Responsible AI
Without these Responsible AI attributes, it will be challenging to deploy it in production because users will not
have the required trust and confidence to use the system. The following sections look at each one of those attrib-
utes and explain what they mean.
The Pillars of Responsible AI ❘ 43
Robust AI
You need to design AI systems to be robust, meaning they should be able to perform well even in situations in
which they have not been previously trained. It is especially critical for use cases such as defense, autonomous
vehicles, and robotics, where the consequences could be catastrophic if the system misbehaves.
BEST PRACTICE TIP A robust AI system should perform consistently even when the data
constantly changes, especially in an enterprise setting, where many data sources are feeding
the model. If the system produces unreliable insights, users will lose trust and confidence
when using the system. You need to test the system for any surprising attacks or situations.
Collaborative AI
The idea behind collaborative AI is to leverage the strengths of both AI and humans. While humans have creative
thinking, leadership, judgment, and emotional and social skills, robots can crunch large amounts of data and help
humans with physical labor, information gathering, speed, and routine tasks such as customer service. By combin-
ing the strengths of both humans and AI, you can build better systems to improve business processes. Figure 4.3
shows examples of this collaboration, such as chatbots, cobots, virtual assistants, and other intelligent machines.
These intelligent machines work alongside humans to make decisions and carry out tasks collaboratively.
Human
Data Crunching
Cobots Creative Thinking
Physical Labor
Virtual Machines Leadership Artificial Intelligence
Information
Chatbots Judgment
Gathering
Intelligent Emotional Skills
Speed AI and Humans
Machines Social Skills
Routine tasks Working Together
FIGURE 4.3: Collaborative AI: Enhancing human capacity with AI
The critical takeaway from collaborative AI systems is that AI will not replace humans but will complement and
enhance human decision-making capacity. While taking care of mundane, repetitive tasks, they help humans focus
on unique, high-value tasks and increase their capacity to deliver on more tasks.
Collaborative AI has created the need for humans to train AI systems so that they behave correctly, as well as
explain why AI systems behave in a certain way. For example, in the case of car accidents, you need to explain
why an accident happened or why the car failed to prevent one. It has also created the need for human supervi-
sion to ensure that AI systems work responsibly.
It’s been shown that companies that replace humans can realize only short-term gains. In contrast, companies that
use AI to improve human capacity to make decisions and carry out mundane tasks faster at scale can redesign
business processes and reap greater long-term benefits.
Use Case Example
Some examples of collaborative AI are HSBC and Danske Bank using AI to improve the speed and accuracy of
fraud detection. Another example is that of Mercedes Benz, which has replaced robots with AI-enabled cobots,
which are robots working with humans to achieve flexibility in the process. This redesigned their human-
machine collaborations.
Trustworthy AI
You need to build trustworthy systems, meaning they should behave consistently, reliably, and safely. They should
also protect the user’s well-being, safety, and privacy. The goal is to build trust between the user and the system.
You should also be able to explain to the user how the system took its actions and be accountable to the user
regarding how it concluded or made a prediction. Figure 4.4 shows the key elements of a trustworthy system.
Accuracy
Human Oversight and
Accountability
Transparency Explainability and

Interpretability
Trustworthy
AI
Fairness Privacy
Security Reliability
Safety Robustness
FIGURE 4.4: Key elements of trustworthy AI
Use Case Example
The Google search engine is an excellent example of a trustworthy system. It has algorithms to ensure that
results are shown based on the relevance of those websites to the search query. It also ranks the results and
filters all the spam and malicious websites when presenting the results. It provides transparent and explainable
results to the user regarding how it came up with those results. Google has also developed a set of ethical
guidelines and principles and has explained to the users how they collect the data, protect, and use it by being
transparent to the users. Google has built a reputation among users as a trustworthy search engine AI system.
Scalable AI
You must build AI systems that can handle large amounts of data without compromising performance during
operation. For a system to be scalable, you need to ensure not just the scalable management of data and models
but also scalability during development and deployment, as well as scalable algorithms and infrastructure (see
Figure 4.5). This section also discusses how to solve data scarcity and collection, data reusability to scale data in
the subsequent chapters on cloud platform architecture, automating operations, and data and model processing.
To address the development and deployment scalability, the book discusses using production pipelines and scal-
able architectures, including distributed cloud computing and edge devices.
Infrastructure Scalability Data Scalability

Deploy and manage AI Store and process large
systems at scale Data amounts of data
• Use cloud-based Scalability • Use distributed databases
infrastructure and and cloud-based storage
nt
caching Deployme solutions
Scala b ili ty
Algorithm Scalability lability Deployment Scalability

hm Sca
Train and deploy algorithms Algorit Deliver AI systems to users
that can handle large without compromising
amounts of data performance
lability
• Use parallel computing c ture Sca
and machine learning Infrastru • Use edge computing and
caching
frameworks
FIGURE 4.5: Scalable AI: Key considerations and case study
SCALABILITY CASE STUDY: GOOGLE
Scalability is essential when building natural language processing–based AI solutions. These

systems must handle large amounts of data without decreasing performance. Google Cloud’s
natural language processing system is an excellent example of a system that has handled this. It
uses natural language processing algorithms to analyze and extract insights from large
amounts of data. It can integrate with other Google services that businesses use for their
operations. One good use case for this is social media platforms. They can use Google Cloud
natural language to process millions of user comments in multiple languages to understand
user sentiments and use it to tailor their end-user experiences and engagement and increase
customer satisfaction without losing performance.
Human-centric AI
You need to build AI systems that are human-centric. Human-centric means that they should respect humans’
values and enable them to make decisions. For this to happen, a conscious effort needs to be made to incorporate
human values and needs in the design. You should factor in social, ethical, and human behavioral impact. These
systems must be built to be transparent, explainable, and accountable to the users. They should understand how
these systems work and be able to make informed decisions based on the system’s recommendations. Figure 4.6
shows some of the key considerations when building a human-centric AI system.
Incorporate Social, Ethical, and

Behavioral Impact
Enable Transparency and

Decision-Making Accountability
Respect Human
Adaptability
Values
HUMAN-CENTRIC
AI
FIGURE 4.6: Building human-centric AI systems with human values at the forefront
Use Case Example
A healthcare system that provides personalized treatment recommendations to patients based on their
historical medical data, genetics, and lifestyle is an excellent example of a human-centric AI system because it
ensures that the treatment is not based on a one-size-fits-all approach but is personalized based on the patient’s
unique needs and circumstances. The developers must also consider the data privacy of the user. They should
also be careful about the potential impact of making wrong recommendations. That is why it would be
necessary for the system to explain how a decision was made. This is critical because it will help the doctors
trust this system more, allowing them to make recommendations to patients much more confidently. Moreover,
you should continuously monitor the system to ensure it operates as expected. It should also adapt to the
changing patient and healthcare system requirements.
HANDS-ON ROLE-PLAYING EXERCISE: IMPLEMENTING RESPONSIBLE AI
Identify potential social, humanitarian, and ethical issues when implementing AI and proposing
ethically sound solutions. Use this understanding to explore the key design elements to ensure
the AI system is robust, collaborative, trustworthy, scalable, safe, and human-centric.
TASK DESCRIPTION TASK OWNER DELIVERABLES
Context setting: Briefly describe your AI Project manager Project details

system’s purpose, users, and current/desired document
potential societal impact.
List potential ethical challenges: List Ethics List of identified

possible prejudices, biases, privacy, security, analyst and team ethical challenges
safety, transparency, or other ethical
considerations relevant to the scenario.
Develop solution for prejudices and bias: Data scientist Strategies for
Propose methods to address biases in data minimizing biases
collection, model training, and deployment.
Develop solutions for privacy and security: Security expert Data privacy and
Design controls to ensure data privacy and security plan
compliance with relevant regulations.
Develop solutions for accountability and Compliance Accountability and

transparency: Outline ways to keep AI officer transparency plan
models transparent and accountable.
Develop solutions for safety and security: QA engineer Safety and

Develop strategies for robust testing and risk security plan
mitigation.
Consider stakeholder needs and wellness: Stakeholder Stakeholder

Explain how the AI system considers the manager consideration
needs and values of stakeholders, including document
users, employees, and the wider community.
Divide into groups: Each group represents a Facilitator Adjusted solutions

different stakeholder. Debate the proposed (team lead) based on role-play
solutions and make adjustments based on
different perspectives.
Summarize the key principles and specific Business analyst Comprehensive

actions: Include reflections on what was document detailing
learned and how it could be applied to future the approach
AI projects.
Discuss challenges and successes: As All participants Reflection

a group or individually, write about the document
challenges and successes experienced during
the exercise, and how these insights might
influence future AI projects.
HANDS-ON EXERCISE FOR STUDENTS
Step 1: Assign a student or a group of students to a topic related to Responsible AI, such as
potential ethical issues, biases, privacy concerns, or other challenges.
Step 2: Have them present a short presentation with their scenario, challenges, and proposed
solutions. Identify the common themes and unique insights and document them for applying it
to their future projects.
WORKBOOK TEMPLATE: RESPONSIBLE AI DESIGN TEMPLATE

Download the “Responsible AI Design” template from the download section of the book (www.wiley.com/
go/EnterpriseAIintheCloud). Use this template to come up with a comprehensive plan to address the
ethical and responsible design of your AI system. Treat this as a working document that keeps evolving with
the project.
SUMMARY
This chapter explained the importance of Responsible AI to ensure AI systems are designed and deployed with
ethical, social, and human-centric principles in mind. Users should be able to understand AI decisions, and the
systems should eliminate bias and prioritize human values. These systems must behave consistently and reliably
and should be scalable to large amounts of data. That means you must ensure adherence to data privacy regula-
tions, implement robust security and testing, and involve stakeholders during design.
In the next chapter, we will discuss developing the AI strategy, the roadmap, and getting alignment from various
stakeholders.
REVIEW QUESTIONS
mation ere. You’ll find the answers in the following section.
1. Transparency in AI design means
A. The models used in AI systems should not introduce bias when they are adapted to changing
conditions.
B. The data used as input to a model should be biased.
C. The users and organizations understand how an AI system operates and makes its predictions so
that they can use it appropriately.
D. It should not be possible for either humans or the programs to change the predefined model devel-
opment process.
2. Which of the following is NOT a best practice tip for Responsible AI?
A. The data used as input to a model should not be biased.
B. The organization implementing AI should focus only on the positive impacts of AI.
C. Humans or programs should not be able to change the predefined model development process.
D. The models used in an AI system should not introduce bias when they are adapted to changing
conditions.
Answer Key ❘ 49
3. What is one of the elements of trustworthy AI?

A. The system should make predictions without any explanation.
B. The system should behave consistently, reliably, and safely and protect the user’s well-being, safety,
and privacy.
C. The system should always agree with the user’s decision.
D. The system should operate independently of any data.
4. What is NOT a characteristic of a human-centric AI system?
A. Transparent and explainable to users
B. Able to operate without any human intervention
C. Accountable to users
D. Understands how the systems work and able to make informed decisions based on the system’s
recommendations
5. What does it mean for an AI system to be robust?
A. It should perform well even in situations in which they have not been previously trained.
B. It should be able to replace human workforce completely.
C. It should be able to operate without human intervention.
D. It should be cheaper than human workforce.
ANSWER KEY
1. C 3. B 5. A
2. B 4. B
Envisioning and Aligning Your
5
AI Strategy
By failing to prepare, you are preparing to fail.
—Benjamin Franklin
This chapter not only sets the tone for your AI transformation but also ensures that every step you take
from now on resonates with your company’s broader objectives.
Targeting business strategists struggling to find their AI direction, technology leaders grappling with busi-
ness and technology alignment, and AI enthusiasts eager to embed AI meaningfully into their business,
this chapter is focused on charting a clear vision for your AI initiatives and aligning it with your business
objectives. Without a clearly defined vision and business case, even the most robust AI implementation can
fail when not aligned with the business goals.
This chapter presents a methodology that clearly lays down the tasks and deliverables you can follow
across various stages of your AI implementation journey. It serves as a bridge between the groundwork
laid in Chapters 1–4 for idea generation and later chapters, teeing it up for developing the business cases in
Chapter 6 to managing strategic changes in Chapter 7 (see Figure 5.1).
By the end of this chapter, you will have a toolkit of strategies, a clear roadmap of tasks, a business case,
and a checklist of tasks vital for your AI journey.
STEP-BY-STEP METHODOLOGY FOR ENTERPRISE-WIDE AI

As you can see in Figure 5.2, you need to manage everything, from strategy to portfolio management to
architecture, infrastructure, operations, governance, trust, ethics, and security. You need to manage the
business, technical, and change management challenges in a systematic, programmatic manner.
NOTE To successfully deploy AI enterprise-wide, you need to address challenges ranging

from people to data issues to security to trust.
I know that implementing enterprise-wide AI is not an easy thing to do. However, to drive business
transformation with AI rather than solve one or two AI use cases, enterprise-wide AI is the answer, and
this book is your end-to-end guide. As shown in Figure 5.2, the chapters in the book are arranged in a
Step-by-Step Methodology for Enterprise-wide AI ❘ 51
manner that should help you manage all these challenges and implement enterprise-wide AI using a systematic
methodology.
02
03

09
Mature AI
strategic change
08
Deploy and
05 Define and articulate a clear AI vision
Monitor while ensuring alignment with enterprise
Models
goals and stakeholders
07
06
FIGURE 5.1: STRATEGIZE AND PREPARE: Envision and align
Continue Your Kickstart Your AI

AI Journey Transformation AI Strategy
and Planning
Monitor and Scale your AI AI/ML Solution
Govern Models Enterprise-wide Selection
AI Infrastructure
and Operations
Develop and Deploy Identify AI/ML Data Prep and

Models in the Cloud Use Cases Model Operations
Continue Your
AI Journey
Evaluate and Choose
Process Data
AI/ML Services
in the Cloud
Integrate Your
Select Your AI/ML
Apps and Systems Automate Your Build Algorithms and
Operations Your Frameworks
Platform
FIGURE 5.2: Steps to implement enterprise AI
The Envision Phase

Your AI journey starts with the Envision phase, which is where you start seriously thinking about how to deploy
AI within the organization.
52 ❘ CHAPTER 5 Envisioning and Aligning Your AI Strategy
Primary Objective: Define Your AI Strategy

The primary objectives of the Envision phase are to define the AI strategy, ensure it aligns with the business goals,
identify different AI initiatives, and prioritize them. I break this down into simple steps, as shown in Figure 5.3.
Envision Phase
Tasks and Deliverables
Identify Business Establish a
Define AI Vision, Obtain Leadership Prioritize AI
Transformation Measurement
Strategy Support Initiatives
Opportunities Framework
• AI vision and goals • List of potential AI • Stakeholder • Metrics framework • Prioritization

• AI strategy initiatives or use analysis identifying with KPIs to track framework to
cases outlining the the key individuals success select the most
problem, solution and their roles and • Reporting and impactful
and benefit responsibilities analytics plan to initiatives
• Initiative Impact • Business case track KPIs, report • List of prioritized
Assessment outlining the costs, and use them for initiatives with
risks, and benefits decision-making high-level
resource,
feasibility, and
communication
plan
FIGURE 5.3: Envision phase: tasks and deliverables
Steps of the Envision Phase

These steps are part of the Envision phase:
Identify opportunities: Brainstorm ways AI can Innovation team, List of potential

Identify your business transform your business product managers AI transformation
transformation across different domains, opportunities
opportunities such as technology, process, across domains
organization, and product.
Align leadership: Obtain Once potential AI initiatives or Senior leadership, Approved

your leadership’s support use cases have been identified, department heads AI initiatives,
run them by the leadership alignment
team to get their buy-in and document with
align them with measurable business outcomes
business outcomes. The
leadership team should include
those who can influence and
drive change within their
respective capability areas.
Step-by-Step Methodology for Enterprise-wide AI ❘ 53
Create a measurement Have clearly defined business Business analysts, Measurement

framework: outcome goals to demonstrate data scientists framework,
Establish your the value that AI and cloud business outcome
measurement framework technology can bring to your goals aligned
company. A measurement with strategy
framework helps ensure your
company’s business goals
are in line with the overall
business strategy.
Prioritize Initiatives: Identify the different areas that Project managers, Prioritized list
Prioritize your AI AI can impact and prioritize strategy team of AI initiatives,
initiatives across their value. impact assessment
multiple areas across areas
CASE STUDY EXAMPLE: IDENTIFYING AI INITIATIVES DURING THE

ENVISION PHASE
Say you are working for a retail company to improve its supply chain. As part of the Envision
phase, you have identified the AI strategy, brainstormed several AI initiatives, obtained
leadership buy-in, established a measurement framework, and prioritized your AI initiatives.
For example:
➤➤ Business strategy: Leverage AI to improve supply chain operations.
➤➤ AI strategy: Use appropriate algorithms to analyze supply chain data and generate insights
to optimize operations. This includes processing data, building models, and leveraging
appropriate AI tools.
➤➤ Business goals: Reduce lead times by 20 percent, increase on-time delivery to 98 percent,
and reduce transportation costs by 10 percent.
➤➤ Measurement framework: This is a set of metrics used to track progress. In this case, it
includes lead times, on-time delivery, transportation costs, inventory turnover, and order
cycle time.
➤➤ AI initiatives: These include demand forecasting for better inventory management, route
optimization to minimize transportation costs, and warehouse automation to reduce
labor costs.
The Align Phase

Now that you have identified the list of potential AI initiatives during the Envision phase, it is time to move on to
the Align phase.
Primary Objective: Build Your Business Case

During the Align phase, the primary objective is to build a business case for AI so that you can get alignment
from the stakeholders to marshal and align the resources to proceed with your AI initiative, which you identified
during the Envision phase. Figure 5.4 illustrates this process.
Align Phase
Tasks and Deliverables
Identify Initiate Change
Assess Current Create a
Remedial Management Develop a Plan
Maturity Levels Roadmap
Measures Processes
• Maturity • Remediation plan • Change • AI initiative plan at • AI initiative

assessment report to address management plan the project or roadmap with
documenting stakeholder with processes to portfolio level milestones,
current maturity concerns ensure stakeholders • Business case resources, and
level across • Process are on board development plan timelines
business, platforms, improvement • Communication, • Risk management • Resource
people plan training, and allocation plan
plan
• Gap analysis report • Tool selection stakeholder • Performance
• Financial plan
with gaps shared plan engagement plans measurement
by stakeholders plan
• Resource plan
FIGURE 5.4: Align phase: tasks and deliverables
Steps of the Align Phase

Here are the steps of the Align phase:
TASK TASK OWNERS DELIVERABLES
Assess your current maturity levels: You focus on Business analysts, Maturity assessment
identifying the current maturity level in your organization, technology leaders report, list of
across business, technology, people, process, security, and concerns and
governance. Document the concerns, limitations, or gaps limitations
shared by your stakeholders.
Identify remedial measures: By understanding the current Process Remedial action

gaps and limitations, you can put in place remedial measures managers, cross- plan, resource
to address stakeholders’ concerns, such as new processes, functional teams allocation plan
procedures, or tools. You may even identify any cross-
organizational dependencies and plan to bring in untapped
resources for this initiative.
Initiate your change management processes: You can then Change managers, Change
put in place organizational change management processes IT partners, management plan,
to adequately address any hurdles. You also talk to different training leads stakeholder support
stakeholders for their support; for example, you may have to documentation
speak with IT partners to procure the right resources or the
training curriculum lead for putting together the right set of
courses. Your goal is to ensure that all the stakeholders are
fully on board and incredibly supportive of this AI initiative.
Summary ❘ 55
TASK TASK OWNERS DELIVERABLES
Develop your plan: During this phase, you capture these Project managers, Project plan, AI
gaps in the form of a plan so that you can work toward strategy team transformation plan
launching the AI initiative in the next phase. There are two
types of plans. One is at the project level to implement the
identified AI POC as part of the Launch phase, and the other
enables enterprise AI from a transformation perspective. The
latter is known as the AI transformation plan.
Create your roadmap: Create a roadmap with clearly Strategic planners, AI initiative
defined milestones with your idea of the resources and project managers roadmap,
timelines. milestone chart
CASE STUDY EXAMPLE: BUILDING A BUSINESS CASE AND PLANNING DURING

THE ALIGN PHASE
Continuing with the retail example, you may perform the following tasks during the
Align phase:
➤➤ Change communication plan: You find out during your gap analysis that some stakehold-
ers are concerned about AI’s impact on their jobs. So you put together a change manage-
ment process that involves periodic communication about the value of the AI initiative
along with reassurances about how it benefits their jobs.
➤➤ Skills gap: You identify remedial action items to address technical skills gaps. One of those
measures is hiring new employees with AI and ML skills and working with the training
curriculum team to develop AI courses for your employees to bring them up to speed with
the technologies.
➤➤ Data issue: During the gap analysis you also realized that your company does not have all
the data required to build and train the models. You assess different options to generate
or acquire necessary data from various marketplaces. While discussing this with your IT
leads, you also learn that you need to build certain aspects of the AI platform.
➤➤ Roadmap: You consolidate all the action items into a plan and define a roadmap with
clear timelines and milestones.
➤➤ Business case: At the end of this phase, you build a strong business case to launch the pro-
ject and ensure all the impacted stakeholders are supportive and committed to the success
of this AI initiative.
WORKBOOK TEMPLATE: VISION ALIGNMENT WORKSHEET

Download the “Vision Alignment Worksheet” template from the download section of the book (www.wiley.
com/go/EnterpriseAIintheCloud). Use this template to ensure alignment between the AI goals, organiza-
tional vision, and mission.
SUMMARY
In this chapter, you took a tour of the Envision and Align phases for implementing enterprise AI across the
organization. By now, you must be aware of the challenges in implementing enterprise AI and why a detailed step-
by-step methodology is not just helpful but also essential.
In Chapter 10, you will learn about the Launch phase and in subsequent chapters, I will dive into how to scale
your AI adoption across the enterprise. If you have ever wondered how to make AI work for your company, these
four phases—Envision, Align, Launch, and Scale—must be your allies and guide on this journey. Think of these as
the signposts that guide you on your path to becoming an AI-first organization.
In particular, the essence of strategically implementing AI and aligning your AI initiatives with your business goals
cannot be underplayed. Your journey has just begun. A clear roadmap, coupled with strategic alignment, is a must
for success.
REVIEW QUESTIONS
1. Which one of the following is not a fundamental building block in an enterprise AI initiative?
A. Data
B. People
C. Business
D. Marketing
2. What is the primary objective of the Envision phase?
A. To launch a pilot to evaluate the idea
B. To define the AI strategy and prioritize AI initiatives
C. To take an iterative approach to learn from the pilot and scale it
D. To align the stakeholders for AI readiness
3. Which of the following is not a primary objective of the Align phase?
A. To assess current maturity levels
B. To identify remedial measures
C. To implement AI initiatives identified in the Envision phase
D. To initiate change management initiatives
4. What is the primary objective of the Launch phase?
A. To identify pilot projects
B. To implement pilot projects
C. To deploy a full-scale AI initiative
5. Which of the following is not part of the Scale phase?
A. To implement a robust AI/ML platform
B. To implement strong security controls
C. To establish an automated operations process to track and remedy issues
D. To implement a pilot project
ANSWER KEY
1. D 3. C 5. D
2. B 4. B
6
Developing an AI Strategy and
Portfolio
All men can see these tactics whereby I conquer, but what none can see is the strategy out of which vic-
tory is evolved.
—Sun Tzu, The Art of War
This chapter transitions from mere ideation and vision, as laid out in Chapters 3 to 5, to translate that
vision into a compelling business case. Even the most exciting AI ideas will not gain traction, the necessary
resources, or the support in your organization unless you have a robust, strategically aligned business case.
In addition to the business case, this chapter crafts the AI strategy that guides the direction, scope, and
approach for your AI initiatives, thus setting the stage for planning the pilot and building the team (see
Figure 6.1). By the end of this chapter, you will have a well-defined business case and a comprehensive AI
strategy aligned with your organizational goals.
LEVERAGING YOUR ORGANIZATIONAL CAPABILITIES

FOR COMPETITIVE ADVANTAGE
This section discusses how you can leverage nontechnical aspects such as business, governance, and
people to execute your AI transformation effort. Note that you do not have to address all these capabilities
at once. You can identify which ones you want to focus on based on your needs. You can decide based on
what you discovered during the maturity levels assessment of those areas during the Align phase.
Chapter 21 covers conducting an AI maturity assessment.
You can classify your company’s capabilities into business and technical categories. This chapter reviews
the business activities, such as developing an AI strategy in line with your business goals, managing an AI
portfolio to deploy various use cases, and marshaling the people to get involved in your AI implementation.
In Chapter 23, I help you build an AI operating model that serves as a framework to bring your strategy,
portfolio, people, platforms, and operations together to achieve your AI goals.
58 ❘ CHAPTER 6 Developing an AI Strategy and Portfolio
02
03

09
Mature AI
strategic change
08
Deploy and
05 Formulate a compelling business case and
Monitor comprehensive AI strategy to drive
Models
organizational transformation and achieve
07 competitive edge
06
FIGURE 6.1: STRATEGIZE AND PREPARE: Develop business case and AI strategy
CASE STUDY: HOW NETFLIX LEVERAGED THEIR CAPABILITIES FOR

COMPETITIVE ADVANTAGE
In Chapter 2, I talked about how a company like Netflix transformed their business by
leveraging cloud and AI technology. I discussed how technology drove process transformation,
enabling them to become more efficient and scale globally. Process transformation, in turn,
created new opportunities for organizational transformation, enabling them to become
product-centric, customer-focused, and agile. And this organizational transformation resulted
in their ability to identify new products, services, and, eventually, new business models.
To achieve business transformation and build its competitive advantage, Netflix had to lever-
age processes to deploy various resources, tangible and nontangible, from within and outside
the enterprise. Typically, these resources can be a mix of people, processes, and technology. The
ability to bring together all those resources to achieve a favorable business outcome is known
as capability.
In the case of your AI effort, you would do the same to build your AI capability to achieve a competitive advan-
tage. You will leverage a team of skilled data scientists, cloud and data engineers, software engineers, and business
teams using the Agile methodology to develop machine learning models on a complex, scalable cloud infrastruc-
ture employing leading-edge best practices.
Focus Areas to Build Your Competitive Advantage

To build your competitive advantage, you need to bring together your organization’s capabilities in these eight
major areas of focus, as shown in Figure 6.2:
➤➤ Strategy and planning
➤➤ People
➤➤ Data
Initiating Your Strategy and Plan to Kickstart Enterprise AI ❘ 59
➤➤ Platforms
➤➤ Machine learning
➤➤ Operations
➤➤ Security
➤➤ Governance
Different organizations have different strengths and weaknesses in these categories, which determines the success
rate and the path they take toward achieving their business goals. For example, a few companies may be good at
product design, which may help them design new products and drive innovation. In contrast, other companies
may be good at supply chain management, helping them deliver products quickly and cheaply. These capabili-
ties have been developed over time through investments in people, processes, and technology and have become a
competitive advantage for an organization.
Business People
Ensure AI and business strategy alignment Have the right people, skills, and
and AI and business goals alignment, and culture to build and deploy AI
develop use cases
Data Governance
Ensure data is available with high Define AI governance and risk
quality to train accurate models and management framework for ethical
generate insights to drive action and Responsible AI
Enterprise AI
Capability
AI/ML/Generative AI Platform
Develop and use high-performing Suitable and reliable platform to store
models for process automation, and process data, build, and deploy
decision-making, human language models
use cases, and innovation
Security Operations
Ensure data, model, and infrastructure Build and deploy quality models
security against unauthorized access, using high-quality data using
use, or destruction MLOps
FIGURE 6.2: Capability focus areas for enterprise AI
Driving Competitive Advantage Through AI

In the case of an AI initiative, this competitive advantage would come from how good the company is at having
its employees learn and excel at adopting AI technologies; how scalable, robust, and secure their cloud infrastruc-
ture is; how mature their DevOps and ML operations are; and how quickly they can act on these insights in real
time to drive customer experience, reduce cost, and manage risks.
From an AI transformation journey standpoint, your focus is to leverage these capabilities to build your competi-
tive advantage and achieve your business goals. How well you succeed depends on how well you engage with the
stakeholders who own these capabilities and navigate the AI transformation effort. Follow the best practices in
this book to increase your probability of success.
INITIATING YOUR STRATEGY AND PLAN TO KICKSTART

ENTERPRISE AI
The following sections review the business and people aspects. In subsequent chapters, I discuss building the
capabilities within the platform, operations, and security dimensions. See Figure 6.3.
In the context of an enterprise-wide AI initiative, below are the focus areas for you to manage effectively:
➤➤ Define your AI strategy
➤➤ Manage your portfolio
➤➤ Encourage innovation
➤➤ Manage your product lifecycle
➤➤ Develop strategic partnerships
➤➤ Monetize data
➤➤ Draw business insights
➤➤ Maximize machine learning capability
Manage Strategy Manage portfolio

Define AI strategy in line with Prioritize your AI initiatives in
company objectives line with business goals
Data Science Capability Manage Innovation

Solve complex business problems Promote a culture of innovation to
using advanced analytics and ML solicit and implement AI ideas
Business
Enablement of
Enterprise AI
Draw Business Insights Product Management
Gain real-time insights about your Deliver customer-centric AI product
business using AI capability using an agile methodology
Monetize Data Strategic Partnerships

Generate measurable business value Reduce risk, accelerate time to value, increase
from data and insights using AI quality of AI using partnerships
FIGURE 6.3: Business focus areas for strategy and planning
Leveraging the eight business aspects shown in Figure 6.3 effectively will increase the probability of success of
your AI transformation effort, enterprise-wide.
Manage Your AI Strategy, Portfolio, Innovation, Product Lifecycle,

and Partnerships
Table 6.1 shows a list of deliverables that you need as part of the Envision and Align phases.
Define Your AI Strategy to Achieve Business Outcomes

The following case study discusses an example of a company that managed its AI strategy to achieve its busi-
ness goals.
TABLE 6.1: AI Strategy and Portfolio Deliverables
CAPABILITIES DELIVERABLES DESCRIPTION
Define your AI AI strategy document A comprehensive document that captures the strengths,
strategy weaknesses, opportunities, and threats, and the AI vision
and business goals
• For example, evaluating the growth of technology
platforms, customer expectations, regulatory
requirements
Current state assessment Assessment of the company’s current capabilities and

resources related to AI
AI technology strategy To achieve the company’s business goals

• For example, using machine learning to personalize
customer experiences
Technology plan to Plan to implement technology strategy. For example:

implement AI • Adopting an agile methodology and organizing
teams around products
• Setting up an AI COE
AI initiatives A prioritized list of high-level AI initiatives, use cases, and

products
Metrics To measure the success of the AI initiatives
Manage your AI Portfolio analysis Analyze your portfolio based on short- and long-term
portfolio goals
• For example, categorize your list of AI initiatives
into high value, low value, low risk, high risk, high
complexity, and low complexity in terms of delivery
capability
AI portfolio roadmap Sequence your initiatives based on the portfolio analysis

in the previous step
Encourage An innovation Define the goals of the innovation program, including

innovation management charter how the AI solutions will be aligned with organization’s
goals
Innovation roadmap List of innovative ideas, both innovative and incremental.

These could be innovative ideas, prototypes, proof of
concepts, or even finished products and services
Metrics They measure the success of innovative ideas
Senior leadership support Critical to ensuring employee motivation levels
continues
TABLE 6.1: (continued)
CAPABILITIES DELIVERABLES DESCRIPTION
Manage your AI AI product strategy and Develop the product strategy from a customer-centric
product lifecycle roadmap point of view to solve a problem identified through
customers and data
Product backlog and Manage product requirements

product requirements
document
Product development and Helps to have a clear roadmap so all parties are clear
delivery plan about their role; helps to manage plans, resources,
milestones, and risks effectively
Develop your List of potential partners Helpful to identify companies with machine learning,
strategic data analytics, and AI skills
partnerships
Partnership agreements Includes a list of the roles and responsibilities of all the
parties involved
Evaluate partners in terms of reducing risk, accelerating

Metrics
time to value product quality, and increasing the
probability of success
CASE STUDY: HOW A SUPPLY CHAIN MANAGEMENT COMPANY MANAGED

THEIR AI STRATEGY TO ACHIEVE THEIR BUSINESS OUTCOMES
The company wanted to optimize its supply chain operations using AI. During the Envision
and Align phases, the company evaluated the competitor landscape, technology developments,
regulatory changes, and customer expectations to understand the business impact.
By looking at its business goals and the impact assessment, it decided to leverage predictive
analytics to optimize its operations and meet its long-term goals.
The company developed its technology strategy to deploy AI technology to improve its supply
chain operations and meet its business goals. As part of its strategy, it reorganized its teams
around products and value streams to become more customer-centric. To ensure a successful
implementation, it solicited feedback from many stakeholders, such as customers, employees,
leadership teams, and third-party vendors.
In the end, the company achieved its business goals using AI technology.
Figure 6.4 shows examples of business strategy, AI strategy, business goals, and AI initiatives.
Prioritize Your Portfolio

Managing the portfolio you identified for AI helps you have a holistic view of your products and services and
ensure that your AI strategy and plan meet your customers’ needs. When prioritizing your products, you should
focus on improving the customers’ end-to-end experience across all channels and business operations.
CASE STUDY: NETFLIX AND PORTFOLIO MANAGEMENT
Consider the example of Netflix. They employed portfolio management to prioritize their
products to deliver an engaging customer experience and maintain their competitive edge.
Netflix used data about their customers’ behavior to personalize their viewing experience.
They also prioritized the products that optimized their content delivery network to allow
faster streaming and reduce buffer times. They also ensured they had the necessary technical
skills, resources, and infrastructure to deliver the customer experience without compromising
on performance, security, and reliability. Using portfolio management to prioritize their AI
products based on their need to achieve strategic goals to personalize the customer experience,
operational efficiency would have helped them ensure superior performance and their ability
to deliver.
Business Al Strategy Business Goals Al Initiatives

Strategy
Q1 Q2 Q3 Q4
Increase Improving • Increase customer Develop personalized recommendation engine

customer loyalty Customer retention rate by 20%
and revenue Retention in the next 12 months.
Implement a chatbot
• Increase revenue by
15% in the next 12
months.
• Improve customer Analyze customer behavior
satisfaction by 10% in data to identify patterns and
the next 12 months. offer promotions
Increase Automating • Reduce production Develop predictive

efficiency and Manual time by 20% in the maintenance models to
reduce Processes next 12 months. reduce machine downtime.
operational • Decrease production
costs Implement computer vision systems to automate quality
costs by 10% in the
control and defect detection.
next 12 months.
• Improve product
quality by 15% in the Use machine learning algorithms to optimize production
next 12 months. schedules and reduce waste.
FIGURE 6.4: Examples of business strategy, AI strategy, business goals, and AI initiatives
Strategy and Execution Across Phases

Figures 6.5 and 6.6 are representations of how you might phase your deliverables for strategy, portfolio, data,
modeling tasks, and managing innovation. Feel free to customize this based on your organization’s needs and best
practices. Note that this can be waterfall or agile, although an agile approach is recommended.
HANDS-ON STRATEGIC PLANNING AND PROJECT MANAGEMENT EXERCISE TO

LEVERAGE ORGANIZATIONAL CAPABILITIES
Scenario: Assume you are part of an AI team trying to improve supply chain opera-
tions using AI.
Goal: Leveraging your company’s organizational nontechnical capabilities in strategy, portfo-
lio, data, modeling, and innovation.
Step 1: Define your AI strategy.

Step 2: Manage your AI portfolio.
Step 3: Encourage innovation.
Step 4: Manage your AI product lifecycle.
Step 5: Develop your strategic partnerships.
Deliverables:
Develop the following:
➤➤ AI strategy document with the business goals
➤➤ Portfolio analysis and AI portfolio roadmap
➤➤ Innovation management charter and roadmap with success metrics
➤➤ AI product strategy and roadmap along with product backlog
➤➤ List of potential AI partners and partner agreements
ENVISION ALIGN LAUNCH SCALE
01 Manage AI Strategy
STRATEGY 02 Manage AI Portfolio
03 Manage Innovation
03 Innovative 01 Potential AI 02 Prioritized AI 03.Innovative 03 Innovative

Ideas Initiatives Portfolio Ideas Ideas
FIGURE 6.5: AI strategy: phases and capabilities
ENVISION ALIGN LAUNCH SCALE
04 Manage Product Lifecycle
05 Strategic Partnerships
EXECUTION
06 Develop Data Science Capability
07 Monetize Data
04 Product 04 Product 04 Product 04 New Product

Requirements Development Launch Requirements
05 Vendor 05 New 05 Vendor 05 Contract

Evaluation Vendor Performance Renewal
Contracts Evaluation
06 Data 06 Data 06 Model Deployment
Collection Preparation and Monitoring
07 Data
Monetization Ideas
FIGURE 6.6: AI execution: phases and capabilities

Answer Key ❘ 65
WORKBOOK TEMPLATE: BUSINESS CASE AND AI STRATEGY

Download the “Business Case and AI Strategy” template from the download section of the book (www.wiley
.com/go/EnterpriseAIintheCloud). Use this template to develop an AI strategy for your company. This
template offers a business-centric view of planning an AI initiative, focusing on financial forecasts, stakeholder
engagement, and alignment with strategic business objectives. Feel free to adapt it to your organization’s
specific needs.
SUMMARY
This chapter delved into the task of translating AI ideas into actionable and impactful AI strategies. This chapter
was a step-by-step guide to craft a compelling AI strategy and a prioritized portfolio.
You also learned how to chart out a comprehensive AI strategy that provides direction and focus for your future
AI endeavors. This chapter serves as the bridge between your overarching AI strategy and execution. In the next
chapter, let’s discuss the critical components to accelerate AI adoption in your company.
REVIEW QUESTIONS
1. The ability to bring together resources to achieve a business outcome is called what?
A. Innovation
B. Achievement
C. Capability
D. Transformation
2. What is the first step in developing an AI strategy?
A. Identify your business goals.
B. Assess your current AI capabilities.
C. Develop a roadmap for AI adoption.
D. Identify potential use cases.
3. What is the purpose of managing your portfolio in the context of an AI strategy?
A. To ensure all projects have the same priority
B. To get a holistic picture of all the products and services to meet customer needs
C. To prioritize products and services based on their cost
D. To focus primarily on customer experience
4. What is the key to leveraging organizational capabilities for competitive advantage in AI projects?
A. Choosing the latest machine learning models and algorithms
B. Increasing the technical skills of your team
C. Focusing on nontechnical aspects such as business, governance, and people
D. Investing in the most expensive AI technologies
ANSWER KEY
1. C 3. B
2. A 4. C
Managing Strategic Change
7
Change is the only constant in life, but with strategic management, we navigate its tides.
—Inspired by Heraclitus
If you are an executive, team leader, change leader, or someone driving strategic change in your organiza-
tion, this chapter is for you. The world of AI, in particular generative AI, is fast-moving, and being able to
manage change is not merely a nice thing to have but a must. While earlier chapters focused on under-
standing the promise of AI, this chapter delves into how to strategically manage and accelerate that change
(see Figure 7.1).
02
03

09
Mature AI
• Manage and
strategic Prepare
change
08
Deploy and
05 Implement a structured approach to drive
Monitor AI adoption, align leadership, and
Models
accelerate strategic change initiatives
07
06
FIGURE 7.1: STRATEGIZE AND PREPARE: Manage strategic change
This chapter guides you through a three-phase approach to accelerate AI adoption, from developing an AI
acceleration charter to ensuring leadership alignment and creating a change acceleration strategy. I also tie
these steps to tangible deliverables that will serve as evidence of your progress. And also note that the steps
outlined here apply to both traditional AI and generative AI implementations.
Accelerating Your AI Adoption with Strategic Change Management ❘ 67
Whether you are trying to reposition a team or redefine an entire company, this chapter provides you with the
tools needed to drive strategic change. Remember that though every organization’s AI journey is unique, the
change management principles are universal.
ACCELERATING YOUR AI ADOPTION WITH STRATEGIC CHANGE

MANAGEMENT
Change acceleration is about accelerating the adoption of innovative technology by addressing concerns that help
you stay ahead of the competition and establish yourself as an industry leader. The fact is, despite implementing
all the measures outlined here, you may still find that the employees are not willing to adopt AI at the speed at
which you want them to. That’s why you need the change acceleration program discussed here.
The next section reviews the change acceleration toolkit shown in Figure 7.2. It brings all the previous people’s
capabilities together into a cohesive plan.
Assemble Team Align Leaders Envision
Develop An AI Acceleration Create a Change Acceleration

Ensure Leadership Alignment
Charter Strategy
Create a Communication
Analyze Value Drivers Assess Stakeholders
Strategy and Plan
Conduct a Change Impact Create an Engagement
ACTIVITIES
Establish AI Governance
Assessment Strategy and Plan
Build an Enterprise AI Conduct Readiness Create a Training Strategy and

Acceleration Team Assessment Plan
Create a Risk Management

Establish AI Initiative Goals Create a Case for Change
Strategy and Plan
Define Your Future State AI

Create a Sponsor Roadmap
Vision
Define Change Acceleration

Create an AI Sustainability Plan
Metrics
Estimate the Budget
FIGURE 7.2: AI change acceleration strategy: phases and capabilities
Phase 1: Develop an AI Acceleration Charter and Governance

Mechanisms for Your AI Initiative
The AI acceleration toolkit consists of assembling the team, aligning the leaders, and envisioning the future. The
AI acceleration charter is a vital tool that helps you assemble the team. It documents the goals, objectives, and
metrics for your AI initiative (see Figure 7.3). Let’s dive into the details.
68 ❘ CHAPTER 7 Managing Strategic Change
Estimate the Budget for AI Projects Develop AI Acceleration Charter

Develop a budget that covers all aspects Outline the goals, objectives, and key
of the AI initiative stakeholders for the AI initiative
Define Change
Acceleration Metrics Analyze Value Drivers
Establish benchmarks to measure Identify the main factors contributing
the success and impact of the Develop AI to the organization’s value, such as
AI initiative on the organization Acceleration operational efficiency and cost reduction
Charter and
Define Your Future Governance
State AI Vision Mechanisms Establish AI Governance
Articulate a vision for the future Set up procedures for
state of the organization, decision-making, issue management,
enabled by AI technology and accountability within the
AI initiative
Establish AI Initiative Goals Build an AI Change Management Team

Set clear, measurable objectives Assemble a dedicated team to handle change
for the AI initiative that align with management, communication strategies,
broader organizational goals and training requirements
FIGURE 7.3: Develop AI acceleration charter and governance mechanisms
Develop an AI Acceleration Charter

Your organization needs an AI acceleration charter that identifies the dependencies and stakeholders involved in
the AI initiative. It focuses on increasing speed, maximizing adoption, and reducing risks. Your stakeholders can
be the IT team, the business team, the data science team, and the legal team. The charter includes a review pro-
cess, stakeholder responsibilities, and metrics to measure AI activities. You can use it to gather input from various
stakeholders and establish initial relationships to achieve business outcomes.
Analyze Value Drivers: For Strategic Planning and Performance Measurement

Value driver analysis helps you identify those factors that drive value, such as operational efficiency, cost reduc-
tion, and revenue generation. You can use these drivers to assess the strategic importance of the initiatives for
planning and enable your managers to focus their activities on what is important.
Establish Governance for the AI Initiative

You should establish a governance mechanism for your AI initiative to set up the right processes for decision-
making, escalations, issue management, assigning ownership, and accountability. This provides clarity and ensures
alignment between the workstreams and the overall organizational governance. It ensures alignment between the
leadership, stakeholders, working teams, and change management teams. You must ensure alignment between
parties and governance by establishing a cadence of meetings and scrum ceremonies that synchronize with the
program-level cadence and reporting mechanisms.
BEST PRACTICE TIP Develop an AI charter, analyze value drivers, establish governance,
build an AI team, and then define the goals, vision, metrics, and budget to implement AI
successfully.
Build an Enterprise AI Acceleration Team for Enterprise AI

You should set up an AI change acceleration team in addition to a center of excellence team who together are
responsible for handling change management issues, organizational changes, role changes, communication strate-
gies, training requirements, and securing executive sponsorship. They help mitigate the risks of AI implementation
from concerns such as poor communication, lack of cross-functional collaboration, and AI implementation-
related challenges such as data handling and model building, trust, and ethics.
Establish AI Initiative Goals

You need to collaborate with various stakeholders to establish clearly defined goals so that you can monitor the
progress of this initiative and align your activities accordingly. You can use these goals to align the people, pro-
cess, and technology to achieve outcomes such as improving customer experience and driving sales.
Define Your Future-State AI Vision

You should define your future-state vision that can be achieved through AI technology and ensure it aligns with
your organization’s goals, structure, processes, culture, and technology. It will help guide your change accelera-
tion program to transform your company’s processes, people, and technology. Leverage analysis techniques such
as decision analysis, process analysis, and business capability analysis. Remember that the future-state model will
evolve as new developments happen in the AI space.
Define Change Acceleration Metrics to Measure the Success of AI Adoption

You can then define change acceleration metrics to measure the success of how people are adapting to the recent
changes in the organization resulting from the introduction of AI initiatives. You should define these early in the
project lifecycle and track them throughout the course of the initiative. You can use this to track where the organ-
ization is struggling to cope with the changes and make the necessary changes to the plan. Examples of change
acceleration metrics are user adoption rate, employee satisfaction, time to proficiency, productivity, and ROI.
Estimate the Budget for AI Activities

Develop a realistic and achievable budget for change management activities. You should include costs for the
team resources, material development, development of skills and knowledge, travel, events, software licensing
costs, hardware costs, and physical costs.
Phase 2: Ensure Leadership Alignment

Leaders play a vital role in the success of any AI initiative. It is therefore important to get their alignment at
the onset of the initiative. Leaders set the vision, motivate employees, manage stakeholders, and guide the
organization.
Ensure Leadership Alignment to Lead by Example, Take Risks, and

Motivate Employees
You must secure leadership alignment at the beginning of the AI initiative. It is a particularly crucial factor to the
success of the AI transformation effort and impacts the ability to transition to the future state. It involves stake-
holder management, leadership support, and ensuring participation in stakeholder updates.
Your leadership needs to communicate your company’s AI vision and strategy to the employees and motivate
them about this AI initiative. Figures 7.4 and 7.5 show the objectives, steps, and tasks mapped to the phases. Feel
free to customize these to your company’s needs.
Assess Stakeholders to Transform Your Workplace with Career Development and

Performance Alignment
Stakeholder assessment is the first step in the process of managing stakeholders, and it involves understanding
their needs, their concerns, and their predisposition to the AI initiative. Understanding these concerns helps you
manage their reactions as you proceed throughout the initiative. You should conduct this assessment in the initial
stages of the initiative and continuously review this throughout the AI initiative.
Envision Align
Leadership Vision Statement AI Governance Framework

Identify leadership roles and responsibilities in Identify the decision-making authority and
driving AI success outline the protocol for leadership support
Leadership Competency Model Culture of Experimentation

Develop a leadership competency model that Encourages leaders to take risks by trying new
outlines the skills and behaviors expected from use cases and challenging the status quo
the leaders
Leadership Awareness Program Leadership Development Program

Develop a leadership awareness program to Develop a leadership program to educate
educate the leaders about the potential of AI leaders about AI, data literacy, and change
and how to lead change management
FIGURE 7.4: Transform your leadership: Envision and Align phases
Launch Scale
Conduct Regular Leadership Check-in Scale Leadership Development Program

Leaders should conduct periodic check-ins to Expand leadership program to other levels
ensure alignment to AI goals and objectives of management
Metrics and Dashboards Leadership Training Program

Have metrics and dashboards in place to Roll out leadership training program to other
measure the current progress in AI adoption levels of management
Leadership By Example
Encourage leaders to lead by example
and display behaviors to support AI adoption
FIGURE 7.5: Transform your leadership: Launch and Scale phases
Transforming your workplace involves assessing your current capabilities in your workforce and producing strat-
egies to fill the gaps. Figure 7.6 shows the steps and deliverables.
Envision Align Launch Scale
CAREER DEVELOPMENT ROLE TRANSITION PLAN IMPLEMENT CAREER REVIEW CHANGE

PLAN DEVELOPMENT PLAN MANAGEMENT AND
Develop a change
PERFORMANCE
Create a career development management plan that outlines In the launch phase, you MANAGEMENT PLAN
path for employees involved in how employees will transition should implement the change
AI roles to facilitate their into the new role comfortably management plan and During the scale phase, you
career progression performance management will refine the change
ACTIVITIES
PERFORMANCE system that was decided in the management plan and the
MANAGEMENT SYSTEM align phase performance management
system as new developments
Develop a performance happen in AI
TALENT PROGRAM
management system that
aligns with the new roles that Develop a talent program that
arise in an AI implementation BRAND COMMUNICATION
includes hiring third-party
STRATEGY
consultant, vendors, as well as
in-house training program Develop brand communication
SKILLS GAP ASSESSMENT
strategy and promote your
Conduct a skills gap analysis brand
that outlines the gaps in skill
sets and areas where
upskilling is required
FIGURE 7.6: Transform your workspace: tasks and deliverables
IBM CASE STUDY: IBM’S NEW COLLAR JOB INITIATIVE: A MODEL FOR
WORKFORCE TRANSFORMATION
IBM’s new collar job initiative is a notable example of workforce transformation. IBM, with
this program, is trying to fill the skills and experience gap in the technology space by helping
those who cannot get a traditional four-year degree to get into these high-tech jobs. With the
advent of modern technologies such as AI, ML, and advanced analytics, there is a need for
people who can maintain and operate these systems. This does not require a four-year degree; a
nontraditional vocational program suffices. IBM created a variety of training and education
initiatives, such as online courses, apprenticeships, and internship programs.
Since its launch in 2017, this has found tremendous success. According to IBM, in 2018, about
50 percent of their new employees came from those who had nontraditional education, and
most of them came from this new collar initiative. More importantly, it has helped IBM diver-
sify its workforce by including people from all kinds of educational backgrounds.
Overall, this has been a new trend in the industry, and this is something that you need to consider as you look for
ways to augment your workforce.
Conduct a Change Impact Assessment

Conduct a change impact assessment to assess how the AI initiative will impact various stakeholders and produce
a training, communication, and change management plan. This helps identify the stakeholders impacted and
decide on the channels of communication your team should use, along with the necessary governance structures,
decision points, and policy reviews. It helps you decide how and when the stakeholders will be involved and pro-
duce specific plans to address those impacts.
Figure 7.7 sums up your journey to get buy-in from the leadership and stakeholders to create a case for change.
Create a
Case for
Organizational
Change
Readiness
Assessment
Change Impact
Assessment
Create a
Assess Stakeholders persuasive
document
Gauge the highlighting the
organization’s benefits of the AI
Launch and Scale
readiness for initiative to gain
Evaluate how the change and
AI initiative will buy-in from
prepare strategies stakeholders
Envision and Align affect different to fill in any gaps
Understand the stakeholders and
needs, concerns, develop a plan to
and attitudes of manage these
Initiate the AI stakeholders impacts
initiative and
scale it up per
Set a clear vision organizational
and align leaders needs
toward the AI
initiative
FIGURE 7.7: Ensuring leadership alignment for AI, including generative AI initiatives
Conduct Organizational Readiness Assessment

Conducting an organizational readiness assessment will help you understand the organization’s willingness,
desire, and ability to cope with change and put measures in place to address any gaps, such as skills training,
communication, sponsorship, and changes to the change acceleration.
Create a Case for Change

The case for change is a document that outlines the benefits that the organization will gain through this AI initia-
tive. You should create it after the stakeholder assessment and use it in various communications to get buy-in
from the employees. It should be clear, concise, and persuasive to the audience and be documented with data.
Phase 3: Create a Change Acceleration Strategy

The change acceleration strategy is a plan that includes the different steps the organization must take to manage
changes resulting from the AI initiative. There should be a built-in collaboration with different stakeholders from
HR, business functions, third-party vendors, leadership, sponsors, and employees. You should create this at the
beginning of the initiative and update it continuously. You should ensure minimum disruption and maximum
adoption. This helps ensure stakeholder alignment, cultural transformation, and business continuity.
Create a Communication Strategy and a Plan

A communication strategy and plan is a document that outlines different stakeholders to be communicated, the
channels to communicate, and the frequency with which to communicate. It should include a communication,
training, and risk management plan.
BEST PRACTICE TIP Create a comprehensive change acceleration strategy, communica-

tion plan, and training plan to ensure successful AI implementation by keeping stakeholders
informed.
Create an Engagement Strategy and a Plan

You should create this engagement strategy and plan after creating the change acceleration strategy and the com-
munication strategy. It helps engage with the stakeholders throughout this initiative when the situation demands.
It helps you manage the pace at which the different stakeholder groups must go through.
Create a Training Strategy and a Plan to Develop AI Skills

The training strategy and plan outline different training programs the employees will have to take as part of the
AI initiative. It should include a learning needs assessment, a training curriculum, and a training delivery plan.
The success of your AI program directly depends on the skills and capabilities of your employees in these leading-
edge technologies. Figure 7.8 shows the steps you can take to build the skills of your employees.
TRAINING PLAN AI/ML SKILLS BASELINE AI/ML TRAINING TRAINING EVALUATION

Develop a training plan to Develop a baseline for AI/ML Provide ongoing AI/ML training Continuously evaluate the
educate employees on AI/ML proficiency in the organization courses such as online, training programs and
and cloud fundamentals, along instructor-led, and self-paced necessary changes
with use cases, benefits, and courses
terminology MENTORING PROGRAMS
ACTIVITIES
Create mentoring programs, CERTIFICATIONS
job rotation, shadowing, and Encourage employees to take
COMMUNICATION PLAN HANDS-ON LABS
coaching so they can learn charge of their own learning
Communicate the business from experts Allow employees to create and get certified in AI
case for AI adoption and how Establish a community of sandboxes to get their hands
it aligns with the business practice such as AI COE to dirty with AI/ML tools
goals communicate AI/ML standards
KNOWLEDGE SHARING
REWARDS PROGRAM
Encourage employees to
SKILLS ASSESSMENT ONBOARDING PROCESS Set up a rewards program for share their knowledge with
Conduct a skills gap Make AI/ML training part of the employees who use AI/ML others and become AI
assessment for AI/ML and process skills in their work evangelists
cloud skills
FIGURE 7.8: AI strategy: phases and capabilities
When you implement these recommendations effectively, you will see a significant increase in the number of
employees using AI/ML and generative AI in their work, leading to increased efficiency, improved customer sat-
isfaction, and revenue. You will notice that your culture has become innovative and agile, with employees taking
ownership of their learning and adopting AI in their workplace. All this will eventually help you become competi-
tive in a highly competitive landscape and even position you as an industry leader.
Create a Risk Management Strategy and a Plan

This communication strategy and plan are essential to ensure that the risks associated with your initiative are
carefully identified, evaluated, and mitigated. The plan helps you identify the action items to mitigate the risks
and to assign ownership to tasks to ensure they complete the deliverables on time, with quality and within the
planned cost.
Create a Sponsor Roadmap

A sponsor roadmap is a document that outlines how to engage and support the sponsors throughout the AI initia-
tive. It holds sponsors accountable to provide leadership support and helps them maintain consistency in their
messaging to achieve business outcomes.
Create an AI Sustainability Plan

You should create a sustainability plan to ensure that everyone is adopting AI continuously, even after you have
completed the initial AI initiative. The plan will continue to evolve as the organizational structure and recent
technology developments in the AI space evolve. You will typically create it once the team has gained sufficient
experience with AI transformation.
HANDS-ON CHANGE ACCELERATION WORKSHOP
This is a type of hands-on workshop focused on promoting discussions and collaboration. It

enables critical thinking and promotes the sharing of different real-world perspectives among
the participants.
Goal: In this exercise, you explore different change acceleration strategies such as leadership
alignment, workforce transformation, and risk management. The goal is to develop a concrete
AI acceleration charter, align leadership, and create a change acceleration strategy.
Step 1: Divide the team into groups representing different functions in your company.
Step 2: Choose a hypothetical scenario such as building a chatbot, predictive maintenance in
manufacturing, or sales forecasting.
Step 3: Ask the teams to execute the steps in the following table using the methodology
in the book.
Step 4: Have the teams present their findings, promote discussions, and summarize the
common themes, ideas, and challenges.
Step 5: Explore how it can be used in your organization.
TASK NAME DESCRIPTION TASK DELIVERABLE

OWNER
Develop an AI Outline the goals, principles, AI strategy A written AI

acceleration and key activities that will lead acceleration
charter define your AI initiative. charter
Ensure leadership Create a presentation or Executive A presentation

alignment document that clearly sponsor or document
conveys the importance of AI and feedback
to your organization and how
it aligns with current business
objectives.
continues
continued
TASK NAME DESCRIPTION TASK DELIVERABLE

OWNER
Create a change Develop a detailed strategy Change A change

acceleration that outlines how your management acceleration
strategy organization will accelerate team strategy
the adoption of AI, including document
key milestones, potential
challenges, solutions, and
a timeline.
Implement Engage in a role-playing Training and Insights and

role-playing exercise where participants development lessons learned
take on different
organizational roles and work
through implementing the AI
acceleration strategy.
WORKBOOK TEMPLATE: STRATEGIC CHANGE MANAGEMENT PLAN

Download the “Strategic Change Management Plan” template from the download section of the book (www
.wiley.com/go/EnterpriseAIintheCloud). The strategic change management plan guides your organiza-
tion through the important phases of implementing an AI initiative, with an emphasis on managing change.
SUMMARY
This chapter discussed the critical components of accelerating AI adoption in your company. By building the AI
acceleration charter, achieving leadership alignment, and developing a systematic AI acceleration strategy, you can
galvanize your employees around AI to achieve business outcomes.
You learned that you could achieve organizational transformation by adopting a systematic, programmatic
approach to business transformation and people empowerment.
In the next chapter, let’s discuss finding the right AI/ML use cases for your company.
REVIEW QUESTIONS
1. Which of the following is not included in Phase 1 of the change acceleration program?
A. Analyze value drivers
B. Establish governance for the AI initiative
C. Develop an AI acceleration charter
D. Ensure leadership alignment
2. What is a crucial step in securing leadership alignment at the beginning of an AI initiative?

A. Conducting a change impact assessment immediately
B. Ignoring stakeholder concerns
C. Communicating the company’s AI vision and strategy to the employees
D. Excluding leadership from stakeholder updates
3. Which of the following is not a step in creating a change acceleration strategy for an AI initiative?
A. Conducting a change impact assessment
B. Creating a communication strategy and plan
C. Creating an engagement strategy and plan
D. Developing AI skills for employees
4. Which of the following components is not involved in ensuring proper organizational alignment for
AI adoption?
A. Aligning business strategies with technology strategies
B. Aligning business strategy with organizational structure
C. Implementing an AI system without considering cultural alignment
D. Prioritizing AI projects appropriately
ANSWER KEY
1. D 3. D
2. C 4. C
PART III
Planning and Launching a Pilot
Project
This part covers all the challenges and tasks centered on planning and launching a pilot project, including
identifying use cases for your project, evaluating appropriate platforms and services, and launching the
actual project.
8
Identifying Use Cases for Your
AI/ML Project
The only way to do great work is to love what you do.
—Steve Jobs
In the next few chapters, let’s discuss planning and launching a pilot (see Figure 8.1). This chapter focuses
on identifying the right use cases. Having addressed challenges, approached design responsibly, and charted
the AI strategy and vision, it is time to identify the right use case for your AI/ML project.
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Your AI Govern Your
journey Team
09 Plan and Launch

ENTERPRISE AI • Identify Use Cases for Your AI/ML Project
Evolve and Setup 04 • Evaluate AI/ML Platforms And Services
Mature AI JOURNEY MAP Infrastructure
and Manage • Launch Your AI Pilot
Operations
08 Scale and Process Data Objective

Transform and Modeling
AI
Deploy and 05 Systematically identify and prioritize AI/ML use cases
Monitor aligned with business objectives, addressing pain
Models points and capitalizing on industry trends
07
06
FIGURE 8.1: PLAN AND LAUNCH: Identify Use Cases for Your AI/ML & Gen AI Project
80 ❘ CHAPTER 8 Identifying Use Cases for Your AI/ML Project
Whether you are a business executive looking to solve business problems or a developer trying to develop state-
of-the-art applications, this chapter is for you. Selecting the right use case is more than just a decision—it’s the
compass that sets your AI journey in the right direction. It will not only help your team solve the business prob-
lem by choosing the suitable set of technologies, but it will also lead to increased operational efficiency, reduced
costs, and even new business opportunities for a competitive edge.
You also learn to prioritize your use cases, which will help you focus all your energies on the most pressing
problems and maximize business impact. It’s your path to positive stakeholder results and a higher probability of
success. Let’s get started.
TIP Prioritize your use cases that have the highest business impact to reduce costs, increase
efficiency, and gain a competitive advantage.
THE USE CASE IDENTIFICATION PROCESS FLOW

Figure 8.2 shows a process flow you can use to identify use cases for your company. It starts with identifying the
business problems and pain points that you want to solve.
Decide on the
business Review Conduct a
objectives applications PoC to Conduct
and success in various Prioritize the choose the Develop a vendor
metrics industries use cases right model project plan evaluation
Document Research Identify the Assess data Design the Build versus
your industry right use availability solution buy
business trends case for the and quality architecture
problem or business
pain points problem
FIGURE 8.2: Use case identification process flow
Educate Everyone as to How AI/ML Can Solve Business Problems

Defining the business problem is one of the first steps in your use case identification process. To do that, the
implementation team should understand the various business problems that need to be solved. A significant part
of this chapter is all about that, and I cover that in greater detail in the following few sections.
Consider the business problems your company is facing as it relates to customer service, fraud detection, risk
assessments and predictions, supply chain optimization, predictive maintenance, quality control, marketing, and
product development. All of these are candidates for AI/ML.
The Use Case Identification Process Flow ❘ 81
Remember that machine learning helps you identify patterns in your dataset to predict outcomes. This feature/
capability can be used to improve existing processes while also enabling new opportunities. Processes that are
well suited to AI/ML are those
➤➤ Where decisions are based on data
➤➤ Where decisions happen repeatedly, either thousands or millions of times
➤➤ That use software to automate
➤➤ That are slow
Define Your Business Objectives

As shown in Figure 8.3, you should start by identifying the business objectives you want to aim for based on the
business problem you are trying to solve. Examples of business objectives are increasing customer satisfaction,
revenue, efficiency, reducing costs, and enhancing productivity. Examples of pain points include customer churn,
high number of claims/returns, equipment failures, managing inventory, and so on.
STAGE 1 STAGE 2 STAGE 3 STAGE 4
Education Pain Points Business Value Value Realization
Identify Business
AI/ML Benefits Quantify Potential Measure Business
Objectives/ Celebrate!
Education Business Value Value
Pain Points
Spread awareness and Outline the key Estimate the potential Establish metrics like
understanding of the business objectives or business value that ROI, payback period,
potential benefits of pain points that AI/ML could be realized by or net present value to
AI/ML and Gen AI and Gen AI implementing AI/ML measure the actual
across the organization could address & Gen AI solutions business value gained
from AI/ML & Gen AI
implementation
FIGURE 8.3: Defining business objectives for your AI initiative
TIP Start by educating everyone about the benefits of AI/ML and Gen AI. Define the
business objectives and quantify the potential business value to measure success.
Based on these business objectives/pain points, you can quantify the business value that can be realized upon
completion of your use case. Some examples of business value are cost savings, revenue growth, improved
accuracy, automation, better customer experience, and faster decision-making. You should be able to measure the
business value using ROI, payback period, or net present value.
Identify the Pain Points

The next step is identifying the pain points that prevent you from achieving your business objectives. Here are
some examples:
PAIN POINT CATEGORY DESCRIPTION BENEFIT OF ADDRESSING EXAMPLE

WITH AI
Inefficient or You can automate Improves efficiency by A logistics company

slow processes and optimize reducing the overall optimizing its supply
specific processes time and effort to chain routes and delivery
in the company. complete tasks. schedules.
Customer experience Using customer Helps increase customer It’s a common use
improvement data to personalize engagement, satisfaction, case to personalize
recommendations. and loyalty. marketing campaigns
and targeting ads.
Complex data analysis Automating Provides insights to A financial services

analysis of large support decision-making. company that analyzes
data datasets. market data to make
investment decisions.
Fraud detection Analyzing You can notify customers of Many companies across
transactional potential fraud. multiple industries
data to detect analyze transactional
unusual patterns. data to detect
fraudulent activity.
New product Analyzing customer Increases customer Identifying

development feedback to identify satisfaction and sales. opportunities for new
opportunities. product features.
Risk assessment AI can assess risks Proactively manage Assessing risks

in various areas. potential risks. proactively such as
customer default on a
loan, product recall risks,
or natural disaster risks.
Pain points can be manual processes, redundant processes, complex processes, data quality issues, and other inef-
ficiencies in your process.
Start with Root-Cause Analysis

Once you have a good idea of your pain points, you can start the root-cause analysis (RCA) to get to the origin
of the problem. Knowing the root cause will help you shortlist the proper use case. For example, suppose you
find out that customer satisfaction could be better. In that case, your root-cause analysis might point you to poor
website design, lack of personalization, slow speed (response times), customer support issues, and so on.
TIP Identify the pain points that hinder your business, and conduct root-cause analysis to
shortlist your use cases.
Identify the Success Metrics

Once you have identified the root cause, you can choose the proper success criteria. For example, in the previ-
ous scenario, your success criteria could be increasing the speed/response time of your website, adding person-
alization, or improving your website design. You then define your success metrics around customer satisfaction,
increased revenue, or reduction in manual effort.
Success metrics help you measure your AI solution’s effectiveness and ensure it aligns with your business objec-
tive. See Figure 8.4 for some examples of success metrics, such as the following:
➤➤ Cost savings: Cost savings is the apparent metric you will choose when automating business processes,
eliminating errors, or improving process efficiency. For example, a company trying to optimize its supply
chain operations could reduce inventory costs and shipping times.
➤➤ Revenue growth: This is used when you are improving the customer experience, which results in
increased sales and revenue.
➤➤ Risk reduction: This metric is used for use cases such as fraud detection, cybersecurity, and compliance.
➤➤ Customer satisfaction: This metric can be used in use cases where you can provide better customer ser-
vice, resolve customer issues quicker, and provide personalized customer experiences. For example, a
retail company can use AI to provide chatbots that can answer questions 24/7.
➤➤ Employee productivity: AI can automate repetitive employee tasks, reduce errors, and provide real-
time insights. For example, AI can be used to generate leads, thus allowing employees to focus on
closing deals.
Customer Satisfaction
For use cases that improve customer
service, resolve customer issues quicker,
and provide personalized customer
experiences
Revenue Growth Risk Reduction

Useful when improving the Used for use cases such as
customer experience, leading fraud detection,
to increased sales and cybersecurity, and
revenue compliance
Cost Savings
Employee
Useful when
Productivity
automating business
For repetitive
processes,
Define employee tasks,
eliminating errors, or
improving process Success reduce errors, and
Metrics provide real-time
efficiency
insights
FIGURE 8.4: Define success metrics for your AI initiative
TIP Define success metrics that align with your business objectives and use cases to measure
the effectiveness of your AI solution.
Explore the Latest Industry Trends

You also need to assess market trends as part of your use case identification process. I’ll talk about the latest AI
developments in later sections of this chapter. Consider attending other conferences and discussing with other
industry experts to learn more about the latest developments in AI, which is a very fast-moving space. Knowing
the latest trends in the AI space will help you understand the status of the industry and assess where it is going
and therefore decide your use cases accordingly. For example, suppose you’re working in the manufacturing
industry. In that case, you find out that supply chain optimization is prevalent. Therefore, you can also adopt sup-
ply chain optimization use cases, such as using predictive analysis to optimize your inventory management.
It can also give you awareness of the latest potential roadblocks that other companies face so you can also be
prepared. For example, you may find that companies are facing issues with data privacy and regulations. It can
help you be proactive and manage your own risks accordingly.
Review AI Applications in Various Industries

You should also review the applications of AI in other industries. By reviewing these applications, you can learn
from the latest AI industry trends, other industry applications, and other best practices.
ACTION PURPOSE EXAMPLE RESOURCES
Review AI Understanding the breadth of AI Understand how AI is Industry reports,

applications applications across industries to applied in healthcare, case studies
gain insights on how AI is used finance, retail, etc.
Research Gaining insights on other Understand why certain Industry forums,

best practices successful AI strategies and strategies are adopted online communities,
methodologies by investigating in specific industries Kaggle.com,
other best practices AI/ML forums
Learn Preempting and preparing Learn about risk factors Case studies,
from pitfalls for potential challenges from healthcare industry industry reports
by understanding pitfalls applications
faced by others
TIP Look to AI industry trends and other AI applications across other industries to learn
from best practices and avoid potential pitfalls.
HAND-ON EXERCISE 1: STRATEGIC AI/ML USE CASE IDENTIFICATION
This is a strategic planning exercise that involves a combination of research, brainstorming,

analysis, and strategic decision-making thus making it a valuable strategic planning exercise.
Goal: The goal of this exercise is to understand your company’s business objectives, pain
points, and suitable use cases to develop an AI/ML application to address those pain points.
TASK INSTRUCTIONS DELIVERABLES EXAMPLES
Form Groups Get into groups Groups formed Team 1, Team 2,

Break into teams to of three to four Team 3.
collaborate. people.
Educate Everyone Discuss different A list of Predicting customer

The team business problems potential churn, automating
understands the related to customer business responses to
potential of AI/ML service, risk problems common customer
in solving business assessment, supply that can be queries, detecting
problems. chain optimization, addressed by fraudulent
etc. AI/ML transactions, etc.
Define Business Spend 10 to Clearly Increase customer

Objectives 15 minutes to defined and satisfaction score by
identify clear, quantifiable 20 percent
Identify the
business objectives quantifiable business Reduce operating
business objectives objectives costs by 15 percent
that your AI/ML
project should aim to aim for. Increase revenue by
for. 30 percent, etc.
Identify Pain Take 5 to 10 minutes A list of Pain points like high

Points to analyze your identified pain customer churn
Identify issues that business processes points in rate, slow response
are hindering you and identify two to your business times to customer
from achieving your three pain points processes queries, high error
business objectives. such as inefficient rate in transaction
processes, processing, etc.
poor customer
experience,
complex data
analysis issues, etc.
Conduct Root- Conduct a Detailed report If customer churn

cause analysis thorough analysis on the root rate is high, the root
Determine the to understand causes of the cause could be poor
root cause of your the underlying identified customer service,
identified business issues causing the business lack of personalized
problems. identified business problems customer experience,
problems. etc.
continues
continued
Identify Identify three Clearly defined Success metrics

Success Metrics to five success success metrics could be “Reduce
metrics that align for your AI/ML customer churn rate
Define metrics that
by 10 percent”
can measure the with your business project
success of your AI/ objectives and the “Increase customer
ML project. chosen use case,
satisfaction score by
20 percent”
like cost savings,
revenue growth, risk “Increase revenue by
30 percent,” etc.
reduction, customer
satisfaction, etc.
Explore Latest Attend conferences, Knowledge Attending AI/ML

Industry Trends engage in of the latest conferences like
Stay updated with discussions with trends in the AI NeurIPS, ICML, ACL,
the latest trends in industry experts, industry etc., and reading AI/
the AI industry. and read up on the ML research papers
latest developments and articles.
in AI.
Review AI Review how other Insights Reviewing how

Applications in industries are documented healthcare industry is
Various Industries leveraging AI and from AI using AI for disease
Learn from the AI learn from their applications in diagnosis, how
implementations in best practices and other industries finance industry is
other industries. pitfalls. using AI for fraud
detection, etc.
Ideate Use Cases Spend five minutes List of use Predictive

Come up with AI/ ideating use cases cases maintenance,
ML and Gen AI per pain point. personalized
solutions. recommendations.
Choose As a team, choose One use case Predictive

One Use Case one use case to selected maintenance.
Select one use case move forward with.
to focus on.
Map Use Case to Refer to the “Use Chosen use Choosing a use
Business Problem Cases to Choose case that case like “Predictive
Choose a suitable From” section and aligns with Maintenance” if your
use case that can select a use case your business business problem
address your busi- that aligns with your problem is high equipment
ness problem. business problem. downtime.
Prioritizing Your Use Cases ❘ 87
continued
Prepare Prepare a five- Presentation

Presentation minute presentation slides
Summarize use case on use case.
findings.
Present Use Cases Take turns Use case Team 1 presents

Share use cases presenting use presentations predictive
with other teams. cases. maintenance
use case, Team
2 presents
personalized
recommendations,
and so on.
Map the Use Case to the Business Problem

Refer to the later section “Use Cases to Choose From” to choose your list of use cases.
PRIORITIZING YOUR USE CASES

The next step is to prioritize your list of use cases. For this, you must develop a set of impact and feasibility crite-
ria. The following sections outline these steps.
Define the Impact Criteria

Clearly define the criteria for impact by discussing with your stakeholders. Some of these could be as follows:
➤➤ Revenue impact
➤➤ Cost impact
➤➤ Efficiency impact
➤➤ Customer satisfaction
Define the Feasibility Criteria

To evaluate your feasibility, consider factors such as these:
➤➤ Complexity of required technology
➤➤ Cost
➤➤ Availability of skill sets
➤➤ Availability of data
➤➤ Platform readiness
➤➤ Time to implement
➤➤ Risks, if any
As part of your feasibility analysis, you need to consult various stakeholders across your organization from busi-
ness and technology, to identify their concerns and address them appropriately, as well as get buy-in.
Assess the Impact

Using the impact criteria, you can do a cost-benefit analysis or estimate the potential ROI.
Assess the Feasibility

Using the feasibility criteria, you can then evaluate the feasibility of the use cases.
Prioritize the Use Cases

Once you have assessed the impact and the feasibility, you can then create a matrix using these two criteria to
prioritize the use cases, as shown in Figure 8.5.
Bu
sis
sin
on ly
iti na
Hy ess
gn t A
pe Pro High Business Value

co en
r-P ce
Business Value (High to Low)
er ss High Feasibility
Re m
so Au
e nti
na to
ag Se
liz m
at at
Im ee
io io
n
oy
n Low Business Value

pl
High Feasibility
Em
Use Case
Prioritization
M
High Business Value

D fa
an
ise ctu
Low Feasibility
e
u
nc
as rin
ts na
e g
bo inte
Pr Au
t
ed to
a a
Ch M Low Business Value
i c t
it ve
io ati
Low Feasibility
n on
ic
m
ed
Pr
Feasibility (High to Low)
FIGURE 8.5: Business value and feasibility analysis to prioritize use cases
As shown in Figure 8.5, you can classify the use cases based on the business value and feasibility. Your business
value criteria can include factors such as cost reduction, revenue growth, and customer satisfaction, and feasibility
criteria can include factors such as technical feasibility and organizational readiness.
Prioritizing Your Use Cases ❘ 89
➤➤ High business value and high feasibility use cases: Your priority use cases will be those that deliver high
business value and are easier to implement. Predictive maintenance in manufacturing can prevent equip-
ment downtime and extend the life of machinery. The business value is very high, and with the necessary
sensor data and appropriate ML models, it is often easier to implement. Similarly, chatbots are easier to
implement and can have a significant impact on customer service and efficiency.
➤➤ High business value and low feasibility use cases: Under this category fall use cases such as predicting
diseases and full automation of manufacturing processes. While these use cases can save millions of lives
and save costs, they can be challenging to implement due to technology limitations.
➤➤ Low business value and high feasibility use cases: Use cases such as employee sentiment analysis to ana-
lyze employee feedback may be easier to implement but may have little impact on the business. Similarly,
employing AI to tag images automatically may be easy to implement but have a low business impact.
➤➤ Low business value and low feasibility use cases: Hyper-personalization of advertisements may appear
very attractive, but the cost of implementing it may make it highly prohibitive due to data privacy and
technical requirements. Similarly, complete automation of all business processes may appear very attrac-
tive but may be of limited value for a small business and technically and financially prohibitive.
Review and Refine the Criteria

You can then review the criteria used for assessing your impact and feasibility and continue to refine it along with
the ranking criteria, working with your stakeholders to finalize your priority of use cases.
TIP Prioritize your use cases based on their potential impact and feasibility.
Choose the Right Model

Once you have prioritized your use cases, the next step is to identify the right model for your use case. You have
to continue this process iteratively until you reach a point where you’ve finalized your success criteria, feasibility
prioritization, and so on.
To do this, you must consult with many organizational stakeholders to define your use case concept. An essential
part of choosing a use case is being able to justify the value of implementing it.
HANDS-ON USE CASE: PRIORITIZATION WORKSHOP
This is a use case prioritization workshop that involves collaborating with various stakeholders
to compare different use cases and arrive at a decision.
DESCRIPTION HANDS-O N TASK DELIVERABLES EXAMPLE
Form Groups if Get into groups of Groups formed Team 1, Team 2,

Needed three to four people Team 3
Break into teams to
collaborate.
continues
continued
Brainstorm use cases. In groups, come up List of two to

Generate potential with two to three use three use case
use case ideas. case ideas ideas
Define impact Discuss with Clearly defined Improve customer

criteria. stakeholders and impact criteria satisfaction by 30
Brainstorm the define metrics like percent
potential impacts of revenue impact, cost Reduce operating
implementing the use impact, efficiency costs by 20
case. impact, customer percent, etc.
satisfaction, etc.
Define Feasibility Discuss with Clearly defined Implementation

Criteria stakeholders and feasibility cost should not
Identify the define feasibility criteria exceed $100,000
feasibility factors of factors like complexity Should be able to
implementing the use of technology, cost, be implemented
case. availability of skill within 6 months,
sets, data availability, etc.
platform readiness,
etc.
Assess the Impact Based on the defined Detailed Potential revenue

Evaluate the potential impact criteria, carry analysis of increase of
impact of the use out a cost-benefit the potential $500,000
case. Brainstorm analysis or estimate impact Estimated
factors like revenue, the potential ROI. cost savings of
costs, efficiency. $200,000, etc.
Assess the Feasibility Based on the defined Detailed Implementation

Evaluate the feasibility criteria, assessment of cost will be
feasibility of the use assess the feasibility of the feasibility $80,000
case. Brainstorm implementing the use Can be
factors like case. implemented
technology, skills, within four
data. months, etc.
Prioritize Use Cases Create a matrix Ranked list of Use case A: High
Based on the impact using the impact and use cases impact, high
and feasibility feasibility criteria to feasibility
assessments, rank the use cases. Use case B: High
prioritize the use impact, low
cases. feasibility, etc.
Use Cases to Choose From ❘ 91
Review and Refine Consult with Refined criteria Updated impact

Criteria stakeholders to refine and adjusted criteria to
Review the criteria the ranking criteria ranking of use include customer
used for assessing and adjust the ranking cases retention rate
impact and feasibility of use cases as Updated ranking
and refine as necessary. of use cases
necessary. based on refined
criteria, etc.
Choose the Right Consult with data Chosen ML Chose a decision

Model scientists or ML model for the tree model for use
Identify the experts to choose the use case case A
appropriate model right model for the Chose a
for the use case. use case. convolutional
neural network for
use case B, etc.
Iterate Till Conduct multiple Finalized use Finalized use

Finalization iterations of case, success case A with
Repeat the process discussions with criteria, success criteria
until the use case, stakeholders, refining feasibility of increasing
success criteria, criteria and rankings, assessment, customer
feasibility, and choosing models, etc. prioritization, satisfaction by 30
prioritization are and chosen percent, feasibility
finalized. model assessment
showing
implementation
within four
months, high-
priority ranking,
and decision tree
model chosen,
etc.
Present Priorities Justify your Presentation of

Share and discuss prioritization. priorities
prioritized list.
USE CASES TO CHOOSE FROM

Listed in this section are close to 80 use cases to choose from across multiple industries and business processes.
You can expect more use cases as the technology evolves.
AI Use Cases for DevOps

AI for DevOps leverages ML to innovate your DevOps process by solving security, defects management, resource
allocation, and concurrency issues. Some examples include predictive analytics, defects management, continuous
delivery, and resource management.
Success Stories: Microsoft implemented ML in DevOps to detect bugs, and Google used it for Chrome testing.
Tools: Amazon CodeGuru Reviewer for code quality, Amazon DevOps Guru for application performance and
scalability, Azure DevOps, Azure Pipelines for CI/CD, Google Cloud Build for CI/CD.
TIP Use AI for DevOps to automate, optimize, and improve the DevOps process.
AI for Healthcare and Life Sciences

Healthcare and life science organizations face physician burnout, complex health data, and rising healthcare costs.
The benefits of AI in healthcare include personalized treatments, better disease diagnosis, and higher efficiency.
Some use cases include medical imaging, drug discovery, personalized treatments, forecast diseases, and clini-
cal trials.
Success Stories: The University of San Francisco California used data analysis to identify sepsis early; IBM
Watson Health identified cancers through medical image analysis; and startup Atomwise developed an AI tool to
treat Ebola.
Tools: Amazon HealthLake, Amazon Transcribe Medical, Amazon Comprehend Medical, IBM Watson Health,
Google Cloud Healthcare API, Microsoft Healthcare Bot.
TIP Ensure data privacy and compliance with regulations like HIPAA to build trust and
safeguard patient information.
AI Enabled Contact Center Use Cases

Traditional contact centers are helping automate and streamline processes as well as improve overall customer
satisfaction. Some examples include virtual agents, speech analytics, predictive analytics, call routing, and cus-
tomer churn.
Success Stories: HR Block and Capital One optimized costs with chatbots; Intuit’s analytics enhanced support for
11,000 agents; and Wix’s sentiment analytics boosted visibility from 12 to 100 percent.
Tools: Amazon Connect can be used for NLP, automated speech recognition, and machine learning, Google
DialogFlow, IBM Watson Assistant.
BEST PRACTICE TIP AI should enhance customer experience, streamline processes, and
complement rather than replace human agents to maintain a personal touch.
Business Metrics Analysis

Machine learning can be used to analyze large amounts of business data to track business metrics to learn from
the past, current, and planned data. Benefits include being able to analyze large volumes of data, more easily
identify trends and patterns, forecast future performance, detect anomalies, determine root causes, and adapt to
changing conditions.
Sales, marketing, and CX insights can help identify growth and strategy opportunities; IT monitoring can help
manage disruptions; inventory and workforce planning can optimize resources; and AI for Data Analytics (AIDA)
enhances existing analytics with ML.
Success Stories: Foxconn boosted forecast accuracy by 8 percent with AWS Forecast, saving approximately
$500,000. Digitata used Amazon Lookout for Metrics and saved 7.5 percent in sales revenue in hours
instead of a day.
Tools: Amazon’s Business Metrics Analysis ML solution helps analyze business metrics dynamically to changing
business conditions, Azure ML, and Google BigQuery.
TIP Use AI to analyze business metrics, adapt to changing business conditions, and identify
new growth opportunities.
Content Moderation
To create high-quality content in large quantities, you need scalable platforms and extensive resources, which
the cloud provides. AI has been facilitating personalized content creation at scales beyond human capacity, with
machine learning that produced varied content. AI-enabled content creation allows for the following:
➤➤ Personalized content
➤➤ Automated generation
➤➤ Machine learning models
➤➤ Massive scale
Success Stories: News media companies such as the Associated Press, the Washington Post, and Forbes have used
content generation tools.
Tools: GANs, neural language processing and computer vision tools, ChatGPT, Claude, WriteSonic, Perplexity,
BARD, and Copy.ai.
TIP Use AI to generate more personalized and engaging content at scale.
AI for Financial Services

AI can be used in financial services to improve core business processes such as fraud detection and claims process-
ing and to improve customer experiences by sending personalized offers. They can be used in small startups as
well as large enterprises. Some use cases include risk management, customer service, investment management,
personalized finance, loan underwriting, trading, and insurance claims.
Success Stories: JP Morgan uses the COIN platform for document validation; BlackRock uses Aladdin to recom-
mend tailored investment decisions; and PayPal’s system detects real-time transaction frauds to minimize losses.
Tools: IBM Watson Financial Services, Alteryx, and Darktrace.
TIP Financial services companies can use AI for applications such as risk management, per-
sonalized finance, and trading.
Cybersecurity
Cybersecurity is the practice of protecting systems, networks, and data from external hacking, unauthorized
access, disclosure, modification, use, or destruction of systems. It involves the use of systems, processes, and poli-
cies to protect assets from cyber-threats such as viruses, malware, hacking, phishing, and ransomware attacks.
Cybersecurity ensures business continuity to prevent downtime and aids in compliance with regulations such
as GDPR. It uses tools like IBM’s Qradar for network security and Symantec’s software for AI-driven endpoint
protection.
Success Stories: FICO uses AI for fraud detection; Splunk uses SOAR to address security incidents; Firefly’s Helix
detects malicious activity; and Dell uses Cylance for cybersecurity.
Tools: You can use natural language processing to detect malware, filter emails, and identify fraudulent trans-
actions, while deep learning can detect malicious images, analyze videos for unauthorized access, and recog-
nize scams.
TIP Use AI to improve the security of systems, networks, and data by proactively identifying
and preventing cyber-attacks.
Digital Twinning
The digital twin is a virtual replica of a physical system or object. As shown in Figure 8.6, you can use that replica
to simulate system behavior and identify any potential issues or improve the performance of those systems. They
are created by gathering the data from multiple sources or sensors and then using that data to create a virtual
model of that system.
FIGURE 8.6: Digital twin of jet airplane engine

Source: chesky/Adobe Stock
Success Stories: Rolls Royce, GE Healthcare, Johnson Controls, Siemens, and ENGIE utilized digital twins to
optimize performance, reduce equipment failures, and improve energy efficiency in their respective industries.
Tools: AWS IOT twin maker, Google Cloud IOT, Core and Microsoft’s Azure IOT hub can be used to create digi-
tal twins. Combined with these twins, you can use machine learning models to detect patterns for further action.
TIP Use digital twins to improve the efficiency, safety, and performance of physical systems
by creating a virtual replica and simulating behavior to identify issues and optimize opera-
tions.
Identity Verification
Identity verification is essential for customers creating new accounts or accessing different accounts, systems, and
information. Identity verification is used in the healthcare industry to access patient records, in the financial ser-
vices industry to access or open new accounts, and in the government sector to access services based on authenti-
cation. You can verify identities through document and biometric authentication methods like passports or facial
recognition or by using personal information for authentication.
Success Stories: Goldman Sachs uses Google Cloud Vision for verifying transactions; Gamestatix uses Azure’s
facial recognition for user access; and HealthTab utilizes Amazon Rekognition to authenticate medical profes-
sionals to access patient records.
Tools: You can use Azure cognitive services and Active Directory for identity verifications, Amazon Rekogni-
tion for facial and text recognition. Google provides Google Cloud Identity Platform and Google Firebase
Authentication.
TIP Use identity verification to protect individuals and organizations from fraud and iden-
tity theft.
Intelligent Document Processing

Intelligent document processing (IDP) can automate manual processes to extract valuable information from
various documents, such as claims, invoices, receipts, contracts, forms, and so on. IDP uses machine learning algo-
rithms to automatically extract this data using natural language processing to classify the data and then verify,
validate, and integrate it with various backend systems for further processing.
Benefits include accelerated document processing, reduced manual work, improved data quality, and faster
insights and decisions.
Success Stories: KPMG utilizes Amazon Textract for cost-effective document processing; HSBC and Johnson &
Johnson use Google’s Document AI for tasks like invoice automation using OCR and NLP. Swiss Re and Chevron
use Microsoft’s Azure Form Recognizer to minimize manual data entry.
Tools: Google Cloud Vision API analyzes images for data extraction, while Amazon Textract processes documents
and Azure Cognitive Services offers various APIs for IDP tasks.
TIP Use IDP to automate manual processes and extract valuable information from docu-
ments; this frees employees’ valuable time.
Intelligent Search
Intelligent search is a type of search that uses AI to understand the meaning of the search term and the context
behind that search and to rank the search results based on relevance. Figure 8.7 shows how by understanding the
context and the meaning of the search queries, intelligent search improves the search accuracy, efficiency, and user
engagement. It’s used in ecommerce, customer service, and knowledge management.
Improved Accuracy
By understanding the meaning
and the context behind the
search term, the system can
provide more accurate search results.
Improved Efficiency
By providing the search results
ranked by relevance, it involves
Intelligent
less interference from the users
Search to get the results that they want
and thus improving efficiency.
Engaging User Experience

By providing the user with
relevant search results, they’re
able to engage the user and
enhance their satisfaction.
FIGURE 8.7: The benefits of intelligent search
Success Stories: Coursera uses Amazon machine learning to provide the right recommendations to their students
on their platform.
Tools: You can use Amazon Kendra for intelligent search; Microsoft Graph for relational data; and IBM Watson
Health to analyze medical data for doctor recommendations. AWS Comprehend can help with text analysis using
NLP, and you can use unsupervised machine learning and NLP techniques for relevant result retrieval and contex-
tual understanding of search terms.
TIP Use intelligent search to provide more relevant and accurate search results, save users’
time, and avoid frustration.
Machine Translation
Machine translation involves translating from one language to another using AI and ML technologies.
Success Stories: It is used in education, government, and healthcare to communicate between people speaking
different languages. It helps with enhanced learning, collaboration, government services, travel, safety, and better
healthcare through better communication.
Tools: Machine translation uses AI for text or speech conversion, while AWS Translate, Microsoft Azure Transla-
tor, and Google Cloud Translate all provide translation services.
TIP Use machine translation to translate text or speech from one language to another,
leading to greater collaboration, learning, and safety.
Media Intelligence
The demand for media in the form of images, video, and audio is increasing at an unprecedented rate in industries
across education, entertainment, and advertisement. Companies must engage with their customers through con-
tent, and a lot of manual work is required to label, tag, review, and come up with headlines, summaries, and so
on. By adding machine learning to content workflows, you can increase these assets’ lifetime value, reduce costs,
and increase speed in delivering content.
Some of the use cases include audience analysis, market research, public relations, and social media monitoring.
Benefits include faster content processing, reduced costs, increased lifetime value, and higher quality experiences.
Success Stories: Netflix employs machine learning for personalized content streaming, while the New York Times
uses media intelligence to curate engaging topics for its readers.
Tools: AWS Elemental Media Insight, Amazon Rekognition, Google Cloud Language, Microsoft Cognitive Ser-
vices, and Azure HDInsight enable sentiment analysis, data extraction, and content pattern analysis from media
content such as video, audio, and text.
TIP Use media intelligence to analyze media content and gain insights into audience demo-
graphics, trends, and brand perception.
ML Modernization
ML modernization involves updating existing machine learning models and pipelines to adopt new technologies
and tools and make them better by improving their efficiency accuracy and performance (see Figure 8.8).
Outdated ML
Models
Modernized Upgrading ML Modernization

ML Models Algorithm • Improved accuracy
• Reduced bias
• Increased scalability
• Faster model development
Machine • Improved performance
Reducing
Model
Learning Incorporating
Modernization New
Development
Technologies
Time Framework
Technologies for ML
Modernization
• AWS SageMaker
• Google AI Platform
Scaling • Azure Machine Learning
Retraining
Model Platform
Models
Development
Preventing
Bias
FIGURE 8.8: Machine learning modernization framework, benefits, and technologies

Some of the focus areas of ML modernization are upgrading the algorithm, incorporating new technologies,
retraining, preventing bias, scaling model development, and reducing model development time.
Success Stories: Intuit uses AWS SageMaker for accurate fraud detection in financial transactions. Airbus utilizes
the Google AI platform to enhance airplane wing design, achieving cost savings and better fuel efficiency.
Tools: AWS offers AWS SageMaker, which is a comprehensive ML platform with prebuilt models and AWS inte-
grations. Google Cloud Auto ML is useful for those with limited ML knowledge, whereas Google AI Platform is
for custom model development with Google Cloud integrations. Microsoft’s Azure Machine Learning Platform is
an all-encompassing platform integrated with Azure storage and serverless computing functions.
TIP Use ML modernization to keep your machine learning models and pipelines up-to-date
with the latest technologies and tools, improving their accuracy, performance, and efficiency.
ML-Powered Personalization
ML-powered personalization involves analyzing user demographic and behavioral data to predict their interests
and provide them with personalized product services and content. It is used in recommendation engines, search
engine results, social media posts, and ecommerce products.
Success Stories: Yamaha Corporation boosted sales by 30 percent using Amazon Personalize. YouTube enhanced
user engagement by 60 percent with Google Cloud Recommendations API, and Air France raised sales by 13
percent by adopting Azure Machine Learning and Azure Personalizer.
Tools: Using machine learning, natural language processing, and computer vision, services like Amazon Personal-
ize, Amazon Forecast, Google Cloud Auto ML, and Microsoft Azure Learning offer personalized content and user
experiences.
TIP Use ML algorithms to predict user interests and provide personalized product services
and content, such as recommendation engines, search engines, and social media.
Computer Vision
Computer vision involves the ability of computers to interpret and understand visual data from the world around
them such as images, videos, and other visual data to identify objects, track movements, and analyze scenes.
Figure 8.9 illustrates the various applications of computer vision, use cases, and the technologies that can be used
to implement it.
Computer vision can be used to detect objects, understand scenes, recognize actions in videos, identify diabetic
retinopathy for proactive screening, and spot defects in manufacturing.
Success Stories: Amazon uses computer vision in its Go store to monitor customer behavior and automatically
adds products to carts.
Tools: Amazon uses Rekognition to analyze images and videos and SageMaker for building models. Google
provides Google Cloud Vision and GAN and CNNs for image analysis. Azure Cognitive Services provides similar
capabilities.
TIP Use computer vision in applications for object detection, scene understanding, facial
recognition, action recognition, and more.
Computer
Vision
Technologies
Use Cases
(Track behaviors, understand)
Object Scene Action Diabetic Defect Amazon Rekognition

Detection Understanding Recognition Retinopathy Detection AWS preferences
SageMaker
Azure Cognitive Services
Google Cloud Vision
GANs
CNN
FIGURE 8.9: Applications and technologies behind computer vision
Personal Protective Equipment

Personal protective equipment involves using machine learning algorithms to detect if a person is wearing protec-
tive equipment such as masks, helmets, and gloves. It became especially important during COVID-19.
Computer vision is used in the healthcare, manufacturing, and construction industries to ensure workers are
wearing protective equipment such as goggles, masks, and gloves.
Success Stories: Honeywell used computer vision to enforce PPE compliance during COVID-19.
Tools: Google Cloud Vision API can be used for image analysis. Similarly, Azure Cognitive Services is useful for
computer vision use cases.
TIP Use machine learning for personal protective equipment detection to ensure workers are
wearing the proper safety gear in manufacturing, healthcare, and construction industries.
Generative AI
Generative AI refers to a type of machine learning where it can generate new types of data like the data it was
trained on. It can be used to generate new content such as images, speech, natural language, and data augmenta-
tion. This feature makes generating new product descriptions, news articles, and other types of content possible.
You can also use it to generate new types of videos and images, which is helpful in marketing.
Generative AI can keep users engaged through marketing campaigns and by creating new fashion designs in
the fashion industry. It has been making waves ever since ChatGPT, discussed in more detail in Chapter 24,
was released in November 2022. I include a few use cases to employ generative AI, but note that this is a very
fast-moving space, and new use cases are continuously emerging. Figure 8.10 shows the different types of data
supported currently by generative AI.
Here are text-related generative AI use cases:
USE CASE DESCRIPTION EXAMPLE
Content creation You can create content such as blogs, News agencies create news articles
articles, and reports based on keywords or sports reports.
or questions.
Drafting emails AI suggests full sentences or even Gmail Smart Compose feature
paragraphs to draft an email.
Advertising content Creating engaging promotional content Product descriptions on

for products and services. ecommerce websites
Creative writing Writing stories, scripts, or poetry based Scripts for movies or commercials
on user prompts.
Personalized Personalized recommendations based on Recommend news articles based on

recommendations user history or preferences. reader browsing behavior
Language translation Translating between languages. Google Translate
Sentiment analysis Identifying sentiments in a text. Analyzing customer comments or

reviews on websites
Screening résumés Assessing résumés based on Screening job applications

certain criteria. for companies
Text summarization Create summaries for large pieces Summarizing large articles
of content. and reports
Chatbots Generate human-like responses in a Chatbots used for customer service,

conversation. sales, HR, etc.
Here are code-related use cases:
Generate code Generate code based on Microsoft’s CoPilot and AWS Code Whisperer.
keywords or requirements
Detect bugs Identify bugs DeepCode can detect bugs in static code.
Refactoring Suggest improvement in code Codota now known as Tabnine suggests code
improvements.
Review code Assist in code review process Code Climate provides automated code review
by suggesting improvements or for test coverage and maintainability.
identifying issues
Code Generate DeepAffects Code Summarizer generates

documentation documentation for code summaries for code blocks.
Unit testing Generate unit test cases Diffblue Cover generates unit tests using AI.
Code Text
Models can be used to s Models like GPT-4 generate
le
generate code, debug, and ecu texts based on user
review code ol Text prompts
M
Data
Data Images
Im
Models can generate DALL-E 2 and Stable
ag
synthetic datasets for Data Types
es
Diffusion generate images
training other models Supported from text
by
Generative
AI
Co
Molecules Audio
Audio
ed
Models can generate novel Generative models can generate
molecular structures with synthetic audio such as music and
desired chemical properties speech, used for voice assistants,
TTS (Text to Speech) systems, and
3D o music composition
de
3D Vi Video
Models can generate 3D shapes and Generative AI models can generate
objects like cars, furniture from text synthetic video, which can be used to create
descriptions deepfakes and movie production from text
prompts
FIGURE 8.10: Data types currently supported by generative AI
Here are some image-related generative AI use cases:
Image synthesis Generates new images DALL-E by OpenAI can create images from text
resembling the original images descriptions.
Style transfer Applies the style of one image DeepArt can recreate photos in the artistic style
to another of any painting.
Image Improves the image quality by Let’s Enhance can increase the image resolution
enhancement increasing resolution, removing without losing quality.
noise, etc.
Image Identifies individual objects Google’s DeepLab uses AI for semantic image
segmentation within an image segmentation.
Image colorization Adds color to black-and-white DeOldify colorizes black and white images
images or videos and videos.
Image captioning Generates a textual description Google’s Cloud Vision API provides a text
of an image description for an image.
Facial recognition Identifies or verifies a person in Face++ can do facial recognition in security
a digital image or a video frame systems, payment gateways, etc.
Object detection Detects instances of Amazon Rekognition can identify objects, people,
objects in images text, scenes, and activities in images.
continues
(continued)
Image Converts low-resolution images Enhanced Super-Resolution Generative

super-resolution to high-resolution Adversarial Networks (ESRGAN) used for
enhancing the quality of low-res images.
Augmented reality Integrates digital information Snapchat’s lenses can overlay digital content onto
with the user’s environment the physical world.
in real time
Here are some audio-related use cases:
Speech synthesis Generates human-like speech Google Text-to-Speech and Amazon Polly convert
text into lifelike speech.
Music composition Generates new music OpenAI’s MuseNet composes new music in
compositions various styles.
Voice cloning Replicates a person’s voice after ElevenLabs and Resemble AI’s voice cloning
learning from a small sample technology can create unique voices for virtual
assistants or voiceover.
Speech Enhances the quality of speech Krisp, an AI-based app, can mute background
enhancement signals by reducing noise noise in any communication app.
Sound Recognizes and categorizes Google’s AudioSet, a dataset of sound

classification audio events events, can be used to train models for sound
classification.
Speech-to-text Converts spoken language into Google’s Cloud Speech-to-Text converts audio to
written text text for transcription services.
Audio Identifies unique components of Shazam uses AI to recognize music tracks by

fingerprinting a sound for recognition their sound.
Here are some video-related use cases:
Deepfake Synthesizes human-like videos DeepFaceLab is an open-source software that

generation from images and other videos creates deepfakes.
Video synthesis Generates new video Nvidia’s GANs can create videos of virtual
content based on specific environments, such as cityscapes.
instructions or data
Video colorization Adds color to black-and- DeOldify is an AI model for colorizing and
white videos restoring old images and videos.
Super resolution Increases the resolution of Video Enhance AI by Topaz Labs improves video
video content quality using AI.
Motion transfer Transfers the motion of one First Order Motion Model for Image Animation
person in a video to another enables animating new faces in images.
person in a different video
Video compression Compresses videos without H.266/VVC is a video codec standard enabling
losing quality high-quality video compression.
Predictive video Predicts the next frames in a PredRNN is a recurrent network for predictive
generation video sequence video generation.
Video inpainting Fills in missing or removed parts Free-Form Video Inpainting with 3D Gated
of a video Convolution and Temporal PatchGAN restores
missing video parts.
Finally, here are some 3D use cases:
3D object generation Creates 3D models from NVIDIA’s GANs can generate 3D models of
descriptions or data objects from 2D images.
Virtual reality Generates immersive VR Artomatix can generate texture maps for
content creation environments virtual reality.
3D animation Automates 3D animation DeepMotion does procedural character

animation in games.
Predictive Predicts the 3D shape of a AlphaFold, a system from DeepMind, predicts

3D modeling protein or a molecule the 3D structure of proteins.
3D reconstruction Creates 3D models Pix3D, a dataset and method by researchers at

from 2D images MIT and ETH Zurich, generates 3D models from
2D images.
3D printing Predicts the results of Autodesk’s Project Dreamcatcher uses

3D printing generative design for optimized 3D printing.
Generative design Suggests design alternatives Autodesk’s Generative Design tool provides
for designers design alternatives for maximum efficiency.
WORKBOOK TEMPLATE: USE CASE IDENTIFICATION SHEET

Download the “Use Case Identification Sheet” template from the download section of the book (www.wiley
.com/go/EnterpriseAIintheCloud). Use this template to help your team brainstorm, evaluate, and prior-
itize potential AI use-cases. Use this to make informed decisions considering factors such as business impact, cost,
and technical complexity.
SUMMARY
This chapter provided a detailed review of how to find the right AI and ML use cases for your company. It began
by exploring the process flow for identifying use cases, starting from educating people about the potential of AI/
ML and Gen AI, and how it can solve business problems. It covered defining business objectives, identifying pain
points, conducting a root-cause analysis, identifying the success criteria, and exploring the latest AI industry trends.
Armed with this information, you can identify your use cases, prioritize them, choose the suitable model, and
finalize them through an iterative process.
In the next chapter, let’s discuss cloud-based AI/ML services offered by different cloud providers to bring your AI/
ML use cases to life.
REVIEW QUESTIONS
1. Identify the type of processes that are ready for AI/ML.
A. Automated processes
B. Processes that depend on data and happen repeatedly
C. Processes that are easier to automate by software and are fast
D. All the above
2. The purpose behind conducting a root-cause analysis is
A. To identify the pain points that are hindering your business
B. To quantify the business value upon completion of the use case
C. To measure the success of your use case using ROI, payback period, or NPV technique
D. To understand the root cause to identify the proper use case
3. Choose the success metric that would result in cost savings.
A. Improved customer satisfaction
B. Reduced errors
C. Increased revenue growth
4. What is the primary purpose of mapping the use case to a business problem?
A. To assess the potential impact of an AI solution
B. To ensure the use case aligns with the business objectives
C. To identify potential pitfalls in the solution
D. All the above
5. Why is it important to iterate until you finalize the use case?
A. To get a solid understanding of the problem being solved
B. To get feedback from the stakeholders
Answer Key ❘ 105
C. To proactively identify risks and challenges

D. All the above
6. Which of the following can be used to improve code quality by automating code reviews during
application development
A. Amazon CodeGuru Reviewer
B. Amazon HealthLake
C. Amazon DevOps Guru
D. Amazon Transcribe Medical
7. What benefit can AI provide in log analysis for DevOps?
A. Identify patterns that could turn into issues proactively
B. Increase the speed and reliability of software deployment
C. Automate testing for bug identification
D. Optimize resource allocation based on usage
8. What is document verification?
A. Using personal information such as driver licenses to verify identity
B. Using documents such as passports, driver licenses, ID cards, biometric verification using finger-
prints, facial recognition, or voice recognition to verify identity
C. Using facial recognition, emotional analysis, and text recognition for verification
D. Using image and video analysis that uses computer vision technology for verification
9. What is intelligent document processing (IDP)?
A. The process of extracting valuable information from various documents using machine learning
algorithms
B. The process of creating new documents using artificial intelligence and machine learning
C. The process of verifying personal information using documents such as passports, driver licenses
D. The process of automating manual tasks using artificial intelligence and machine learning
10. Which of the following is a use case for media intelligence?
A. Upgrading media intelligence models
B. Analyzing financial intelligence
C. Tracking mentions on social media
D. Predicting user interests
ANSWER KEY
1. B 5. D 9. A
2. D 6. A 10. C
3. B 7. A
4. B 8. B
Evaluating AI/ML Platforms and
9
Services
Information is the oil of the 21st Century, and analytics is the combustion engine.
—Peter Sondergaard
Welcome to the exciting world of evaluating AI/ML platforms and services. Whether you are a data scien-
tist, business analyst, developer, or decision-maker for AI/ML implementation, this chapter will help you
choose the right platform and services.
Choosing the right platform and services is pivotal to the success of your AI/ML as well as Gen AI imple-
mentation. This chapter covers the various AI and ML services that the cloud service providers such as
AWS, Microsoft, and Google provide. The idea is to leverage what’s available to get your project kick-
started. Now is an excellent time to look at the different portfolios of AI/ML services and map them to the
business problems you are trying to solve. This includes evaluating specialized services for healthcare and
industry solutions so that you can embark on a path that resonates with your business goals and regulatory
landscape.
It isn’t just about choosing the right tool; it is about aligning with the fastest time to market, handling high-
intensity workloads, maintaining cost-effectiveness, and staying at the cutting edge of technology. Once you
have identified the right use cases, you need the right set of tools, setting the stage to launch your pilot, the
focus of the next chapter (see Figure 9.1).
Given the plethora of AI/ML platforms and services, each with its own features and offerings, it can be
quite overwhelming. This chapter guides you through a comprehensive evaluation process and presents the
choices from the three major cloud providers to help you choose the right platform and services.
This chapter can help you to turn your possibilities into reality. Let’s get started.
Benefits and Factors to Consider When Choosing an AI/ML Service ❘ 107
9 StepsThin Arrows Power Point Diagram

02
01 Strategize Plan and
and Prepare Launch Pilot
03
Continue Build and
Your AI Govern Your
journey Team
09 Plan and Launch
ENTERPRISE AI • Identify Use Cases for Your AI/ML Project
Evolve and Setup 04 • Evaluate AI/ML Platforms & Services
Mature AI
JOURNEY MAP Infrastructure
• Launch Your AI Pilot
and Manage
Operations
08 Objective
Conduct a comprehensive assessment of leading
AI Deploy and 05
Monitor
AI/ML platforms and services to determine the
Models most suitable solutions for your enterprise's
07 specific use cases and business goals
06
FIGURE 9.1: PLAN AND LAUNCH: Evaluate AI/ML Platforms & Services
BENEFITS AND FACTORS TO CONSIDER WHEN CHOOSING AN

AI/ML SERVICE
This section covers the various factors you need to consider when choosing an AI/ML service.
Benefits of Using Cloud AI/ML Services

Cloud service providers provide several AI/ML services that offer a wide range of capabilities and functionali-
ties that you need to build your AI solutions, especially at scale in production. They provide access to several
sophisticated algorithms along with pretrained models that can save you a lot of time and money during the
model-building process. They also offer several development and deployment tools that accelerate the develop-
ment process. See the following table for more details:
BENEFIT DESCRIPTION
Faster time to market You can speed up your implementation time by leveraging these services because
you do not have to build these algorithms from scratch. Instead, you can focus
on solving the specifics of your complex business problems. These services come
as part of preconfigured environments, which allows for faster prototyping and
experimentation, faster time to market, and more cost-effective implementation.
Tip: Focus on solving complex business problems instead of building algorithms
from scratch.
continues
108 ❘ CHAPTER 9 Evaluating AI/ML Platforms and Services
(continued)
BENEFIT DESCRIPTION
Intensive workloads These AI services are designed to handle high computational demands and data
processing workloads that are typical of enterprise AI. They provide a scalable
infrastructure along with high-performance capabilities for these AI services to be
developed, trained, and deployed efficiently.
By leveraging the scalability, reliability, security, and performance offered by these AI
services, you can address the common challenges faced by typical machine learning
projects easily. The providers offer large GPUs and TPUs that are optimized for large
ML workloads and reduce training and inference times.
Tip: Use these services to handle large data volumes and spikes in processing
demands by dynamically allocating cloud computing resources based on need.
Cost- By leveraging these services that are hosted in the cloud, you do not have to invest
effective solutions up front or build in-house expertise. Thus, you save time and money.
By leveraging the cloud provider’s economy of scale, the pay-as-you-go pricing
model, you can avoid up-front infrastructure costs, reduce operating costs, and
implement cost-effective AI solutions.
Tip: Monitor the usage of cloud resources to effectively manage costs.
Collaboration and These services allow sharing of models, code repositories, and version control
integration systems. Additionally, since they can easily be integrated with other cloud services,
databases, and APIs, it is possible to leverage additional capabilities and data
sources to meet the needs of the use cases.
Better security These AI services can also be compliant with industry standards and regulatory
compliance requirements and data privacy laws. You can implement security measures such as
data encryption and access control.
Rich capabilities These services come with capabilities such as natural language processing, computer
vision, speech recognition, and recommendation engines in the form of pretrained
models and APIs.
Tip: Integrate these capabilities into business applications, thus empowering your
company to deliver personalized experiences, automate processes, and leverage
valuable business insights that drive customer-centric actions.
Better tool support AI services include popular frameworks such as TensorFlow, PyTorch, and AutoML
platforms that simplify and speed up the model development process.
Up-to-date with Cloud providers also invest a lot in innovation and keep releasing new services and
innovation updates to existing services, which will help you stay up-to-date and benefit from the
latest research and development.
Factors to Consider When Choosing an AI/ML Service

This section covers some of the factors to consider when choosing an AI/ML as well as Gen AI services.
KEY FACTORS WHEN CHOOSING DESCRIPTION

AN AI/ML SERVICE
Identify your business problem Before choosing an AI service, you need to first identify
the business problem you are trying to solve, including the
business objectives, the desired outcomes, and the impact
expected on your business operations, customer experience,
and decisions.
Explore AI/ML service capabilities Explore the various capabilities offered by these AI services,
such as natural language processing, computer vision, and
speech recognition, and ensure that they meet your use case.
Evaluate performance and scalability needs Evaluate the performance and scalability needs of your ML
solution and ensure that your service can handle them. Factor
in training and inference times, processing speeds, and
dataset processing needs.
Tip: Consider training and inference times, processing
speeds, and dataset processing needs.
Assess development and integration needs Assess the availability of software development kits and the
ability to integrate with your company’s backend systems,
applications, databases, and processes.
Assess model development tools Assess the tools available for model development, such as
SageMaker, PyTorch, AutoML, TensorFlow, and scikit-learn.
Evaluate service’s security capabilities Evaluate your security needs and ensure that the provider can
meet your needs.
Evaluate service pricing Evaluate the pricing models and ensure that they meet your
budget expectations.
Tip: Factor in costs for training, data storage, and API calls to
avoid future surprises.
Evaluate vendor support Evaluate vendor support factors such as their support SLAs,
their responsiveness, their uptime guarantees, and their track
record with you.
Identify industry regulations If your company has any industry-specific regulatory

requirements, factor them in as well.
Keep up with innovations Make necessary allowances for the growth of your use case,
such as increasing data volumes, innovation, and changing
business conditions.
HANDS-ON AI/ML SERVICE EVALUATION EXERCISE
This is a hands-on exercise that needs the team to participate, collaborate, research, and arrive
at a solution.
The goal is helping the team identify a business problem, evaluate various AI/ML services
available from different cloud providers to address that problem, and then develop an imple-
mentation plan for the chosen solution.
At the end of this exercise, the team should be aware of the process involved when selecting an
AI/ML service and consider various business and technical factors.
STEPS TASK DELIVERABLES RESOURCES TIPS/EXAMPLES

OWNERS
Identify a Project Develop a Company Include both

business manager document data, business short-term
problem outlining objectives and long-term
the business from company objectives
problem, strategy, and
objectives, other company
planned sources
outcomes, and including
impacts leadership
Assess AI/ Data Understand List of AI/ML Factor in the

ML service scientists and develop a services, online specific problem
capabilities report with the research requirements
details of the
capabilities of
each service
Evaluate Data Estimate and Project Plan for future

performance engineers prepare a requirements, growth
and document with data volume
scalability performance estimates
needs and scalability
requirements
Assess Software Document SDKs, backend Ensure the

development developers capturing the system selected AI/ML
and development documentation service can be
integration and integrated with
needs integration existing systems
needs
STEPS TASK DELIVERABLES RESOURCES TIPS/EXAMPLES

OWNERS
Assess Data Document Documentation Factor in the

model scientists comparing of tools like team’s expertise
development different SageMaker, and comfort with
tools model PyTorch, different tools
development AutoML,
tools TensorFlow,
Google Cloud
Platforms, Azure,
and scikit-learn
Evaluate Security Document Government Factor in

service’s analyst capturing and company industry-
security security needs, security specific security
capabilities security standards and requirements
architecture, protocols
and service’s
capabilities
Evaluate Financial Document Pricing models Factor in hidden

service analyst outlining costs of the services costs like data
pricing storage, API calls
Evaluate Project Document Vendor’s SLAs, Consider factors

vendor manager comparing track record like uptime
support different guarantees,
vendor responsiveness,
support and skillsets
Identify Compliance Compliance Industry Ensure

industry officer document regulations regulatory
regulations compliance to
avoid fines and
legal issues
Keep up with Project Document Trends in the AI/ Consider both

innovations manager discussing ML space immediate and
growth long-term needs
potential of
AI/ML
AWS AI AND ML SERVICES

AWS provides these ML services in three categories based on the user’s expertise—high-level AI Services, Sage-
Maker, and AI frameworks—as shown in Figure 9.2.
Pre-trained AI services that provide

intelligent solutions for well-known use 1 Core and Sp
ecialized AI
cases such as fraud detection Services
Tool to execute the entire machine learning

pipeline 2 Machine Learning Services
Major frameworks, such as TensorFlow, ructure

nd Infrast
eworks a
MXNet, PyTorch, and Caffe 2 3 ML Fram
FIGURE 9.2: AWS AI/ML stack
AI Services
AWS provides high-level AI services, which are pretrained AI services that provide intelligent solutions for well-
known use cases such as fraud detection, personalized recommendations, contact center intelligence, document
processing, intelligent search, business metrics analysis, forecasting, quality control, cybersecurity, and so on.
These AI services can be incorporated into the applications and workflows and used to build end-to-end
solutions in AWS.
NOTE AWS also provides prebuilt, industry-specific AI services, such as the industrial and
healthcare industries.
Amazon SageMaker
In the middle layer, AWS provides SageMaker. This tool is geared toward experts, such as data scientists and ML
engineers, who can use it to execute the entire machine learning pipeline, all the way from data ingestion, labe-
ling, data preparation, feature engineering, model building, model training, model tuning, model monitoring, and
monitoring deployment. It can also be used for bias detection, AutoML, hosting, explainability, and workflows.
These topics are discussed in greater detail in later chapters.
AI Frameworks
At the bottom of the stack are the AI frameworks that can be used by expert practitioners. This provides all the
major frameworks, such as TensorFlow, MXNet, PyTorch, and Caffe 2. AWS also provides high-performance
instances for annual training, such as P4D instances powered by NVIDIAs, TensorFlow, and core GPUs. For infer-
ences that typically constitute 90 percent of the ML cost, AWS provides EC2 Inf1 instances powered by inferential
chips. Please note that this a fast moving space and you should watch out for the latest developments.
Core AI Services ❘ 113
Differences Between Machine Learning Algorithms, Models, and

Services
It’s important to understand the differences between machine learning algorithms, models, and services. These are
listed in detail in Table 9.1.
TABLE 9.1: Differences Between Machine Learning Algorithms, Models, and Services
ASPECT ML ALGORITHMS ML MODELS ML SERVICES
Definition Set of mathematical An algorithm trained Cloud-based platform

instructions to analyze data with data that provides models and
for learning algorithms as APIs
Purpose To analyze data and learn To make predictions To provide tools and
from it to solve a business and decisions infrastructure for building ML
problem apps
Development Developed by an expert, Created based on Developed by a cloud service

such as a data scientist algorithms trained on company, such as AWS,
data Google
Training Trained on labeled or Generated through a Platform used to train models

historical data training process or perform ML operations
Customization Parameters can be Models can be ML services can be configured

customized to fit specific customized per needs for specific needs
requirements
Complexity Can be simple or complex Depends on the A varied range of complexity

chosen algorithm based on capabilities and use
cases
Deployment Integrated into an Deployed to make Accessed through APIs or

application predictions cloud platforms
Examples of algorithms are linear regression, logistic regression, decision trees, and support vector machines.
Examples of models are deep neural networks, and examples of AI services are AWS SageMaker, Azure Machine
Learning, and Google Cloud ML Engine.
The next sections review the AI services. For the sake of easy understanding, they are categorized into core and
specialized AI services.
CORE AI SERVICES
This section discusses the core AI services offered by AWS, as shown in Figure 9.3.
Amazon Translate Amazon Rekognition

Amazon Comprehend s AWS Panaroma
Amazon Textract ent
um V
id
Do
eo
&
Text
Core AI
Services
Ch
h
bo
at
ec
ts e
Amazon Lex Sp Amazon Polly
Amazon Transcribe
Amazon Transcribe
Call Analytics
FIGURE 9.3: AWS core AI services
Text and Document Services

This section explains some of the text and document services, including AWS Comprehend, AWS Textract, and
AWS Translate.
Amazon Comprehend
Amazon Comprehend is a natural language processing service. It uses machine learning models to classify words
and phrases into categories. It can be used to solve business problems such as the following:
FUNCTIONALITY DESCRIPTION
Key phrase Extract critical phrases and entities such as people, organizations, and products
extraction from text.
Accuracy Increase the accuracy of your search results, product recommendations, and security.
Sentiment analysis Identify the sentiment in a text document, whether it is neutral, positive, or negative.
Text classification Classify text such as product reviews, customer feedback, customer service tickets,
news articles, etc.
Language detection Detect abusive language, phishing attacks in text, and fake news.
Compliance risks Understand customer feedback better and identify compliance risks.
To get started, you need to create an AWS account, enable the Amazon Comprehend service, upload your text
data to Amazon S3, and use Amazon Comprehend to extract insights from it.
AWS Translate
AWS Translate is a fully managed service that can translate content and be integrated into applications. It can
translate various text formats such as HTML, JSON, XML, plain text, and speech. It can be used in the follow-
ing domains:
MARKET APPLICATION
Ecommerce Translate website content to increase sales.
Customer support Translate tickets for better customer support.
Marketing Translate marketing materials to reach a wider audience.
Online sales Translate product documentation.
Training Translate employee training materials for better training.
Healthcare Translate to local languages for patients.
Amazon Textract
Amazon Textract is a machine learning service that uses optical character recognition to extract information such
as text and tables from documents. The document types can be either PDF, image, or image-based documents. It
can be used for the following:
FUNCTIONALITY DESCRIPTION
Document analysis Identify different types of content such as text, tables, and images.
Form extraction Extract information from forms such as names, addresses, phone numbers, etc.
Data extraction Extract information from documents such as invoice numbers, dates, amounts, and
recipients.
Text recognition Recognize text such as authors, document titles, and date created.
Image recognition Recognize images such as company logos and signatures.
Amazon Textract can save time and money extracting data from various documents. These use cases are shown in
Figure 9.4.
Financial Healthcare Legal Industry Human Government

Industry Industry Resources
• Loan Tax Document • Medical record • Contract extraction • Resume and • Improve citizen
Analysis analysis • Document processing application review services
• Automated Invoicing • Patient care • Recruitment
• Eliminate manual improvement acceleration
errors
FIGURE 9.4: Amazon Textract use cases

Amazon Textract can help many industries, including the following:
INDUSTRY APPLICATION
Finance Extract data from loan tax documents and other documents to speed up the loan
approval process and eliminate errors due to manual data entry.
Healthcare Review medical treatment plans, patient medical records, etc., to provide better
patient care.
Legal Extract data from legal documents such as contracts and court filings to reduce the time
required to process this information manually.
Human resources Use it to review resumes and applications and speed up the recruitment process to
recruit the top candidates.
Government Use it to speed up their services provided to their citizens.
TIP Be sure that you thoroughly analyze and understand your specific use cases and require-
ments before choosing AWS Text and Document services such as Amazon Comprehend,
AWS Translate, and AWS Textract.
Chatbots: Amazon Lex

Amazon Lex is used in Alexa, which we are all so used to. Amazon Lex helps build conversational interfaces
to applications using voice and text. It uses the same power as Alexa to build those interfaces. As shown in
Figure 9.5, it can be used for the following:
Customer Support
Answer questions Home Security
Process orders Voice-activated control
Manage accounts Alarm systems
Make recommendations Music playback
Travel Sector Book Reservations

Reservations Virtual assistant
Travel policies Booking and appointments
Flight status
CHATBOTS
HR Applications
Financial Services
Employee inquiries
Time off scheduling Account balances
Information updates Payments
Fund transfers
Education Healthcare
Coursework assistance Scheduling appointments
Recommendations Prescription refills
Q&A Health advice
FIGURE 9.5: Amazon Lex: Conversational interfaces using voice and text
USE CASE DESCRIPTION
Chatbots Can create chatbots that can handle customer questions, process orders, check
shipping status, manage accounts, and make product recommendations.
Voice To control home security systems, alarms, clocks, and play music.
activated systems
Virtual assistants Can take orders and book reservations and appointments.
Here are some industry use cases for Amazon Lex:
SECTOR APPLICATION
Financial Use chatbots to check account balances, make payments, and transfer funds.
Healthcare Schedule appointments for patients, refill prescriptions, and make generic
recommendations related to health.
Education Help students find coursework, make course recommendations, and answer course-
related questions.
Human resources Answer employees’ questions about benefits and policies, schedule time off, and update
personal information.
Travel Make reservations, learn about travel policies, and check flight status.
Speech
Amazon provides Amazon Polly and Amazon Transcribe for speech related services.
TIP You should ensure accurate and high-quality speech-to-text and text-to-speech conver-
sions for optimal performance and best user experience.
Amazon Polly
Amazon Polly is a cloud-based text-to-speech service that uses deep learning to create authentic lifelike voices.
These voices can be added to applications to create innovative speech-enabled products. Amazon Polly can be
used for the following:
USE CASE APPLICATION
IVR systems Develop IVR systems to interact with customers to answer their queries
using NLP.
Enterprise applications Help make applications more accessible by allowing users with visual
impairments and reading difficulties.
continues
(continued)
USE CASE APPLICATION
E-learning and Add audio to courses to make the learning experience pleasant for users who
training platforms prefer to listen rather than read.
Voice-enabled IoT devices Integrate Amazon Poly with a thermostat to control that device and make the
experience much more pleasant.
Amazon Transcribe
Amazon Transcribe is a speech-to-text cloud service capability from AWS that businesses can use to convert large
volumes of audio into text for subsequent processing.
Call centers Convert call center call recordings into text and analyze them to understand customer needs,
preferences, and problems.
Media Convert audio and video content into text to be easily searchable by users and can be
analyzed to understand user engagement and preferences.
Legal Transcribe large volumes of court hearings and proceedings to help law firms and legal
departments.
Healthcare Transcribe notes and records to provide better healthcare.
Vision Services
This section discusses some of the available vision services, such as Amazon Rekognition and AWS Panorama.
Amazon Rekognition
Amazon Rekognition helps businesses add video and image analysis capabilities to their applications. It can detect
objects and analyze people and scenes in videos and images. It is used in the following situations:
SECTOR APPLICATION
Security and Identify people in real-time who are trespassing on the premises.
surveillance
Content moderation Remove inappropriate content in online communities and other online marketplaces.
Marketing and Personalize marketing campaigns by analyzing videos and images to understand
advertising customer demographics, behavior, and preferences.
SECTOR APPLICATION
Retail Categorize products based on images to create catalogs, saving much time.
Manufacturing Detect product deficiencies during manufacturing.
Healthcare Detect diseases proactively by looking at medical images.
TIP You can leverage vision services like Amazon Rekognition and AWS Panorama for
enhanced security, personalized marketing, and optimized operations.
AWS Panorama
AWS Panorama allows you to add computer vision to your applications. It enables you to detect objects, people, and
scenes in videos and images and create 3D models of objects and scenes. It has uses in the following industries:
Retail Track customers and based on their behavior, recommend products they are likely to
buy. Use Panorama to determine where customers spend more time in the store by
tracking their movements. Use this information to redesign the store layout so that
customers can quickly find their products.
Manufacturing Identify defects in the products. Also, track the movement of products within the
manufacturing assembly and improve the manufacturing process.
Supply chain Identify the movement of goods through the supply chain, identify potential areas
management where delays are happening, and identify opportunities to improve.
Warehouse Identify the movement of people and vehicles within a warehouse and identify
management potential opportunities to improve the logistics process.
HANDS-ON PRACTICAL EXERCISE: APPLYING AWS CORE AI SERVICES
This is a hands-on exercise focused on project-based learning experience to leverage AWS AI

core services for practical use.
The goal is to use AWS AI/ML tools to solve real-world problems using services such as text
and document analysis, speech services, vision services, and chatbots.
STEPS TOOLS TASK DELIVERABLES RESOURCES EXAMPLES/TIPS
OWNERS
Task 1: Text AWS Data Develop a AWS Comprehend Use AWS Comprehend for sentiment
and document Comprehend, scientist document AWS Translate analysis on customer reviews.
analysis Translate, detailing results Use AWS Translate to translate content
AWS Textract
Textract and potential use such as customer support tickets
cases. Sample data
Use AWS Textract to extract data from
a set of invoices.
Task 2: Speech AWS Polly ML Develop a AWS Polly Use AWS Polly to develop IVR systems
services and engineer document AWS Transcribe and add audio to e-learning courses.
implementation Transcribe detailing results Use AWS Transcribe to convert call
Sample text and
and potential use center recordings into text for analysis.
audio data
cases.
Task 3: Vision AWS Data Develop a AWS Rekognition Use AWS Rekognition for security
services Rekognition scientist document AWS Panorama surveillance to identify trespassers.
implementation and detailing results Use AWS Panorama to track customer
Sample images
Panorama and potential use behavior in retail stores.
and video data
cases.
Task 4: Amazon Lex AI Develop a working Amazon Lex Use Amazon Lex to build a chatbot that
Chatbots engineer chatbot and Sample prompts handles customer questions, processes
document its for testing orders, and checks shipping status.
functionalities.
Task 5: Project Develop a Use the feedback Collect feedback and results from all
Evaluation and manager comprehensive from task owners, task owners.
report writing report on the objectives, and Evaluate the results against the
implementation, key results (OKRs) objectives and key results (OKRs).
evaluation, and
Prepare a comprehensive report
potential business
detailing the implementation, results,
use cases of AWS
potential business use cases, and future
AI Services.
improvements.
Specialized AI Services ❘ 121
SPECIALIZED AI SERVICES
This section covers the specialized AI services from AWS, as illustrated in Figure 9.6.
Code + DevOps
Amazon DevOps
Amazon DevOps Guru
Search Industrial
Amazon Kendra Amazon Monitron
Amazon Lookout for Equipment
Amazon Lookout for Vision
Business Process Healthcare

Amazon Personalize Amazon HealthLake
Amazon Forecast Amazon Comprehend Medical
Amazon Fraud Detector Amazon Transcribe Medical
Amazon Lookout for Metrics
Specialized AI
Services
FIGURE 9.6: AWS specialized AI services
Business Processing Services

This section covers the various business processing related services that AWS provides, including Amazon Person-
alize, Amazon Forecast, Fraud Detector, and Amazon Lookout for Metrics.
Amazon Personalize
Amazon Personalize is a fully managed service that can personalize customer experiences at scale. It can be used
to make recommendations for websites to improve search results and in marketing campaigns by recommending
the products to customers.
It uses collaborative filtering, content-based filtering, and rule-based filtering. User, behavioral, and product data
are used to make these recommendations.
RECOMMENDATION METHOD DESCRIPTION
Collaborative filtering Uses data from users interested in these products to recommend them to
other users
Content-based filtering Uses data from the products that the user has already interacted with to
recommend other related products
Rules-based filtering Is based on the rules that you specify to recommend the products
Amazon Forecast
Amazon Forecast is a fully managed service that uses machine learning to generate accurate time-series forecasts.
Here are some places where it can be leveraged and a view into some of its features and workings:
RECOMMENDATION METHOD DESCRIPTION
Collaborative filtering Uses data from other users interested in these products to recommend
them to other users
Content-based filtering Uses data from the products that the user has already interacted with to
recommend other related products
Rules-based filtering Is based on the rules that you specify to recommend the products
Figure 9.7 shows how Amazon Forecast works by using latest algorithms to predict future time-series data using
historical data and needs no machine learning experience on the user’s part.
Forecasts can be
Historical Data viewed in a console
Sales, customer,
website, app usage,
financial, weather,
social media, supply
chain data
Upload Data Amazon Forecast Customized Forecasts can be
Upload data into Amazon Forecast does Forecasting exported for
Amazon Forecast the following Amazon Forecast downstream analysis
• Inspects the data uses the trained
• Identifies key model to
attributes generate
• Selects right forecasts
algorithms Forecasts can be
Related Data
• Trains a custom accessed via API
Promotions, events,
model
holidays, product
data, and so on
FIGURE 9.7: How Amazon Forecast works
ARIMA
ARIMA stands for Auto-Regressive Integrated Moving Average and is a statistical model that uses past data to
forecast future values. It is used for the following:
Forecast Can be used to forecast financial data such as stock prices and exchange rates
financial data
Predict Capable of predicting sales or demand during different times of the year such as
sales or demand holidays or seasons, due to its ability to capture trends and seasonality in data
EDS Prophet
EDS (Extended Kalman Filter Dynamic Linear Model) Prophet is a time-series forecasting tool developed by
Facebook. Some of its features include the following:
FEATURE DESCRIPTION
Additive models Uses additive models to capture seasonality trends and other data that affect time-
series data when making predictions
Use cases Useful when you have multiple sources of data such as IoT and sensor data, for
example, to predict when a machine may require maintenance
DeepAR
DeepAR is a deep learning based supervised learning technique. Some of its features include the following:
FEATURE DESCRIPTION
Recurrent Neural RNNs are used to model complex time-series data and can handle multiple time-
Networks (RNNs) series data with different levels of granularity.
Use cases RNNs are used where there are multiple factors that can influence the outcome.
Customer behavior Often used to forecast customer behavior such as website traffic and sales.
Stock prices Can be used to predict stock prices based on past data, current events, and
other factors.
ETS
ETS (Exponential Smoothing State Space Model) is a time-series forecasting method. Some of its features include
the following:
FEATURE DESCRIPTION
Smoothing approach ETS uses a smoothing approach by generating a weighted average of past data to
predict future values.
Seasonality ETS works well with seasonal data and trends and is particularly useful for fast-
and trends changing situations.
Short-term trends ETS is useful to predict short-term trends such as forecasting demand for perishable
goods, or the demand for different types of medications based on historical and
other data.
Note that no forecasting model is perfect. All these models have some degree of error, and you’ll get the best
results by combining predictions from multiple models. Table 9.2 compares and contrasts these forecast models.
TABLE 9.2: Comparison of Amazon Forecast Models for Various Use Cases
MODEL TYPE USE CASES PROS CONS
ARIMA Statistical Sales data, website Simple and easy to Assumes linear
traffic, and customer interpret. Handles relationships. Can
behavior data seasonality and trends well. be sensitive to
outliers.
EDS Statistical Forecasts demand for Handles seasonality and Requires

Prophet products and services trends well. Handles missing manual feature
data well. engineering.
DeepAR Deep Forecast customer Handles complex patterns Takes longer to

learning behavior, such as well and automatically learns train and needs
website traffic and features. more data.
sales
ETS Exponential Demand forecasting, Handles seasonality and Assumes linear

smoothing inventory trends well. Flexible relationships. Can
management, and parameterization is possible be sensitive to
workforce planning and simple to use. outliers.
NPTS Takes longer to

Neural point Demand forecasting, Handles complex patterns
train and needs
forecasting time-series analysis, well and supports multiple
more data.
and anomaly related times series.
detection Powerful and accurate.
Amazon Fraud Detector

Amazon Fraud Detector is a fully managed service that can detect frauds like account takeover, payment fraud,
return fraud, and other suspicious activities such as large orders and unusual login patterns. Some of its features
include the following:
FEATURE DESCRIPTION
Prebuilt and You can either use prebuilt models or create your own custom models.
custom models
AWS console You can use the AWS console to create a detector, feed in the labeled data to train
the detector, test it, and, when satisfied, deploy it into production.
Amazon Lookout for Metrics

Amazon’s Lookout for Metrics is a fully managed service that can be used to monitor metrics or a set of metrics
and detect any anomalies and generate alerts. You can then quickly investigate and diagnose issues. Some of its
features include the following:
FEATURE DESCRIPTION
Detect anomalies Used to detect anomalies in websites, traffic, machine downtime, sales data inventory
levels, customer satisfaction, and operational data.
Automatic detection Automatically detects and labels anomalous data points and even identifies the root
cause behind those anomalies.
Integration Designed to be fully integrated with other AWS data storage solutions such as
Redshift, S3, RDS, and other third-party databases.
Kendra for Search

Amazon provides Kendra for the Search capability. Amazon Kendra is a fully managed service that can automati-
cally index your documents using machine learning algorithms and thus provide more relevant search results.
Kendra makes it easy to find documents quickly.
Figure 9.8 shows some of the inner workings of Amazon Kendra.
Ingest data Intranet

Sync multiple data Incremental search
sources using connectors learning
Relevance
tuning
Website
Data Source Pre-trained search
Connectors domains
S3, Salesforce,
Slack, etc. Custom content
attributes
Amazon Kendra Optimize results Contact
Search enabled by Techniques to Custom center
machine learning improve search synonyms agent assist
for websites and accuracy
apps use Custom Data Source Custom
document
Connectors
enrichment
S3, Salesforce, Slack, Embedded
etc. search apps
FIGURE 9.8: How Amazon Kendra works
Some of Kendra’s features include the following:
Employee knowledge base Set up an enterprise search solution that includes content from databases,
websites, and other internal documentation such as FAQs, employee
handbooks, and product documentation.
Customer support knowledge For customer support teams that include FAQs, troubleshooting guides,
management system product documentation, and customer support tickets.
continues
(continued)
Partner management Index partner agreements, FAQs, and sales content.
Customer satisfaction Improve employee productivity, customer satisfaction, and partner

satisfaction.
Ecommerce Create a custom search engine where customers can find their products
quickly, thus increasing customer service satisfaction and sales.
Healthcare Create a medical search engine that enables doctors and professionals to
find their information quickly; this can help doctors make better decisions
and improve patient outcomes.
Legal Create a legal search engine for lawyers to find their information quickly,
which can improve their service to customers by building more robust
legal cases and providing better legal advice.
Code and DevOps

This section covers some of the code and DevOps-related AWS services, including Amazon CodeGuru and Ama-
zon DevOps Guru.
Amazon Code Guru

Amazon CodeGuru leverages AI and ML algorithms to do the following (see Figure 9.9):
Write Code
01
Run App Review Code

Amazon CodeGuru 05 A
Ama
Am
ma
m zon CodeGuru
Amazon 0
022 Amazon CodeGuru
Profiler Uses M
Machine Reviewer
Lea
e arning
rning
Learningg to Impr
Improve
p
Ap
pp
p Code
App
Test App
Deploy App 04 03
03 Am
Amazon CodeGuru
Reviewer
FIGURE 9.9: Amazon CodeGuru uses machine learning to improve app code.
BENEFIT DESCRIPTION
Improve Conduct code reviews, automated suggestions, and remediations to enhance the
code quality quality of the code.
Find and fix bugs Reduce development time by providing suggestions to improve the code.
Improve security Find security bugs and help comply with regulations by providing suggestions to
improve the code.
Improve Provide suggestions to code so that you can leverage the appropriate infrastructure
performance to improve performance.
TIP Use Amazon CodeGuru to enhance code quality, security, and performance, and use
DevOps Guru to proactively monitor DevOps processes and detect issues.
Amazon DevOps Guru

Amazon DevOps Guru uses machine learning to identify potential operational issues before they impact the cus-
tomers. Some of its capabilities include the following:
FEATURE DESCRIPTION
Anomaly detection Leverages operational data from AWS CloudWatch, CloudTrail, and other AWS
services, continuously analyzes them, and detects anomalies that are then used
to generate alerts for subsequent reactive or proactive action
Real-time alerts Generates alerts through integration with AWS Systems Manager, Amazon
EventBridge, Amazon SNS, and other third-party incident management systems
Performance, security, Helps improve performance, provides security, and ensures high availability by
and availability preventing downtime
TIP The integration capability of DevOps comes in handy, especially for AI and ML appli-
cations, because they are often complex, require considerable computer resources, and are
challenging to monitor in real time.
HANDS-ON PROJECT-BASED LEARNING EXERCISE
This is a type of learning exercise that is focused on applying the skills and knowledge gained
about AWS services to a series of tasks in the real world.
continues
continued
OBJECTIVE DELIVERABLES RESOURCES EXAMPLE
Get access to AWS AWS Account Internet Create an AWS

Services. connection account.
Email ID
Learn to create solutions Understand Amazon Use Amazon

and campaigns, import, Amazon Personalize Personalize
or create datasets, and Personalize AWS Console to make
generate personalized recommendations
recommendations for based on user
customers. behavior.
Learn to import time- Forecast model Amazon Forecast Use ARIMA,

series data, select the using Amazon Time-series Prophet,
appropriate predictor Forecast dataset DeepAR, ETS
algorithm (e.g., ARIMA, to generate
Prophet, DeepAR, ETS), forecasts for
and visualize forecasts. sample dataset.
Learn to use historical Fraud detection Amazon Use Amazon

event data and known model Fraud Detector Fraud Detector to
fraud labels to train a Fraud identify potential
model, create rules, and labeled dataset fraudulent
make decisions using transactions in
Amazon Fraud Detector. sample data.
Test the model with
sample transactions to
identify potential fraud.
Monitor a sample dataset Anomaly Amazon Set up an

for anomalies. Learn detection Lookout anomaly detector
how to set up anomaly model for Metrics and identify
detectors, add datasets, Dataset anomalies in the
and interpret anomaly sample data.
detection results.
Create an intelligent Intelligent Amazon Kendra Create an

search service using search service Document intelligent search
sample documents. Learn using Kendra dataset service for a
to create an index, add set of sample
data sources, and test the documents.
search functionality.
OBJECTIVE DELIVERABLES RESOURCES EXAMPLE
Stop or delete all services Cleaned up AWS Console Ensure to

to prevent any additional AWS services stop or delete
charges. services to avoid
any additional
charges.
Learn how to apply these Reflection on N/A Reflect on what

learnings to your project learning you learned in
for the benefit of your this hands-on
company. exercise and how
you can apply
these services in
your business or
project.
Industrial Solutions
Amazon provides a few industrial solutions, including Amazon Monitron, Amazon Lookout for Equipment,
and Amazon Lookout for Vision (see Figure 9.10). These solutions apply artificial intelligence (AI) and machine
learning (MI) to vast amounts of industrial data generated by sensors and systems to generate insights so they can
optimize industrial processes and improve quality and safety.
Predictive
maintenance,
Amazon operational efficiency
Monitron
01
02
Amazon Detect defects and
Amazon Lookout anomalies
Lookout for
for Vision
Equipment
03 Reduce downtime,
improve reliability
FIGURE 9.10: Amazon industrial AI solutions

Amazon Monitron
Amazon Monitron is an industrial solution that can detect anomalies in industrial equipment. Some of its features
include the following:
FEATURE DESCRIPTION
Monitoring Monitron monitors temperature, vibration, and sound.
Predictive Combines sensors, gateways, and machine learning to reduce unplanned downtimes
maintenance and improve predictive maintenance.
Operational Leverages industrial IT and edge-creating technologies to integrate with industrial

efficiency equipment and systems to provide real-time insights, improving operational efficiency.
Amazon Lookout for Equipment

Amazon’s Lookout for Equipment is another service focused on preventing downtime and improving the opera-
tional efficiencies of equipment for industrial companies. Some of its features include the following:
FEATURE DESCRIPTION
Issue detection Detects vibration, flow, temperature, sound, and pressure to identify potential issues.
Equipment care Improves equipment reliability, reduces maintenance costs, and improves safety.
Industry adoption Can be used in several industries, such as manufacturing, transportation, and energy
automotive, to monitor equipment such as pumps, motors, bearings, valves, and more.
Use cases Industrial use cases include quality control, energy optimization, predictive maintenance,
and asset management.
Amazon Lookout for Vision

Amazon’s Lookout for Vision is used to detect defects in industrial images. This includes defects such as cracks,
scratches, and missing parts. Some of its features include the following:
FEATURE DESCRIPTION
Detect This service or tool can identify visual damages including dents and water damage and
visual damages safety hazards due to oil spills, leaks, blocked walkways, and product quality issues such
as color and texture anomalies.
Use cases This tool is particularly useful in industries such as manufacturing, packaging, and
logistics.
Healthcare Solutions
AWS provides several healthcare-related services, including Amazon HealthLake, Amazon Comprehend Medical,
and Amazon Transcribe Medical, each of which is discussed in the following sections.
Amazon HealthLake
HealthLake is a managed service that stores, transforms, and analyzes healthcare data to improve healthcare
quality, reduce costs, and improve the efficiency of healthcare organizations. Here are some of its features:
FEATURE DESCRIPTION
Proactive care HealthLake can look into patients’ data and try to identify those with the risk of
developing complications and then proactively treat or educate them.
Cost reduction Hospitals can use HealthLake to understand where they are spending money, such
as on prescription drugs, and then negotiate that with the drug companies for
better pricing.
Appointment Hospitals can automate scheduling appointments using HealthLake, which will not
scheduling only impart efficiency to the hospital’s operations but also result in better patient
satisfaction.
Clinical trials HealthLake can analyze large datasets to identify anomalies, which can help medical
researchers identify potential participants for clinical trials and lead to new therapies
and treatments.
Clinical HealthLake can also act like a clinical support system by providing physicians with
support system the latest clinical research and insights to help them change their treatment plans or
identify other potential treatment options.
Population health Healthcare organizations can also track the trends in population health over time and
then proactively produce means to prevent diseases or improve overall health.
TIP You can leverage Amazon healthcare services for proactive treatment, cost optimization,
automation, clinical support, and secure medical transcription.
Amazon Comprehend Medical

Amazon Comprehend Medical is like Amazon HealthLink in that it also helps reduce costs, improve patient care,
and improve hospital operations. Here are some of its features:
FEATURE DESCRIPTION
Population health Helps with population health management by better understanding the patient’s health
and developing personalized treatments.
Medical coding You can also use Amazon Comprehend Medical to automatically code medical diseases
and procedures for billing purposes, thus saving the time and effort required for
medical coding.
Amazon Transcribe Medical

Amazon Transcribe Medical is a HIPAA-eligible service that uses Automatic Speech Recognition to allow physi-
cians, vendors, and payers to securely transcribe medical speech, such as physician-patient conversations, medical
dictations, and clinical notes. Here are some of its features:
FEATURE DESCRIPTION
Compliance HIPAA-eligible
Technology Uses automatic speech recognition
Users Physicians, vendors, and payers
Use case Securely transcribes medical speech, such as physician-patient conversations, medical
dictations, and clinical notes
Clinical notes Includes progress notes, discharge summaries, and consultation notes
HANDS-ON EXERCISE: APPLYING AMAZON INDUSTRIAL AI SOLUTIONS

USING AMAZON MONITRON, LOOKOUT FOR EQUIPMENT, AND LOOKOUT
FOR VISION
The focus of this exercise is to provide a practical experience to using these Amazon Industrial
solutions to improve operational efficiency, predictive maintenance, and quality control in an
industry setting.
OBJECTIVE STEPS AND RESOURCES EXAMPLE

DELIVERABLES
Assume Step 1: Configure ➤➤ AWS account Set up Amazon

that you are Amazon Monitron Monitron to monitor
➤➤ Amazon Moni-
working with service and connect a hypothetical
tron
an industrial it to the company’s conveyor belt system
manufacturing equipment. ➤➤ Documentation in a warehouse.
company Step 2: Monitor
experiencing temperature, pressure,
equipment vibration, sound data.
downtime.
Step 3: Analyze this
The goal is to
data using Amazon
monitor the
Monitron to identify
equipment to
potential issues.
identify the
underlying
causes.
OBJECTIVE STEPS AND RESOURCES EXAMPLE

DELIVERABLES
The goal is to ➤➤ Configure Amazon ➤➤ AWS account Use Amazon Lookout

move away Lookout for Equip- ➤➤ for Equipment
Amazon Lookout
from manual ment and connect to analyze the
for Equipment
inspections to it to their data historical data of
developing sources.
➤➤ Documentation a hypothetical
their ➤➤ ➤➤ Sample industrial motor to detect any
Ingest historical
predictive dataset anomalies.
data and detect
maintenance
anomalies.
capability.
➤➤ Evaluate the
effectiveness of
the new system by
comparing it with
the accuracy of the
predictions during
manual approach.
The goal is to ➤➤ Configure Amazon ➤➤ AWS account Use Amazon

improve the Lookout for Vision ➤➤ Lookout for Vision
Amazon Lookout
quality control service. to analyze images of
for Vision
process. ➤➤ a production line to
Connect it to the ➤➤ Documentation detect any defects in
company’s produc-
tion line cameras.
➤➤ Sample industrial products.
images
➤➤ Analyze images for
defects.
➤➤ Monitor the ser-
vice performance
and optimize the
parameters to
improve accuracy.
Apply it in the Develop a detailed ➤➤ Results from the ➤➤ Evaluate the

real world. report evaluating the previous steps number of true
results obtained from ➤➤ positives, false
Evaluation met-
Amazon Monitron, positives, true
rics relevant to
Amazon Lookout negatives, and
the use case
for Equipment, and false negatives
Amazon Lookout for detected.
Vision. ➤➤ Compare the
benefits and
challenges while
using these
AWS services.
MACHINE LEARNING SERVICES

This section covers the ML services. Within the AWS ML services section, there are the SageMaker Canvas,
SageMaker itself, and the SageMaker Studio Lab, as shown in Figure 9.11.
A fully managed machine learning service

to build and deploy ML models
1 Amazon Sa
geMaker
No-code ML drag-and-drop feature in

SageMaker for nonexperts 2 SageMaker Canvas
Interactive tutorials that help users to Lab

er Studio
learn machine learning quickly 3 SageMak
FIGURE 9.11: AWS ML services
Amazon SageMaker
Amazon SageMaker is a fully managed machine learning service provided by AWS to build, train, and deploy
machine learning models. It provides end-to-end support for the entire machine learning lifecycle, as shown in
Figure 9.12. It allows developers and data scientists to deploy models at scale in production.
s end to end su
rt ovide ppo
Capabilities Sta Pr rt f
or
include D ata Stor
e fu
ll
a r e F e
d
notebooks, ep atures
Pr
En
model training,
ac
es e
vic dg
hin
De
hosting,
De ge E
el
te
autopilot,
c
a
ear
tB
an
integration
ias
M
ning
lifecycle
Manage &
Notebook
Build with
Monitor
Amazon
SageMaker
els
Ex icti
Pr
od
ed
pl on
M
ian s
Speeds up development
ain
Enables developers and

by allowing data
Tr
Dep
lo e data scientists to deploy
scientists and Prod y in Tun ters
uctio e models at scale in
developers to focus on n P aram production
core tasks
FIGURE 9.12: SageMaker’s capabilities

Machine Learning Services ❘ 135
Some of its features include the following:
CATEGORY FEATURES
Capabilities ➤➤ Uses Jupyter notebooks to develop and train machine learning models
➤➤ Provides training environments that can be local cloud or on edge
➤➤ Offers deployment options such as real-time or batch inference
➤➤ Provides monitoring options such as metrics, logs, and traces
Benefits Helps companies to reduce costs, improve efficiency, and make better business decisions
Amazon SageMaker Canvas

SageMaker Canvas is a new drag-and-drop feature in SageMaker that helps democratize the use of SageMaker
for non-experts. Here are some of its features:
FEATURE DESCRIPTION
Collaboration Enables more collaboration between developers, data scientists, and business analysts.
Speed Helps organizations speed up the machine learning workflow and deploy
models quickly.
Democratization Business analysts can use SageMaker with limited machine learning knowledge.
Use Cases Includes fraud detection, product recommendations, and customer churn prediction.
SageMaker Studio Lab

Amazon SageMaker Labs are interactive tutorials that help users learn machine learning quickly. They offer an
opportunity to learn with a hands-on approach. Here are some of its features:
FEATURE DESCRIPTION
Upskill An essential tool for organizations to upskill their employees
Hands-on Helps employees to learn machine learning development through guided, hands-on
experience experiences
Experimentation Allows users to learn and experiment with different models without worrying about
the underlying hardware and infrastructure
Collaboration Allows users to collaborate and share their work
TIP Use Amazon SageMaker’s end-to-end support for ML lifecycle, democratize ML with
SageMaker Canvas, and upskill employees with SageMaker Studio Lab.
THE GOOGLE AI/ML SERVICES STACK

AI/ML services can be categorized into three major use cases—for the data scientist, for the developers, and for
the infrastructure specialists. These first two use cases are covered in this chapter, and the infrastructure-level
services are covered when I talk about the cloud platform in Chapter 13.
For Data Scientists

The following AI/ML services are available for data scientists, as shown in Figure 9.13.
End-to-End Workflow
Data Preparation
Model Training, Deployment, Management
Vertex AI
Integrated Development Environment

Rapid Prototyping
Model Development and Deployment Vertex AI Workbench
Model Sharing and Deployment

Collaboration Tools
Access to Datasets and APIs AI Hub
FIGURE 9.13: Google AI/ML stack for data scientists
Vertex AI
Vertex AI is a cloud-based machine learning platform that helps companies build, train, and deploy models in
production. It helps with the end-to-end machine learning workflow, including data preparation, model training,
deployment, and management.
Vertex AI Workbench
A Vertex AI Workbench is a single development environment for data scientists who can develop and then deploy
models. Here are some of its features:
FEATURE DESCRIPTION
Easy transition from development Integrated with Vertex AI and other tools, which helps minimize the
to production transition from development to production
Rapid prototyping Facilitates rapid prototyping and model development

The Google AI/ML Services Stack ❘ 137
AI Hub
Google’s AI Hub is a centralized platform where data scientists, machine learning engineers, developers, and other
AI practitioners discover, share, and deploy models from within the hub. Here are some of its features:
FEATURE DESCRIPTION
Share assets Gain access to other assets such as datasets and APIs.
Collaboration tools Access other tools for collaboration, such as version control, code reviews, and
issue tracking.
GOOGLE’S AI/ML EXERCISE FOR MACHINE LEARNING
This hands-on exercise is focused on helping data scientists and AI practitioners to use
Google’s AI/ML services such as Vertex AI, Vertex AI Workbench, and AI Hub. It will help the
team members to create and deploy models and prototype solutions.
Task 1: Assume you are working for a retail company to improve the sales using a predic-
tive model.
TOOLS STEPS
Vertex Gather and prepare necessary data such as historical sales data and data
AI related to external factors that could impact the sales.
Train a model for sales predictions using Vertex AI.
Deploy the trained model and create a mechanism for new data input.
Monitor and adjust the model’s performance as needed.
Task 2: Experimenting with different recommendation systems to personalize the customer’s

shopping experience using rapid prototyping.
TOOLS STEPS
Vertex AI Use Vertex AI Workbench to prototype various recommendation

Workbench system models.
Select a model to integrate with the company’s e-commerce platform

for a pilot test.
Monitor the effectiveness of the recommendation system and adjust

as necessary.
Task 3: Streamlining asset sharing and collaboration across various teams.

continues
continued
TOOL STEPS
AI Hub Set up a workspace on AI Hub for sharing assets such as datasets,

models, and APIs.
Establish workflow for version control, code reviews, and issue

tracking.
Monitor and adjust team collaboration and asset utilization.
For Developers
The AI/ML services listed in Table 9.3 are available to developers.
TABLE 9.3: Google AI/ML Services for Developers
AUTOML CLOUD NATURAL DIALOGFLOW IMAGE AND VIDEO MEDIA

LANGUAGE SERVICES TRANSLATION
Automated Sentiment Conversational Video AI Speech-to-text

model analysis Interfaces ➤➤ Video tagging ➤➤ Audio to text
development ➤➤ Entity ➤➤ Chatbots/ recommenda- conversion
➤➤ AutoML recognition voice assis- tions ➤➤ Real-time or
Vision tants ➤➤
➤➤ Syntax Transcription prerecorded
➤➤ AutoML analysis ➤➤ Natural and translation audio
Natural language ➤➤
➤➤ Content Content Text-to-speech
Lan- understand-
classification moderation ➤➤ Text-to-speech
guage ing
➤➤ Text ➤➤ Label detection conversion
➤➤ AutoML ➤➤ Media
classification ➤➤ Personalized
Tables Translation Vision AI
communications
➤➤ ➤➤ Object and face
Multi-device
support detection Translation AI
➤➤ ➤➤ Cloud
Text
recognition translation API
➤➤ ➤➤ Google
Image search
Translate
➤➤ Landmark
➤➤ Neural machine
detection
translation
AutoML
AutoML is a set of tools and techniques that facilitates the development of models accessible by automating
many tasks. Some of its features include the following:
FEATURE DESCRIPTION
Easier model Intended to make model development easier for those with limited knowledge of
development machine learning
Development tools Provides several tools, such as AutoML Vision, AutoML Natural Language, and
AutoML Tables, to solve specific ML problems
Cloud Natural Language

Cloud Natural Language is a cloud-based natural language processing service offered by Google Cloud for devel-
opers to analyze and understand human language in text format. Some of its features include the following:
FEATURE DESCRIPTION
Natural Helps with sentiment analysis, entity recognition, syntax analysis, content
language tasks classification, and more.
Application Developers can apply natural language processing to applications using this API.
integration
Sentiment analysis Developers can train machine learning models to classify, extract, and detect
sentiments.
Dialogflow
The Google Dialogflow platform enables developers to create conversational interfaces for applications such as
chatbots, voice assistants, and other conversational agents. Some of its features include the following:
FEATURE DESCRIPTION
Device support Works across multiple devices and platforms.
User requests Leverages machine learning and natural language understanding techniques to
understand user requests and provide responses.
Human-like Facilitates natural interaction and human-like capabilities and offers enterprise-wide
interactions scalability.
Application Helps developers develop such applications very quickly.

development
TIP Use Vertex AI for data scientists, AutoML and Dialogflow to streamline development,
and Cloud Natural Language and Vision AI to enhance applications.
Media Translation
Media translation enables real-time translation of audio or video content into text in the language of your choice.
Currently, it supports 12 languages. The following sections discuss the services that AWS offers under the media
translation umbrella.
TIP Media translation can be used to make your content more accessible to a global audi-
ence.
Speech-To-Text
FEATURE DESCRIPTION
Applications Applications such as dictation, podcast, voice commands, and more.
Google Speech to Text plays a vital role in the Google Contact Center AI by converting
Contact Center AI spoken words into written text, which helps real-time transcription, sentiment
analysis, call analytics, and interactive voice response (IVR).
Text-to-speech Provides real-time text conversion into speech and can be used for video
conversion captioning, voice-enabled chatbots, transcribing meetings and calls, and so on.
Supports This includes English, French, Spanish, and Japanese. It can understand multiple
multiple languages dialects and accents.
Personalized Can be used to personalize communications and engage with customers via voice-
communications based user interfaces in devices and applications.
Translation AI
Google offers this translation service, which can be used to translate text and speech between different languages.
Some of its features include the following:
FEATURE DESCRIPTION
Global reach Helps reach customers globally and provides compelling experiences by
engaging with them in their local languages.
Translation options Google provides three translation options: Cloud Translation API, Google
Translate, and Neural Machine Translation.
Language support Google Translate is free and can be used for nearly 100 languages. It translates
text, images, speech, and web pages.
Google Translate API A paid service that can be used to integrate translation into your applications
and services.
Neural machine A much more accurate translation service that uses neural networks for
translation translation.
Image and Video Services

This section covers some of the AI/ML image and video services offered by Google.
Video AI
Google Video AI is used in many applications such as Google Photos, YouTube, and Google Cloud. It provides
automatic video tagging, video recommendations, video transcription and translation, and content moderation.
Vision AI
Google Vision AI is a cloud-based image analysis service that detects and analyzes objects, faces, text, and other
visual entities from videos and images. It is helpful for label detection, face detection, text recognition, optical
character recognition (extracting text for scanned documents), image search (finding similar images), and land-
mark detection.
PROJECT-BASED EXPLORING GOOGLE AI/ML SERVICES FOR DEVELOPERS
In this exercise, the project team will build an application using Google AI/ML
services using image recognition, natural language processing, translation, and
chatbot functionality.
STEP NO. GOOGLE AI/ML HANDS-O N TASK EXPECTED OUTCOME
SERVICES
1 Automated Use Google’s AutoML A functional machine

Machine to develop a machine learning model
Learning learning model for a developed with minimal
(AutoML) simple problem such coding effort
as image classification
or text analysis.
2 Cloud Natural Develop a simple An application capable

Language application that uses of analyzing sentiments
the Google Cloud from text
Natural Language API
to analyze sentiments
in user-provided text.
3 Dialogflow Create a simple A chatbot capable

chatbot using of conversational
Dialogflow that interactions with users
understands user
intents and responds
accordingly.
4 Media Integrate Google’s An application capable

Translation Media Translation into of real-time translation
the application and across various
test it with different languages
input languages.
continues
continued
STEP NO. GOOGLE AI/ML HANDS-O N TASK EXPECTED OUTCOME

SERVICES
5 Speech-to-Text Implement a feature An application feature

in the application capable of converting
that converts spoken voice inputs to text
words into written
text for transcribing
meetings or voice
commands.
6 Translation AI Use Google’s An application feature

Translation AI to capable of translating
translate text from text into a specified
one language to language
another, preferably by
auto-translating user-
generated content.
7 Vision AI Implement Google’s An application feature

Vision AI in the capable of analyzing
application to analyze and identifying objects
images or videos, or components in an
such as identifying image or video
objects within an
uploaded image.
Bonus Integration of Integrate these A seamless user journey

challenge Services functionalities into a across various AI/ML
cohesive user journey services in a single
within the application. application
THE MICROSOFT AI/ ML SERVICES STACK

The Microsoft AI/ML services stack can be broadly categorized into four areas that enable more advanced AI/ML
practitioners and the novice. The Azure Applied AI services are geared toward data scientists and machine learn-
ing engineers, while Azure machine learning is a platform for developers.
Azure Applied AI Services

Azure Applied AI services are capabilities that developers, data scientists, and other AI practitioners can use in
their applications without needing expertise in machine learning. Figure 9.14 shows an overview of the various
Azure-applied AI services and their capabilities.
Azure Cognitive Services

Azure Cognitive Services is a suite of cloud-based APIs and services that helps developers integrate AI and ML
capabilities into their applications, websites, and devices.
The Microsoft AI/ ML Services Stack ❘ 143
Natural language
Extract entities from audio
processing, computer
and videos
vision, speech recognition
Azure Video Azure
Indexer Cognitive
Services
Extract entities such as key Azure Develop conversational

value pairs, tables, and text Applied Al
Azure Form Azure Bot chatbots
from forms Recognizer Services Service
Azure Azure Draw insights in real time

Use Al and NLP to search
Cognitive Stream from various sources such
for data such as images and
Search Analytics as loT devices, social media
text
clickstreams
FIGURE 9.14: Azure-applied AI services
FEATURE DESCRIPTION
Purpose Helps developers integrate AI and ML capabilities into their applications, websites,
and devices
Capabilities Includes natural language processing, computer vision, speech recognition, and more
Use cases Can improve customer service, fraud detection, product recommendations, image
classification, and natural language processing
Method Just make an API call to embed the ability to speak, hear, see, search, understand, and
thus include decision-making into the apps
Specific services Can be used for speech-to-text, text-to-speech, speech translation, and speaker
recognition
Cognitive Services for Language

Cognitive Services for Language offers capabilities such as entity recognition, language translation, sentiment
analysis question answering, conversational language understanding, and translation.
Cognitive Services for Vision

Cognitive Services for Vision analyzes content within images and videos and provides custom image recognition.
It has the following features:
FEATURE DESCRIPTION
Cognitive service Provides content moderation and personalization of experiences for the users, and
for decisions addresses device problems earlier, all contributing to better decision-making
Open API service Provides an open API service for advanced language models and coding
Azure Bot Service

Azure Bot Service is a cloud-based platform for developing conversational chatbots that interact with users across
websites, messaging apps, and voice assistants.
TIP Use Azure Applied AI services to integrate AI capabilities into your applications and
unlock the power of natural language processing, computer vision, real-time analytics, and
intelligent search.
Azure Stream Analytics

Azure Stream Analytics is a cloud-based analytics platform that draws insights in real-time from various sources,
including IoT devices, social media clickstreams, and more. It can integrate with Azure events hubs, Azure IoT
hub, and Azure functions to create complex streaming data pipelines.
TIP Azure Stream Analytics can be used to monitor and analyze streaming data in real time,
which can help you make better decisions and take action more quickly.
Azure Cognitive Search

Azure Cognitive Search allows developers to integrate enterprise-grade search into their applications. Some of its
features include the following:
FEATURE DESCRIPTION
AI and NLP Uses AI and Natural Language Processing to search for data such as images and text
Search features Provides semantic search, faceted navigation, fuzzy matching, and geospatial
search features
Unstructured data Can extract insights from unstructured data such as sentiment analysis, key phrase
extraction, and entity recognition
Azure Form Recognizer

Azure Form Recognizer helps to extract entities such as key-value pairs, tables, and text from forms. It enables
data entry to be faster and error-free as compared to manual data entry. It can be used on documents such as
invoices, receipts, and custom documents.
Azure Video Indexer

Azure Video Indexer is a cloud-based platform that extracts entities from audio and videos. Once you upload
the content to a portal, it extracts information using speech-to-text, facial recognition, sentiment analysis, scene
detection, and so on. Users can then search through this information, either through a portal or through an API.
Using Azure Video Indexer is easy; just go to videoindexer.ai and upload a video after signing up. Figure 9.15
shows the results for a video that I uploaded. It came up with a list of relevant topics, transcripts, keywords,
named entities, and emotions.
The Microsoft AI/ ML Services Stack ❘ 145
FIGURE 9.15: Using Azure Video Indexer tool
Azure Machine Learning

Azure Machine Learning is a cloud-based machine learning platform. Developers can use it to build, train, and
deploy machine learning models into production. It allows developers to prepare data, build and train models,
validate, deploy, manage, and monitor the results.
PHASE FEATURE
Data preparation During data preparation, developers can label training data, manage label projects, use
analytics to explore data, and access and share datasets.
Build and Provides notebooks, automated machine learning capability, the ability to run
train models experiments, accelerate training using a scalable compute engine, access to a secure
compute instance, and access to open-source libraries such as Scikit-learn, PyTorch,
TensorFlow, Keras, and more.
Validate and Provides the ability to deploy model endpoints for real-time and batch inference,
deploy models pipelines and CI/CD workflows, prebuilt instances for inference, model repository, and
hybrid and multi-cloud deployment options.
Manage and Provides the ability to track logs, analyze data and models, detect drift, maintain
monitor models model accuracy, debug models, trace machine learning artifacts for compliance, and
ensure continuous monitoring of security and cost control features, including quota
management and automatic shutdown.
EXERCISE: BUILDING, DEPLOYING, AND MONITORING AN AI MODEL WITH

AZURE SERVICES
This exercise is to give your team practical experience with building, training, and deploying a
machine learning model using Azure Machine Learning. It includes important aspects such as
data preparation, model training, validation, deployment, and monitoring.
Remember to clean up resources at the end of the exercise to avoid unnecessary charges. Also
note that the roles in the exercise can vary in your organization.
STEP TASK OBJECTIVE INSTRUCTION DELIVERABLE

NAME OWNER
Data Data Prepare data Import a dataset A cleaned

preparation engineer for model and use Azure’s and
training data preparation preprocessed
tools to preprocess dataset ready
the data. for model
training
Model Data Build and Choose a suitable A trained

building scientist/ML train a ML algorithm machine
and training engineer machine and use it to train learning
learning a model on the model
model prepared dataset.
Monitor the
training process
using Azure’s
tracking tools.
Model Data Evaluate Evaluate Evaluation

validation scientist/ML the model’s the model’s metrics and
engineer performance performance using analysis
a holdout validation
set and appropriate
evaluation metrics.
Model MLOps Deploy the Deploy the model A deployed

deployment engineer/ model for use using Azure’s machine
ML deployment learning
Engineer options for real- model
time or batch
inference.
Monitoring MLOps Monitor Use Azure’s Regular

engineer the model’s monitoring tools to performance
performance track model usage, reports and
over time data drift, and updated
retrain the model models as
as necessary. necessary
Other Enterprise Cloud AI Platforms ❘ 147
OTHER ENTERPRISE CLOUD AI PLATFORMS

This section covers several other enterprise cloud AI platforms that are worth knowing about and considering.
Dataiku
Dataiku is an enterprise-class data science platform that allows data preparation, machine learning, and collabo-
ration all in one place. It enables use cases such as customer segmentation, fraud detection, product recommenda-
tion, inventory management, and supply chain optimization.
Dataiku has key capabilities for explainable AI. It includes interactive reports for feature importance, partial
dependence plots, subpopulation analysis, and individual prediction explanations.
DataRobot
DataRobot’s specific capabilities include AutoML, which automates the machine learning development and
deployment lifecycle, and explainability, which explains how the models work and helps businesses build trust
around their models. It facilitates collaboration by allowing users to share their work, including code, models,
and workflows, in a collaboration platform. Its governance platform provides the tools to manage risks with
machine learning models to comply with regulations and protect data.
KNIME
KNIME is a powerful open-source data analytics tool that helps with data preparation and machine learning
modeling. In addition to typical data preparation feature extraction and machine learning modeling tools, it pro-
vides a web extraction tool that extracts data from websites.
IBM Watson
IBM Watson is a robust set of AI/ML tools that businesses can leverage to carry out several use cases around nat-
ural language processing, speech recognition, machine vision, and so on. It provides capabilities such as natural
language processing, machine learning to make predictions, deep learning for highly accurate speech recognition
and image recognition, speech-to-text, text-to-speech, visual recognition, and virtual agents.
TIP When choosing an enterprise cloud AI platform, it is important to thoroughly assess the
features, scalability, security, and integration capabilities of the platform.
Salesforce Einstein AI
Salesforce Einstein AI is a robust set of AI and machine learning tools developed by Salesforce to enhance the
CRM experience for businesses. Some of its key capabilities include predictive analytics, NLP, image recognition,
automated insights, and intelligent automation for workflows.
Oracle Cloud AI
Developed by Oracle, this product helps developers and businesses to integrate AI into their applications and
workflows. It provides capabilities such as machine learning, an autonomous database to manage the data
automatically, chatbots that can integrate with messaging platforms and websites, intelligent apps that provide
recommendations and automate tasks, and data analytics.
HANDS-ON EXERCISE: EVALUATE PLATFORMS
This hands-on exercise is focused on researching the capabilities of the platforms to implement
the use cases under consideration. The team will be evaluating the platforms against the fol-
lowing: features such as AutoML, explainability, collaboration, security, scalability, integration,
and deployment options.
STEP OBJECTIVE INSTRUCTIONS DELIVERABLE
Research Learn about Research key Summary

platforms assigned platforms capabilities and use of platform
cases. capabilities
Define evaluation Determine metrics Identify criteria such List of six

criteria to assess platforms as features, scalability, evaluation criteria
security, integration,
ease of use, and
pricing.
Evaluate Analyze platforms Evaluate platforms Completed

platforms based on criteria based on each criteria. evaluation table
Present findings Share evaluation Present assessment Presentation

with team within the allotted summarizing
time. platform
assessment
Discuss and Decide on top Discuss platforms, List of top one

finalize platform(s) as a identify gaps, and to two platforms,
team determine next steps. plan for piloting
WORKBOOK TEMPLATE: AI/ML PLATFORM EVALUATION SHEET

Download the “AI/ML Platform Evaluation Sheet” template from the download section of the book (www.wiley
.com/go/EnterpriseAIintheCloud). Use this template to help your team evaluate various AI/ML platforms
for your company’s needs. It will help you to evaluate various platforms based on various criteria.
SUMMARY
This chapter reviewed the benefits and factors to consider when choosing a cloud-based AI/ML service. This
chapter explained the criteria to consider when choosing an AI/ML service from one of the major cloud service
providers—AWS, Google, and Azure.
It began by discussing the benefits of using cloud AI/ML services, including faster time to market, handling inten-
sive machine learning workloads, cost effectiveness, collaboration, integration capabilities, security compliance,
rich AI capabilities, tool support, and keeping up with innovations.
To assist you with the evaluation process, the chapter reviewed several factors, including identifying the business
problem, exploring service capabilities, reviewing performance and scalability needs, and considering develop-
ment and integration needs, security capabilities, price points, vendor support, and regulations.
It also examined various AI/ML services offered by the three major cloud providers and obtained a glimpse into
the capabilities and tools available for various use cases and industries.
In the following chapters, you continue your journey toward deploying enterprise AI in the cloud. You explore
data processing, building AI platforms, modeling pipelines, ML operations, and governance practices, and delve
into AI/ML algorithms to train models.
REVIEW QUESTIONS
1. Choose the correct answer regarding AI/ML services.
A. The ML technology was limited to a few companies in the past but is now spreading to more
companies due to the advent of AI/ML services.
B. The ML technology is still in the periphery and has not reached the core of business applications.
C. AI and ML are only popular with data scientists and not so much with other team members.
D. Cloud service providers are still working on providing the right set of AI and ML services
for projects.
2. Which of the following is a high-level AI service?
A. SageMaker
B. MXNet
C. Fraud detection
D. TensorFlow
3. Choose the correct machine learning algorithm.
A. SageMaker
B. Azure ML
C. Decision trees
D. Google Cloud ML Engine
4. Choose the AWS service that helps developers identify the phrases and entities in a document.
A. AWS Comprehend
B. AWS Rekognition
C. Amazon Textract
D. SageMaker
5. How can Amazon Polly be used in enterprise applications?
A. To develop IVR systems
B. To make applications more accessible to the visually impaired
C. To create lifelike voices in apps
D. To add audio to e-learning courses
6. What can AWS Panorama be used for in retail industries?
A. To personalize marketing campaigns
B. To track customer behavior and recommend products
C. To detect diseases proactively

D. To improve warehouse management
7. What type of filtering does Amazon Personalize not use?
A. Content-based filtering
B. Collaborative filtering
C. Role-based filtering
D. Community-based filtering
8. Which forecasting model in Amazon Forecast is suited for fast-changing situations and short-
term trends?
A. ARIMA
B. ETS
C. DeepAR
D. EDS Prophet
9. How can data scientists use Vertex AI?
A. Data visualization and exploration
B. Model development and deployment
C. Data preparation and cleaning
D. Code version control
10. What is the purpose of the AI Workbench?
A. Model development and deployment
B. Data preparation and cleaning
C. Rapid prototyping and model development
D. Collaborative model sharing and deployment
11. How can developers use Cloud Natural language?
A. Real-time translation of audio or video content
B. Conversion of speech into text
C. Analysis and understanding of human language in text format
D. Detection and analysis of objects, faces, and text in images
12. Which service automates the development of models and is meant for those with limited knowledge of
machine learning?
A. Cloud Natural Language
B. Dialogflow
C. AutoML
D. Translation AI
13. Models can be deployed in the cloud with scalable and distributed machine learning workflows using
which of the following?
A. Deep Learning Containers
B. Deep Learning VM Image
C. GPUs
D. TensorFlow Enterprise
14. What capabilities are offered by Azure Cognitive Services for Language?
A. Real-time analytics and intelligent search
B. Entity recognition and sentiment analysis
Answer Key ❘ 151
C. Speech-to-text and text-to-speech

D. Semantic search and faceted navigation
15. Which open-source data analytics tool includes a web extraction tool for extracting data from websites?
A. Dataiku
B. DataRobot
C. KNIME
D. IBM Watson
ANSWER KEY
1. A 6. B 11. C
2. C 7. D 12. C
3. C 8. B 13. A
4. A 9. B 14. B
5. B 10. C 15. C
Launching Your Pilot Project
10
The people who are crazy enough to think they can change the world are the ones who do.
—Steve Jobs
This chapter is a pivotal moment in your AI journey, where you transition from strategy to execution. You
are on the cusp of introducing AI/ML into your company operations.
Launching a pilot is no small feat, fraught with many challenges and requiring precise coordination, and
that’s where the methodology comes into play. This chapter assists you in navigating the essential stages of
launching, from initial planning to the nitty gritty details of a machine learning process.
This chapter brings all the strategy and vision discussed in the previous chapters to reality through a pilot
launch, starting from business objectives to model development, deployment, and monitoring. I make this
real through a hands-on project implementation exercise that bridges theory with practice. Ultimately, the
goal is to launch a successful pilot that makes your strategic vision a reality. See Figure 10.1.
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Your AI Govern Your
journey Team
09 Plan and Launch

ENTERPRISE AI 04
• Identify Use Cases for Your AI/ML Project
Evolve and Setup • Evaluate AI/ML Platforms And Services
Mature AI
and Manage • Launch Your AI Pilot
Operations

AI
Deploy and 05 Initiate a focused AI project ensuring alignment with
Monitor business objectives, technical feasibility,
Models and stakeholder engagement
07
06
FIGURE 10.1: PLAN AND LAUNCH: Launch your AI pilot

Launching Your Pilot ❘ 153
NOTE Understanding your company’s vision and goals while conducting a maturity
assessment helps identify gaps and enables you to create an AI transformation plan. This
in turn helps you move from pilot to production and make AI a core part of your business
strategy.
LAUNCHING YOUR PILOT

During the Align phase in Chapter 5, you learned that you need to conduct a maturity assessment of AI in the
company, which will help address the current state of AI across dimensions such as business, people, data/models,
platforms, governance, and security.
Trying to understand your company’s vision, business strategy, and goals and reviewing that against the current
state is important. This will help you identify gaps that need to be addressed when implementing a proof of con-
cept during the Launch phase.
Once you have completed the current state assessment as part of the Align phase, you should have a clear under-
standing of the current state of AI adoption. Your capabilities across AI strategy, talent capabilities, data manage-
ment, platforms, machine learning, operations, governance framework, data/infrastructure, and network security
should be clear.
Armed with this knowledge, you can create a roadmap with the list of initiatives, resource allocations, and fund-
ing to move forward from the Align phase into the Launch phase, and you can focus on launching a pilot.
Planning for Launch

Your focus is now on putting together a project plan to launch the pilot to implement a particular use case that
has been shortlisted through the prioritization of your AI initiatives. It will boil down to putting together a plan
to implement a machine learning project.
As a part of the maturity assessment exercise, you now have a list of action items across the business, people,
platform, operations, governance, machine learning, and security. Part of these action items are short-term and
focused on initiating the launch. In contrast, others are long-term focused on increasing the adoption of AI across
your entire organization. The long-term action items go into your AI transformation plan, while the short-term
items need to bake into your project plan focused on executing your pilot. I discuss the AI transformation plan in
greater detail in Chapter 23. Figure 10.2 shows some of the tasks needed to transition from envision to launch-
ing a pilot.
Recap of the Envision Phase

At this point, you have a list of projects for initiatives that you have identified working with various stakehold-
ers that constitutes your AI portfolio. You have also prioritized those initiatives and shortlisted one or two use
cases to implement as part of your pilot. In addition, you have defined a few KPIs that you will use to measure
the success of your launch when it is completed. Here are some additional activities that you can initiate during
this phase:
Change You have initiated some change-management activities to get leadership alignment
management and communicated the AI vision to the key stakeholders to get their support. You
activities have also established a smaller version of COE, a group of A-class players from various
departments working on the pilot as a core team.
Talent You also have to assist with any gaps in the talent required to implement the pilot, such
management as data scientists, machine learning engineers, or cloud engineers who will be required
to establish the platform.
154 ❘ CHAPTER 10 Launching Your Pilot Project
ML platform From an ML platform perspective, you have identified the steps required to set up the
AI platform and any gaps that need to be addressed during the launch.
Data capability You have assessed the current state of the data, identified any data gaps, and the
data infrastructure required to collect, ingest, clean, and process the data to feed
the models.
Governance Lastly, from a governance perspective, you have procured funding for this launch
by working with a sponsor, ideally an executive who understands the value of the
technology, is behind this launch, and is looking forward to ensuring its success.
Ethical You should consider the principles to be factored in during the design of the
considerations models to ensure trustworthy, ethical, and socially responsible implementation of
your AI model.
• Complete maturity model assessment

• Shortlist a use case for launch
• Identify short- and long-term action items 3
• Identify initiatives
t
ec
• Initiate change management initiatives
2
oj
Pr
• Form a small COE team
L
M
ch
un
La
1
n
lig
A
n
sio
• Develop ML project plan

vi
• Define project KPIs

En
• Define Responsible AI design principles
FIGURE 10.2: Activities to move from Envision to Align to launching a pilot
NOTE After completing the AI maturity model assessment, use the insights to develop a
roadmap and project plan for your AI launch.
Planning for the Machine Learning Project

Now that you have completed the Align phase, you have a clear set of action items to follow, and ideally, you
have a project plan to execute the launch of your pilot. With that in mind, the next section discusses the machine
learning lifecycle, some of the important considerations when taking up a machine learning project, and how
to develop the business goals, define the KPIs, and identify if machine learning is indeed the right solution for
the problem.
Following the Machine Learning Lifecycle ❘ 155
FOLLOWING THE MACHINE LEARNING LIFECYCLE

When implementing a machine learning project, you follow a certain lifecycle, which is an iterative process with
some defined steps that you should follow to make it a successful implementation.
You can break that cycle down into the following phases:
➤➤ Defining the business goal
➤➤ Framing the problem as a machine-learning problem
➤➤ Processing the data (collection, preprocessing, and feature engineering)
➤➤ Building the model (testing, tuning, and evaluation)
➤➤ Deploying the model (inference and prediction)
➤➤ Monitoring the model
Note that these phases are not necessarily sequential and that there are feedback loops between some phases, such
as problem definition and monitoring, data preparation and model building, and model development and model
monitoring.
The next sections briefly examine each of these phases shown in Figure 10.3, discuss in greater detail the business
goal, and frame the machine learning problem.
Phase 1 Phase 2 Phase 3 Phase 4 Phase 5 Phase 6 Phase 7 Phase 8

Identify Frame the Collect Data Preprocess Engineer Train, Tune, Deploy Monitor
Business ML Problem Data Features & Evaluate Model Model
Goals
FIGURE 10.3: Machine learning phases
Business Goal Identification

Business goal identification is covered in detail in Chapters 5 and 6. During this phase, you do the following:
Identify the First, you should have a clear idea of the problem you’re trying to solve and the
business problem business value you get by solving that problem.
Measure the You should be able to measure the business value against specific business objectives
business value and success criteria. The business objective is what you want to achieve, while the
business value is what you expect to achieve at the end of this implementation.
Machine Learning Problem Framing

During this phase, you do the following:
Identify the machine Your primary focus is to translate your business problem into a machine learning
learning problem type problem. You try to identify the type of ML problem, such as whether it is
a classification problem, regression problem, anomaly detection problem,
and so on.
Identify the label Given a dataset, you will identify one of the attributes as a label, which is the
attribute that needs to be predicted.
Define the performance Focus on optimizing the performance and error metrics.
and error metrics
Follow a defined machine learning lifecycle that includes defining business goals, framing the problem as a
machine learning problem, data processing, model building, model deployment, and model monitoring, with
iterative feedback loops between some of the phases.
Data Processing
Data processing is covered in greater detail in Chapter 15. Data must be input into a model during the build
process in the correct format.
Data collection This involves identifying the different sources for the data—internal and external—as
and preparation well as collecting data from multiple sources. You must also preprocess the data, such
as addressing the missing data problem, formatting data types, and addressing any
anomalies or outliers.
Feature Feature engineering involves creating new data, transforming data, extracting, and
engineering selecting a subset of the data from the existing set.
Model Development
Model development is covered in greater detail in Chapter 17. Model development involves the following:
Model building, tuning, and Building, testing, tuning, and evaluating the model before it can be
evaluation deployed into staging and production.
Setting up CI/CD pipelines for It also involves setting up a CI/CD pipeline for automatic deployment
automatic deployment into staging and production environments based on a set of criteria.
NOTE A CI/CD pipeline is a set of tools that automate the process of software deployment
from source code control to production deployment.
Following the Machine Learning Lifecycle ❘ 157
Model Deployment
Once the model has been tested, tuned, evaluated, and validated, it is ready to be deployed into production to
make inferences and predictions. Chapter 18 covers this.
Model Monitoring
Model monitoring involves continuously monitoring the model’s performance so that it performs well against a
particular set of desired parameters and proactively detects and mitigates performance degradation. Chapter 19
covers this.
TIP When a model performance tends to degrade, it is important to retrain the model on
new data that is representative of the current data distribution.
HANDS-ON PROJECT IMPLEMENTATION EXERCISE: PLANNING AND

LAUNCHING YOUR AI PILOT
Your ultimate goal is to implement AI effectively and scale it across the enterprise. However,
you need to start somewhere, and that’s the goal of this exercise—to help you plan and launch
your pilot project.
TASK DESCRIPTION TASK OWNERS DELIVERABLES
Identify Business Problem Business analyst, Business problem

Select and define a business project manager statement, objective
problem that can be addressed with
machine learning.
Refer to Chapters 6 and 7 regarding
business goal identification.
Change Management Project manager, Change

Develop a strategy to align leadership team management plan
leadership and stakeholders with
the AI vision.
Assess Capabilities and AI team, data AI maturity assessment,

Identify Gaps scientists gap analysis
Evaluate current AI maturity and
identify areas of improvement.
Create Project Plan Project manager Project plan, roadmap

Develop a project plan including
action items, resource allocations,
and roadmap.
Ethical Considerations AI team, ethical Ethical guidelines

Outline ethical principles oversight committee
for the model’s design and
implementation.
Refer to Chapter 4 for more details
regarding ethics.
Define Machine Learning Problem Data scientists ML problem

Translate the business problem into definition, metrics
a machine learning problem.
Data Processing Data engineers, data Processed data set

Collect, preprocess, and prepare scientists
data for model building.
Model Building and Tuning Machine Trained model,

Train, test, tune, and evaluate the learning engineers evaluation report
chosen machine learning model.
Set Up CI/CD Pipeline (Optional) DevOps team CI/CD pipeline setup

Create or simulate a CI/CD pipeline
for automatic deployment.
Deploy Model into Production DevOps team, Deployed model

Deploy the validated model into machine
staging or production environment. learning engineers
Implement Model Monitoring Data scientists, Monitoring plan,

Outline and implement a strategy operations team monitoring tools/setup
for ongoing monitoring of the
model's performance.
Reflect and Document Entire project team Reflection report,

Reflect on lessons learned, complete project
successes, and challenges, and documentation
document the process.
WORKBOOK TEMPLATE: AI/ML PILOT LAUNCH CHECKLIST

Download the “AI/ML Pilot Launch Checklist” template from the download section of the book (www.wiley
.com/go/EnterpriseAIintheCloud). Use this template to ensure that you have taken care of all the key
activities and milestones for your AI/ML project to ensure a successful pilot launch. It will help you to tackle
various challenges and tasks when launching an AI/ML project.
Answer Key ❘ 159
SUMMARY
This chapter covered the various steps you take as you move from the Align phase to the Launch phase. The
chapter discussed how to prepare for the launch of your machine learning project. It discussed the different
phases of a typical machine learning project: business goals, problem framing, data processing, model develop-
ment, deployment, and monitoring. The chapter also included a hands-on exercise to plan and launch an AI pilot.
In the next chapter, let’s discuss putting into motion some organizational change management steps to ensure
your people wholeheartedly embrace your AI plans.
REVIEW QUESTIONS
1. The focus of the core team implementing the AI pilot is
A. To establish the platform
B. To implement AI initiatives
C. To conduct a maturity model assessment
D. To establish a COE
2. Why do you need a label in machine learning problem framing?
A. To make predictions.
B. To evaluate the model.
C. It identifies the type of machine learning problem.
D. It is a feature created during feature engineering.
3. What is the primary focus of the machine learning problem framing phase in the machine learning
lifecycle?
A. Data processing
B. Model definition
C. Defining the business goal
D. Identifying the type of machine learning problem
4. Feature engineering is used
A. To evaluate the model’s features
B. To collect and preprocess features
C. To create new data and transform existing data
D. To identify the business data
5. Which phase involves translating a business problem into a machine learning problem?
A. Data processing
B. Model deployment
C. Machine learning problem framing
D. Business goal identification
ANSWER KEY
1. A 3. D 5. C
2. A 4. D
PART IV
Building and Governing Your
Team
People make magic happen! Part IV explores the organizational changes required to empower your
workforce. I guide you through the steps to launching your pilot and assembling your dream team. It’s all
about nurturing the human side of things.
11
Empowering Your People
Through Org Change
Management
Culture eats strategy for breakfast.
—Peter Drucker
This chapter is a pivotal step in the AI implementation methodology. By now, you have a good grip on the
AI cloud ecosystem, use cases, and AI strategy. But here comes the crucial part: ensuring your AI plans are
wholeheartedly embraced by your people (see Figure 11.1).
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Your AI Govern Your
journey Team
09 Build and Govern Your Team

ENTERPRISE AI • Empower Your People Through Org
Setup 04
Evolve and
Change Management
Mature AI
and Manage • Build Your Team
Operations
08 Objective
Foster a transformative organizational culture that
AI
Deploy and 05 champions AI, prioritizing people-centered
Monitor
Models approaches to ensure seamless AI adoption and
07 alignment with your business goals
06
FIGURE 11.1: BUILD AND GOVERN YOUR TEAM: Empower Your People Through Org Change
Management
164 ❘ CHAPTER 11 Empowering Your People Through Org Change Management
The chapter starts by exploring how to evolve your company culture to foster AI adoption, bring innovation, and
enable meaningful change. Knowing that technology alone won’t cut it, the methodology delves into strategies to
redesign your organization to enable agility and renewed thinking around adopting AI at the core. It is built on
three strategies.
➤➤ Evolving the culture
➤➤ Redesigning the organization
➤➤ Aligning the organization
This is where the key deliverables—such as the cultural evolution program and organizational redesign plan
discussed in this chapter—come in.
You need to get your entire organization aligned, starting from top executives to frontline staff, around AI adop-
tion. In a world where change is constant, mastering org change management is a must. This chapter helps you
with the insights and tools needed to lead change confidently so that your AI initiatives not only survive but
thrive. Let’s get started.
SUCCEEDING THROUGH A PEOPLE-CENTRIC APPROACH

It is the people who are going to decide whether the initiative is going to be a success. Getting them on board
with the initiative helps accelerate your AI journey by making them receptive to change while promoting a
culture of continuous learning and growth. The eight people-oriented aspects discussed here focus on enabling
the right culture, setting up suitable organizational structures, garnering leadership support, and transforming the
workforce.
Figure 11.2 shows the key issues you need to focus on from a people perspective.
➤➤ Evolving you culture
➤➤ Transforming your leadership
➤➤ Developing your cloud fluency
➤➤ Developing AI skills
➤➤ Transforming your workforce
➤➤ Accelerating your change
➤➤ Redesigning your organization
➤➤ Aligning your organization
Case Study: How a Company Transformed Itself with AI Using a People-

centric Approach
A retail company wanted to leverage cloud and AI technologies to improve its retail operations
and gain a competitive edge. Their gap assessment during the Align phase revealed that they
needed to address the needs and aspirations of the people to ensure a successful
implementation.
They Evolved Their Culture

Their gap assessment revealed that their organizational structure was very hierarchical and
siloed. This was an impediment to their adoption of cloud and AI technologies. To resolve this,
they decided to build a culture of collaboration and innovation. They achieved collaboration
by promoting cross-functional teams and encouraging innovation through a culture of
experimentation and learning.
Succeeding Through a People-centric Approach ❘ 165
They Transformed Their Leadership

They also realized they had to change their leadership to keep pace with the changing times
and adopt innovative technologies such as the cloud and AI. They brought in new leadership
that was well-versed with cloud and AI technologies. They could lead by example, inspire their
teams, and bring about innovation.
They Developed Their Cloud and AI Fluency

They realized that they needed to educate their workforce to leverage the latest developments
in cloud and AI technology. They offered training courses, reward programs, certifications, and
on-the-job training opportunities for their employees to learn new skills.
They Transformed Their Workforce

They realized that to implement and operationalize new AI projects, they had to transform
their workforce. They achieved this by assessing their current skillsets and competencies and
filling the gaps by reskilling their employees through training programs and career develop-
ment opportunities.
They Accelerated the Change Process

They noticed that the change process was terribly slow and that they had to speed up things.
They achieved this by creating a change acceleration framework. This involved implementing a
change management plan, communicating the change vision to the employees, soliciting
employee feedback, addressing their concerns, and communicating the progress periodically.
We discussed in great detail in Chapter 7 titled “Manage Strategic Change.”
They Redesigned Their Organization

They noticed that communication and decision-making could have been more efficient. They
addressed this by redesigning their organizational structure by creating cross-functional matrix
teams, changing the lines of reporting, and enabling decentralized decision-making.
They Aligned the Organization for Business Success

They brought their strategy, processes, systems, org structure, and culture together to increase
the probability of success. They aligned the projects with the business vision and strategy. They
allocated resources accordingly so that the organization was invested in the most important
priorities.
BEST PRACTICE TIP To successfully implement AI at the enterprise level, prioritize a

people-centric approach by evolving your culture, transforming leadership, developing cloud
and AI fluency, transforming your workforce, accelerating the change process, and redesign-
ing your organization.
As you can see, you need to take care of the people when implementing the cloud and AI at the enterprise level.
In the following two sections, you see how to systematically address people aspects using a comprehensive change
management methodology. As part of this methodology, I address the objectives, tasks, and deliverables for these
capabilities, such as cultural evolution, org redesign, org alignment, change acceleration, leadership transforma-
tion, workforce transformation, and cloud/AI upskilling.
Culture Evolution Transformational Leadership

Align your culture with digital Empower leaders to drive change and make
transformation best practices outcome-focused decisions
Organizational Alignment
Ensure technology change is driven by Cloud Fluency
business to drive business outcomes Develop and hire the skills and
People knowledge to use cloud
enablement
for enterprise
AI
Organizational Design
Design your organization structure to
Ai Fluency
implement AI strategy Develop and hire the skills and
knowledge to use AI technology
Change Acceleration Workforce Transformation

Structured approach using organizational Invest in people and create a skilled
alignment, leadership, and culture workforce receptive to change
FIGURE 11.2: Change management focus areas for enterprise AI
Note that you would also need to implement change management techniques when choosing the AI/ML plat-
forms and services. AI platforms come with their own training modules, may impact daily workflows, and can
act as catalysts for collaboration. AI may also require additional security controls. It is therefore important to get
the stakeholders’ buy-in when adopting a new AI/ML platform or service and adopt these change management
practices enabling growth and development.
Evolve Your Culture for AI Adoption, Innovation, and Change

As shown in Figure 11.3, Gartner reports in their research that transformations bring cultural tensions that fall
under these seven broad categories.1 They even found that many employees face more than one tension.
None of these
Other
Need for unity versus diversity at work
Consistency versus empowerment
Focus on compliance versus taking risks
Centralization versus local autonomy
Innovation or keep things running with efficiency
Commercial versus people focus
Deliver things quickly or focus on quality
FIGURE 11.3: Percentage of employees experiencing cultural tension due to change

Succeeding Through a People-centric Approach ❘ 167
Regardless of the type of stress, the fact is it is going to impact their performance. It is, therefore, important to
assess the organization’s current culture and how the people react to change. Based on this assessment, you must
take a certain number of steps, as shown in Figures 11.4 and 11.5. The focus of these steps is to bring in a change
that enables innovation, get them aligned on transformational values, and build an open culture to encourage tak-
ing risks. You should also provide resources to build their capability, and foster collaboration between teams.
TIP If you manage your company culture well, you will see greater employee participation,
higher motivation levels, and increased productivity, which translate into innovation and
customer satisfaction.
CURRENT STATE DEVELOP CULTURE

CULTURE ASSESSMENT COMMUNICATION PLAN
Conduct a current state Develop a cultural evolution
assessment of the culture and communication plan to get
n A
identify areas to improve for AI sio employee buy-in
lig
vi
adoption
En
n
Evolve Your
Culture
h
al
Sc
nc
MONITOR AND MEASURE e u IMPLEMENT CULTURE
La
EVOLUTION PLAN
Monitor the effectiveness of the
culture evolution plan and make Implement a cultural
necessary improvements to evolution plan to target
increase its effectiveness employee groups through
training programs,
workshops, and mentoring
FIGURE 11.4: Evolve your culture: phases and tasks
CURRENT STATE COMMUNICATION TRAINING MEASUREMENT

ASSESSMENT: The company developed a The company imparted The company evaluated the
A company found during their communication plan to training to both leadership and success of its culture evolution
assessment that employees socialize culture evolution plan frontline workers about AI. program through employee
were afraid of innovative to employees through town feedback and surveys. It made
technologies such as AI due to halls workshops and training the necessary adjustments to
MENTORING
potential job displacement and programs.
ACTIVITIES
improve the program by

lack of skills. They also created an internal The company paid employees improving training and
AI community to promote with experienced professionals mentoring.
RESULT collaboration and learning. to provide mentoring and on-
To address that, the company the-job training. RESULT
developed a cultural evolution The company was able to
program that focused on RESULT REWARDS sustain its culture evolution
increasing the awareness of AI efforts and sustain for the
The company got the They used rewards to
and training programs to build long-term resulting in
employees to buy into the encourage risk-taking and
their skills in AI/ML improvement in AI adoption
communication plan. innovation.
technology. across the company.
FIGURE 11.5: Evolve your culture: phases and deliverables

Redesign Your Organization for Agility and Innovation with AI

The organizational design aims to develop the appropriate structure for your organization to promote agility
and innovation and ensure the successful completion of your AI initiatives with appropriate roles, structures, and
responsibilities. Figure 11.6 captures the various steps and deliverables you can focus on for this track.
CURRENT ORG STRUCTURE PLAN IMPLEMENT ORG EVALUATE ORG

ORGANIZATIONAL Create a plan for new STRUCTURE PLAN STRUCTURES
STRUCTURE ASSESSMENT organizational structure, Implement the new Evaluate the success of the
Assess the current processes, and roles. organizational structure, new org structure and make
organizational structure processes, and roles. changes as required.
processes and roles for AI
adoption. USE CASE EXAMPLE
ACTIVITIES
To solve the issue of hierarchy USE CASE EXAMPLE USE CASE EXAMPLE
USE CASE EXAMPLE and lack of collaboration, you You may create a new AI The company reviews the
You may realize that your may decide to implement division with a dedicated team progress so far and makes
company is very hierarchical cross-functional teams with of data scientists, machine further changes to the
and siloed. appropriate roles and learning, data analysts, and structure, jobs, and roles.
responsibilities. machine learning engineers.
CURRENT JOBS EVALUATE JOB
JOB PLAN STRUCTURES
ASSESSMENT IMPLEMENT JOB PLAN
Create the plan for new job Evaluate the success of the
Assess the current job Implement the new job
descriptions and role profiles new jobs and make changes
descriptions and role profiles structure and role profiles.
for the new AI adoption. as required.
required for AI adoption.
FIGURE 11.6: Redesign your organization: tasks and deliverables
ALIGNING YOUR ORGANIZATION AROUND AI ADOPTION TO

ACHIEVE BUSINESS OUTCOMES
The focus of organizational alignment is to ensure that the business and technology teams are aligned toward
adopting AI to achieve business outcomes. It involves aligning strategy, processes, systems, culture, team struc-
tures, and talent. Achieving proper organizational alignment involves alignment between business strategies and
technology strategies, business strategy with organizational structure, and properly structured processes and
systems. You should also ensure project prioritization, cultural alignment, and the proper execution of the change
accelerated framework discussed in Chapter 7 titled “Managing Strategic Change.”
NOTE Chapter 5 titled “Envisioning and Aligning Your AI Strategy” discussed aligning the
organization around the AI vision and strategy. That theme continues to play throughout the
book. For example, the people-centric alignment plays an even more vital role at the ground
level when it comes to choosing an AI/ML platform or service. It only continues to increase
in importance as you scale your AI capabilities. You learn about this more in Part VIII of
this book.
Figure 11.7 shows different components that need to be aligned as part of ensuring organizational alignment.
Aligning Your Organization Around AI Adoption to Achieve Business Outcomes ❘ 169
Aligned Business and To ensure technology

01 Technology Strategy serves the business needs
Aligned Business Strategy To ensure everyone is working

02 and Org Structure toward the common goal
Structured Systems To ensure resources are allocated

03 appropriately
Organizational and Process
Alignment
To focus on what’s important
Aligning strategy, 04 Prioritized Projects for the organization
processes, systems,
team structures,
culture, and talent To ensure resources are working
05 Allocated Resources on the most important tasks
To create a supportive work

06 Aligned Culture environment and cope with
change
Accelerated Change To cope with change, speed

07 Framework up adoption, and maintain a
competitive edge
FIGURE 11.7: Organizational alignment
HANDS-ON EXERCISE FOR CULTURAL EVOLUTION AND ORG REDESIGN

TO ADOPT AI
This exercise gives you tangible experience in dealing with org change management practices in
the context of AI adoption.
Assess Cultural Tensions HR department, Cultural report

Identify cultural tensions using surveys, organizational
interviews, etc. design team
You can use Figure 11.3 to identify different

categories of tension.
Develop a Cultural Evolution Program Senior management, Cultural

Create a strategy to evolve culture focusing HR department evolution
on innovation, alignment, risk-taking, program
resource-building, collaboration. and strategy
document
Refer to Figures 11.4 and 11.5 for the steps,
phases, and tasks.
continues
continued
Create an Organizational Redesign Plan Senior management, Organizational

Assess current organizational structure for operations team redesign plan
agility and innovation.
Develop redesign plan for implementing AI.
Develop Alignment Strategies Senior management, Organizational

Create a comprehensive alignment plan strategic planning alignment plan
aligning all parts of the organization with AI. team
Implement and Monitor Senior management, Status reports

Implement the strategies. Continuously operations team,
monitor progress, making adjustments as organizational
necessary. development team
HANDS-ON ROLE PLAY EXERCISE TO INCORPORATE PEOPLE-CENTRIC

APPROACH TO AI (FOR STUDENTS)
Objective: Including people-centric elements to an AI implementation such as evolving the

culture, transforming the leadership, developing the workforce, and aligning the organization
around AI.
Instructions:
Step 1: Divide your team into groups of three to five with each group representing a particu-
lar business unit, such as HR, marketing, strategy, operations, and so on.
Step 2: Assign each group a specific role such as evolving culture, transforming leadership,
cloud/AI fluency, workforce transformation, change acceleration, organizational
redesign, and alignment.
Step 3: Each group should discuss challenges, opportunities, gaps, and strategies for each area
and present their findings and recommendations.
Step 4: Engage in collective discussions with all the groups to drive overall organizational
alignment.
WORKBOOK TEMPLATE: ORG CHANGE MANAGEMENT PLAN

Download the “Org Change Management Plan” template from the download section of the book (www.wiley
.com/go/EnterpriseAIintheCloud). Use this template to manage changes in your company’s processes,
systems, and people roles. It will help you to tackle the changes effectively and ensure successful adoption of AI.
SUMMARY
This chapter explained how companies can empower their people through organizational change management to
build a competitive advantage. The key takeaway from this chapter is that technology won’t be enough to bring
change and innovation at the enterprise scale.
The chapter started by understanding the different cultural tensions that might arise during major transforma-
tional efforts and devised strategies to address them. This chapter provided step-by-step instructions to foster
innovation, risk-taking, collaboration, and alignment with transformational values.
Remember that it’s important to align your organization at all levels to ensure that strategy, processes, systems,
culture, and teams work together to achieve the organizational goals. Prioritizing people and culture to drive
organizational change, continuous learning, growth, and innovation will translate to business success and cus-
tomer satisfaction.
In the next chapter, let’s discuss building a high performing team that can kickstart your AI transformation effort.
REVIEW QUESTIONS
1. Which of the following is not the only factor in the success of an AI project?
A. Leadership support
B. Employee training and development
C. Adequate resources
D. Change management plan
2. What are the challenges of evolving your culture evolution for AI adoption, innovation, and change?
A. Employees are afraid of adopting modern technologies.
B. Employees lack the necessary skills.
C. Employees are resistant to change.
D. It is difficult to measure the ROI of AI projects.
3. Which of the following is NOT one of the eight major areas of focus to build your competitive advan-
tage in an AI initiative?
A. Strategy and planning
B. People
C. Data
D. Location
E. Platforms
F. Operations
G. Security
H. Governance
4. What did the case study company identify as an obstacle to the adoption of cloud and AI technologies?
A. Lack of funding
B. Hierarchical and siloed organizational structure
C. Lack of market demand
D. Technological incompatibility
5. What does organizational alignment focus on in terms of AI adoption?

A. Ignoring the need for change management
B. Ensuring that the business and technology teams are aligned toward adopting AI to achieve busi-
ness outcomes
C. Focusing only on the technology strategy without considering business strategy
D. Replacing all business processes with AI technologies
ANSWER KEY
1. D 3. D 5. B
2. A, B, C 4. B
NOTE
1. www.gartner.com/en/corporate-communications/trends/diagnosing-cultural-tensions#:
~:text=During%20times%20of%20significant%20change,and%20helping%20employees%20build%20
judgment.
12
Building Your Team
Coming together is a beginning, staying together is progress, and working together is success.
—Henry Ford
In this chapter, you embark on the fundamental task of building a collaborative team essential to kick-
starting your AI project. The importance of assembling a synergistic team cannot be understated. This
chapter delves into the various AI/ML roles and responsibilities involved in an AI implementation, be it
the core roles such as data scientists and machine learning engineers or the equally crucial supporting roles
such as security engineers.
Given that AI is taking center stage and evolving so rapidly, even noncore roles are quickly becoming core.
What is essential, though, is to tailor these roles according to your specific business, project, and company
requirements, and that’s where this chapter comes in. You should also get good at transforming traditional
roles, such as system administrators and business analysts, to harness your company’s existing knowledge
base and expertise. This, coupled with judiciously tapping into external skill sets, can bridge any looming
skills gap.
Building a team is almost an art; the goal is to assemble a diverse, cross-functional team with clear roles
and responsibilities to drive effective AI transformation tailored to business objectives. See Figure 12.1.
This dream team will go on to steward your AI operations, starting from data processing in Chapter 15,
model monitoring in Chapter 19, and managing the ethics of AI in Chapter 20.
UNDERSTANDING THE ROLES AND RESPONSIBILITIES IN AN ML

PROJECT
This section reviews the roles and responsibilities that you need in an ML project. As part of this selection
process, the following sections outline some important points to consider.
Build a Cross-Functional Team for AI Transformation

As part of an AI transformation initiative, one of the essential things that you must do, regardless of what
role you play, is identify a team that can help you with not only the overall transformation but also deploy-
ing the AI prototype as part of the Launch phase.
174 ❘ CHAPTER 12 Building Your Team
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Your AI Govern Your
journey Team
09 Build and Govern Your Team

ENTERPRISE AI 04
• Empower Your People Through Org
Evolve and Setup Change Management
and Manage • Build Your Team
Operations
08 Objective
Assemble a diverse, cross-functional team with
AI
Deploy and 05 clear roles and responsibilities to drive effective AI
Monitor
Models transformation tailored to business objectives
07
06
FIGURE 12.1: BUILD AND GOVERN YOUR TEAM: Building your team
When implementing your first enterprise-grade AI/ML project, you will face some challenges in identifying the
right resources. This is because the ML platform has many components, and you need to bring together skill sets
from multiple organizational units in your company. It includes data science, data management, model building,
governance, cloud engineering, functional, and security engineering skills.
Adopt Cloud and AI to Transform Current Roles

You need to assess how the cloud can potentially free current staff to focus on innovation. This, in turn, will drive
how their job roles will change when you migrate the ML platform and AI applications to the cloud.
Because the cloud brings with it the benefits of speed, scale, productivity, and innovation, the team can focus on
larger, more impactful, innovative, and transformational efforts, now and in the future.
Speed comes from the ability to experiment with multiple models, and scale leads to the ability to serve a broader
set of customer segments, geographies, and channels. Adopting infrastructure as code, code deployment pipelines,
and security as code enables better compliance and results in higher productivity from automation.
Some current on-prem roles include IT solutions architect, systems administrator, network administrator, and
desktop administrator. The applications administrator and database administrator will likely have to be reskilled/
redesigned. For example, the system administrator previously responsible for managing the onsite hardware and
infrastructure has to be relieved of some of these responsibilities, as the cloud provider (such as AWS, GCP, or
Azure) now manages them.
TIP Create a cross-functional team with the appropriate skill sets, while also adapting exist-
ing job roles as needed to align with your company and project goals.
Customize Roles to Suit Your Business Goals and Needs

Note that the actual job roles and responsibilities vary based on the company, the project goals, and other factors
appropriate for a given initiative. You’re going to realize that these roles and responsibilities are not very clearly
Understanding the Roles and Responsibilities in an ML Project ❘ 175
defined, and there is often overlap between these roles. I recommend you use this as a guideline but then custom-
ize it so you align your specific needs based on the skill sets that you have in your organization and the hiring
plan that you may have to initiate to fill the gaps.
NOTE Most AI initiatives have a higher risk of failing if they don’t have the right skills.
The following sections review the new roles in the cloud-based AI scenario. Let’s first look at the core team and
then the extended teams they need to work with. Figure 12.2 summarizes the roles involved in a typical machine
learning project.
Core Roles Auxiliary Roles
• AI Architect • Domain Expert

• Data Scientist • Data Analyst
• Data Engineer • Model Risk Manager
• Machine Learning Engineer • Security Engineer
• ML Ops Engineer • Software Engineer
• Business Analyst • Model Owners
• AI Ethicist • IT Auditor
• IT and Cloud Engineers • And So On
The actual composition of your team and the definition of core versus auxiliary depends on your
specific use case and project needs.
FIGURE 12.2: Core and auxiliary roles in a machine learning project
Core AI Roles
Several roles are required on an AI project. Depending on the type of AI project, some roles may be critical for a
project’s success and in other cases may not. In general, the following roles can be considered core roles. Use this
for guidance only, as the actual roles may vary based on your specific needs.
➤➤ AI architect
➤➤ Data scientist
➤➤ Data engineer
➤➤ Machine learning engineer
➤➤ MLOps engineer
AI Architect
The AI architect is an emerging role primarily focused on defining the transformational architecture that AI intro-
duces. You can think of AI architects as the glue between the business stakeholders, the data engineers, the data
scientists, the machine learning engineers, DevOps, DataOps, MLOps, and others involved in an AI initiative.
They are responsible for the following:
TASK DESCRIPTION
Designing the overall AI strategy Identifying the AI application opportunities and use cases, as well as
the underlying business and technical requirements to enable those use
cases. Includes developing the AI roadmap.
Designing the AI architecture Selecting the right AI technologies and ensuring the right data quality
and security.
Defining the overall framework Ensuring that the new AI technology is integrated with the existing
for building and deploying AI business processes and applications.
applications
Data Scientist
At its core, the data scientist role is about taking in data and refining it by creating and running models and
drawing insights to make business decisions. It is a critical role to ensure that the project is successful. Defining
the problem is one of the first things this role must do by working with the business stakeholders.
Data Engineer
The data engineer works closely with the data scientist, to whom they provide the data. They focus primarily on
the following:
➤➤ Data acquisition, integration, modeling the data, optimization, quality, and self-service
➤➤ Ensuring that the data pipeline is appropriately implemented and is working satisfactorily so that the
models get the correct data in the proper format and quality
Machine Learning Engineer

The machine learning engineer role is another crucial role for an AI project, and there is some overlap between a
data scientist role, but there are also some subtle differences.
While data scientists focus more on drawing insights from the models, the machine learning engineer focuses on
deploying models into production based on clean data from the data engineers.
MLOps Engineer
The MLOps engineer builds and manages automation pipelines to operationalize the ML platform and ML
pipelines for fully/partially automated CI/CD pipelines. These pipelines automate building Docker images, model
training, and model deployment. MLOps engineers also have a role in overall platform governance, such as data/
model lineage, infrastructure, and model monitoring.
Cloud Engineer
Cloud engineers aim to ensure the cloud infrastructure is secure, reliable, and cost-effective. They are responsible
for the following:
RESPONSIBILITY DESCRIPTION
Setting up the cloud This includes configuring virtual machines, setting up databases, and storage. It
infrastructure may also include configuring cloud-based services such as Amazon SageMaker,
Azure Machine Learning Studio, and Google Cloud platform.
RESPONSIBILITY DESCRIPTION
Setting up the accounts This includes creating and managing user accounts, setting up access controls, and
and working with security working with security teams to ensure the infrastructure is secure.
Managing the security and This includes monitoring and enforcing security policies, managing user
access controls permissions, and responding to security incidents.
Monitoring the This includes monitoring resource usage, identifying bottlenecks, and
performance of the system implementing cost-saving strategies.
Business Analyst
Business analysts engage early in the ML lifecycle and translate business requirements into actionable tasks. Their
primary role is to bridge the gap between business stakeholders and the technical team.
DATA SCIENTISTS VS. BUSINESS ANALYSTS
Data scientists are responsible for collecting, cleaning, and analyzing data. They use their
knowledge of statistics and machine learning to develop models that can be used to make
predictions or decisions. Business analysts are responsible for understanding the business needs
and requirements for an AI/ML project. They work with data scientists to ensure the project
meets the business’ goals.
Data scientists typically have a strong mathematics, statistics, and computer science back-
grounds. Business analysts typically have a strong background in business, economics, and
communication.
Data scientists are typically more focused on the technical aspects of AI/ML, while business
analysts are typically more focused on the business aspects. However, both roles are essential
for the success of an AI/ML project.
Business analysts need to understand the business requirements and translate them into technical specifications
that the developers and data scientists can use to build the AI/ML system. They also need to analyze the data and
provide insights to the stakeholders on how the system is performing and how it can be improved.
Use Case Example
Let’s say you are working as a business analyst on an AI/ML project for a telecom company. One of
the company’s fundamental business problems is customer churn when customers cancel their
subscriptions and switch to a competitor’s service. The company wants to use AI/ML to predict which
customers will likely churn so they can take measures to retain them.
As a business analyst, your first step is to gather the requirements from the stakeholders. You need to
understand what data sources are available, what business rules need to be applied, and what the
system’s output should look like. You then work with the data scientists and technical team to design
the system architecture and define the data processing pipeline.
Once the system is built, you must work with the data scientists to analyze the data and provide
insights to the stakeholders. You need to monitor the system’s performance and recommend improve-
ments. For example, specific customers are more likely to churn if they experience network issues, so
you should recommend that the company invest in improving its network infrastructure in those areas.
Overall, as a business analyst in an AI/ML project, your role is crucial in ensuring that the AI/ML system is
aligned with the business objectives and that the stakeholders are informed about the system’s performance and
potential improvements.
Noncore or Support Roles

In addition to the core roles, several roles are not critical for the success of an AI project but can facilitate the
process. Please note that as the adoption of AI matures in organizations, some of the support roles are turning
into core roles.
Domain Expert
Domain experts have valuable functional knowledge and understanding of the environment where you implement
the ML solution. They do the following:
RESPONSIBILITIES DESCRIPTION
Requirements gathering Help ML engineers and data scientists develop and validate assumptions and
hypotheses.
Design involvement Engage early in the ML lifecycle and stay in close contact with the engineering
owners throughout the Evaluation phase.
Industry compliance Provide valuable insights and help ensure the project adheres to industry
standards and best practices.
Domain experts have a deep understanding of their respective fields and are essential to any project involving spe-
cialized knowledge. With their expertise, projects can succeed, significantly saving time and money and minimiz-
ing reputation losses. Therefore, a domain expert is critical for any project requiring specialized knowledge.
Domain experts are especially important for healthcare, finance and banking, retail and e-commerce, manufactur-
ing, energy, agriculture, real estate, and so on.
NOTE Not having a domain expert can result in models being created in a vacuum, thus
missing critical industry requirements and resulting in poor model performance.
Data Analyst
A data analyst is a professional who does the following:
Data collection Collects, cleans, analyzes, and draws insights from the data.
and analysis
Business interpretation Helps interpret that data for the business stakeholders to make business
decisions.
Exploratory data analysis Conducts exploratory data analysis to identify patterns and trends and uses data
visualization tools to help interpret the data for the business stakeholders.
Collaboration with data In the context of enterprise AI, may also work with data scientists to select and
scientists preprocess data.
Model testing and Tests and validates the models to ensure they are reliable and accurate.
validation
NOTE Data scientists have a broader scope and dive deep into advanced computations and
predictive analytics, while data analysts are focused on deriving insights from data to guide
business decisions. In small organizations, the roles can be played by the same person.
Model Risk Manager

The primary responsibility of a model risk manager is to ensure that the risks associated with the model are iden-
tified and mitigated. They must ensure that the model is reliable, accurate, and meets regulatory requirements.
Model risk managers ensure the following:
Responsible Ensures the model is deployed responsibly and ethically, protecting the company’s
deployment brand image, and preventing reputational risk.
Data management Ensures the data used to train the model is free from bias, is accurate, and
is complete.
Model testing and Ensures the model is adequately tested and validated before deploying in
validation production by testing it against several scenarios so that it is robust and can
perform well even in unexpected situations.
Compliance and Ensures the model is transparent, explainable, and complies with legal
transparency requirements by working with legal and compliance teams.
Security Engineer
The role of the security engineer is to ensure the security of the AI systems end to end, starting from managing the
data, the models, the network, and the infrastructure. They need to do the following:
Implement Do so at all layers of the AI system, starting from data preparation, model training,
security controls testing, and deployment stages.
Risk management Identify and assess the security risks, implement controls to mitigate risks, and
assess the effectiveness of those controls.
Incident response Respond to security incidents such as ransomware attacks, address data breaches
in real-time, identify the root cause, and implement steps to remediate those risks.
Legal compliance Ensure that the AI infrastructure is legally compliant concerning data privacy and
security by working with legal and risk teams.
Model security Manage the security of machine learning models by ensuring no tampering with
training data and no bias introduced into the data.
Software Engineer
The role of a software engineer is to do the following:
Integration and interface Integrate the AI algorithms/their output into other back-end applications and
development develop front-end user interfaces for users to interact with the models.
Deployment Deploy AI solutions to production and manage different software versions.
Performance and security Ensure these solutions are performant, reliable, scalable, and secure.
Maintenance Continue monitoring the AI solution in production, identify bugs, solve them
when required, and add new features.
Model Owners
The model owner is a new role primarily responsible for the model’s development, maintenance, governance, and
performance. The model owner is typically a technical person who works with the business stakeholders. They
need to do the following:
Safety and responsibility Ensure that the model can perform safely and responsibly.
Business problem Understand the business problem the model is supposed to solve.
understanding
Performance metrics definition Define the performance metrics by which the model will be evaluated.
Data selection and Select the correct data to train the model upon and ensure there is no bias
bias checking in the data.
Algorithm selection and Work with the data scientists to choose a suitable algorithm, configure it,
deployment and deploy it in production.
Model monitoring Monitor the model’s performance to ensure it’s working accurately with
fairness and transparency.
Model maintenance Maintain the model by considering new data, retraining it, and addressing
any issues.
Model Validators
Model validators are responsible for the following:
Model testing and Test and validate the models to ensure they work accurately and responsibly
validation while meeting the business’ objectives and requirements.
Error and bias check Ensure that the models are working without any errors and bias and with
fairness, which may include designing tests for various scenarios and conducting
statistical analysis to check for biases or limitations in the model.
Legal compliance and Ensure that the models comply with legal requirements and capture
documentation documentation about how these models work so they can share it with other
stakeholders, auditors, and senior management.
Model improvement Provide recommendations for improving the model.

recommendations
IT Auditor
The IT auditor is responsible for analyzing system access activities, identifying anomalies and violations, prepar-
ing audit reports for audit findings, and recommending remediations. Their areas of focus include model govern-
ance, data integrity and management, AI applications security, ethical considerations, AI systems performance and
scalability, and continuous monitoring.
HANDS-ON ROLEPLAY EXERCISE: ADOPTING NEW ROLES IN AN AI PROJECT
This exercise is to help everyone understand their new roles better and learn how these roles
collaborate with each other.
TASK DESCRIPTION TASK OWNERS (JOB DELIVERABLES

FUNCTIONS)
Introduction: Brief participants Facilitator Understanding of AI/ML

on the importance of AI/ML roles and importance
team and roles.
Role identification: List all Facilitator Visual

required roles on Post-It notes. representation of roles
Skill mapping: Map primary Participants, facilitator Skills identified

skills required for each role. for each role
Team assembly simulation: Participants in groups Selected candidates

Simulate a hiring panel to select for each role
candidates for each role.
continues
continued
TASK DESCRIPTION TASK OWNERS (JOB DELIVERABLES

FUNCTIONS)
Discussion and justification: Participants in groups, Agreed candidates,

Discuss and justify the chosen facilitator understanding of
candidates. hiring criteria
Role interrelation activity: Participants in groups, Understanding of team

Understand how roles interact facilitator dynamics and collaboration
through a hypothetical scenario.
Document key learnings: Facilitator, possibly an Documented insights to

Document insights and feedback HR representative guide future hiring
for future reference.
WORKBOOK TEMPLATE: TEAM BUILDING MATRIX

Download the “Team Building Matrix” template from the download section of the book (www.wiley.com/go/
EnterpriseAIintheCloud). use this template to build a skills and roles matrix for building your AI team. It
will help you have all the necessary roles and skills needed to implement your AI project successfully.
SUMMARY
The chapter discussed the roles and responsibilities you need to fill for your project to take off and ensure the
efficient development, deployment, and management of your AI systems. It also discussed the core roles and other
noncore roles that you need to fill depending on the specific needs of your project. The discussion of these roles
should help you build a blueprint for assembling a team of professionals who can navigate the complexities of AI
development by taking on responsibilities and driving innovation. Remember to customize these roles according
to your company’s needs.
Understanding of these roles and how they interact with each other will help you plan the team structure, com-
municate the vision of the AI system, and ensure that the AI system is robust, secure, fair, and legally compliant.
In the next chapter, we will discuss setting up a robust enterprise AI cloud platform that will help you take your
model from ideation to production quickly with the least amount of risk.
REVIEW QUESTIONS
1. The primary responsibility of a machine learning engineer is
A. Running models and drawing insights
B. Ensuring the data pipeline is appropriately implemented
C. Designing the overall AI strategy
D. Deploying models into production
Answer Key ❘ 183
2. The primary role of a business analyst is

A. Collecting, cleaning, and analyzing data
B. Developing and validating assumptions and hypotheses
C. Bridging the gap between business and technical teams
D. Developing the project plan
3. What task does a model owner NOT perform?
A. Algorithm selection and deployment
B. Model maintenance
C. Model testing and validation
D. Performance metrics definition
4. Who ensures that the AI models comply with legal requirements and capture documentation?
A. IT auditor
B. Security engineer
C. Model validators
D. Software engineer
5. Which role is primarily responsible for the model’s development, maintenance, governance, and
performance?
A. Security engineer
B. Model validators
C. Model owner
D. IT auditor
ANSWER KEY
1. D 3. C 5. C
2. C 4. C
PART V
Setting Up Infrastructure and
Managing Operations
In this part, you roll up your sleeves and get technical. Part V is like your DIY guide to building your own
AI/ML platform. Here, I discuss the technical requirements and the daily operations of the platform with a
focus on automation and scale. This part is a hands-on toolkit for those who are hungry to get geeky.
13
Setting Up an Enterprise AI
Cloud Platform Infrastructure
Design is not just what it looks like and feels like. Design is how it works.
—Steve Jobs
In this chapter, you get your hands dirty by setting up an enterprise AI cloud platform infrastructure (see
Figure 13.1). By now, you should have a good idea of the strategic vision, have aligned with the stakehold-
ers on the AI/ML use cases and the business metrics, implemented change management processes, built a
team, and may even have a good idea of the ML algorithm that you want to adopt.
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Govern Your
Your AI
Team Setup Infrastructure and
journey
09 Manage Operations
ENTERPRISE AI
Setup 04 • Setup Enterprise AI Cloud
Evolve and JOURNEY MAP Infrastructure
Mature AI Platform Infrastructure
and Manage
Operations • Operate Your AI Platform with
Automation for Scalability and Speed
08 Process Data
Scale and Objective
AI
Deploy and 05 Design and implement a robust, scalable AI
Monitor
Models cloud infrastructure tailored to specific use
07 cases, while weighing the key considerations
of build-vs-buy and multi-cloud strategies
06
FIGURE 13.1: Setting up the enterprise AI cloud platform infrastructure

188 ❘ CHAPTER 13 Setting Up an Enterprise AI Cloud Platform Infrastructure
Building a production-ready, robust enterprise AI cloud platform infrastructure is important because it helps you
take your models from ideation to production in the shortest possible time and with the least risk possible.
This chapter is not just about the theory; it’s about integrating AI into business processes to automate, reduce
costs, increase efficiency, and create new opportunities. To do this, you must create efficient models and build a
repeatable, robust, stable, and secure enterprise AI system adaptable to changing conditions. The right infrastruc-
ture will serve as a solid foundation for serving models with good performance, scalability, and reliability and as a
launchpad to transition from pilot projects to full-blown AI operations. See Figure 13.1.
In this chapter, you
➤➤ Review the typical reference architectures (AWS, Azure, and the Google Cloud Platform) for the most
common use cases
➤➤ Review the general components of an ML platform
➤➤ Learn how to choose between building and buying a ready-made ML platform
➤➤ Learn how to choose between different cloud providers
➤➤ Review hybrid and edge computing and the multicloud architecture
REFERENCE ARCHITECTURE PATTERNS FOR TYPICAL USE CASES

There are close to 200 services provided by AWS and this number keeps growing. This becomes challenging for
companies attempting to choose the correct tools for the right problem and map them to various use cases. This
section reviews a few reference architecture patterns so you can use them to customize your implementation
based on your specific project needs. This can help you create your future-state strategy and gain insights from
big data processing, data warehousing, visualization, and AI/ML to prepare for the future. Those architectural
patterns are related to the following five use cases:
➤➤ Customer 360-degree architecture
➤➤ Event-driven architecture with IoT data
➤➤ Personalized recommendation architecture
➤➤ Real-time customer engagement
➤➤ Data anomaly and fraud detection
Customer 360-Degree Architecture

Customer 360-degree architecture can be used to increase customer retention and profitability by improving
customer service, increasing sales, and improving marketing campaigns. Building a 360-degree customer view
involves aggregating data from multiple data sources into a single unified solution, from which you can draw the
necessary insight about your customers.
Building this architecture, as shown in Figure 13.2, involves four steps.
STEP STEP NAME DESCRIPTION POSSIBLE AWS

NUMBER COMPONENTS
1 Data sources Ingest the data from multiple sources into AWS Data Migration Service,
a single data source. Data can come from AWS Glue
SAP, ERP, CRM, websites, Salesforce, Google
Analytics, and so on.
Reference Architecture Patterns for Typical Use Cases ❘ 189
STEP STEP NAME DESCRIPTION POSSIBLE AWS

NUMBER COMPONENTS
2 Knowledge Build a unified customer profile by Amazon Neptune

graph system implementing a graph-based knowledge
system to identify new customer segments
and build customer hierarchies.
3 Intelligence Build an intelligence layer by offloading the Amazon Redshift, Amazon

layer data into Amazon RedShift or S3 so that S3, Amazon Athena
insights can be drawn using analytical tools
such as Amazon Athena.
4 Activation Use the insights from the intelligence AWS Lambda, Amazon
layer layer and enrich it with recommendations Personalize
or predictions from AI/ML to create the
next best action and personalize customer
experiences.
DATA
SOURCES
• CRM DATA DATA
• ERP CONNECTED CUSTOMER 360
INGESTION HARMONIZATION
• SFDC
• Social media • Transforms,
• Rest API
• Mobile • Data Lake • Catalog Aggregate
• Streaming
• POS • Data • Data Discovery • Deduping,
• ETL (Extract,
• MDM Warehouse • Data Lineage Entity Linking
Transform,
• Sensors • Feature Store • Data Quality • Ontology
Load)
• Chat • Graph Database • Data Lifecycle Creation,
• FTP
• Call center Refinement
• Images
• Videos INTELLIGENCE/
• Emails CONSUMPTION
ACTIVATION
• Third-party
data (e.g.,
Identity Access Management • SQL Query
weather,
Key Management Service • Data • Digital
futures,
Secret Manager Warehouse Channels
commodities,
etc. Logging • Business • Call Center
Auditing Intelligence • Data Exchange
CI/CD Stack • AI/ML
FIGURE 13.2: Customer 360-degree architecture
Figure 13.3 shows an implementation of the customer 360-degree architecture using AWS components. You
should be able to build similar architecture using Google or Azure Cloud components.
NOTE The customer 360-degree architecture has four steps: data ingestion, knowledge
graph system, intelligence layer for insights, and activation layer for personalized customer
experiences.
DATA
CONNECTED CUSTOMER 360
SOURCES DATA DATA
• CRM INGESTION • Data Lake (S3) HARMONIZATION
• ERP • Data • Transforms,
• SFDC • Rest API Warehouse • Catalog Aggregate
• Social media • Streaming (Amazon • Data Discovery • Deduping,
• Mobile • ETL (Extract, Redshift) • Data Lineage Entity Linking
• POS Transform, • Feature Store • Data Quality • Ontology
• MDM Load) (Amazon • Data Lifecycle Creation,
• Sensors • FTP Neptune) Refinement
• Chat • Graph Database
• Call center (Feature Store)
• Images
• Videos
• Emails INTELLIGENCE/
CONSUMPTION
ACTIVATION
AWS Identity Access Management • SQL Query (Amazon

AWS Key Management Service Athena)
• Digital
AWS Secret Manager • Data Warehouse
Channels
Logging (Amazon CloudWatch) (Amazon Redshift)
• Call Center
Auditing (AWS CloudTrail) • Business Intelligence
• Data Exchange
CI/CD Stack (Amazon QuickSight)
• AI/ML
FIGURE 13.3: Customer 360-degree architecture using AWS components (AWS components shown in bold)
EXERCISE: BUILDING A CUSTOMER 360-DEGREE VIEW

ARCHITECTURE USING AWS
This is an experiential learning exercise that allows the team members to apply their theoreti-
cal knowledge of AWS services into practice and acquire AWS skills.
STEP NAME TASK OWNER TASK DETAILS DELIVERABLE
Set up data Data engineer Create mock data that Mock data files
sources represents different
data sources (CRM
system, web analytics,
etc.)
Data ingestion Data scientist Ingest the data from AWS Data Migration
multiple sources into a Service or AWS Glue
single data source on configuration
AWS using AWS Data
Migration Service or
AWS Glue.
Build Data scientist Create a Graph Amazon Neptune

knowledge DB using Amazon DB setup, relations
graph system Neptune. Load the explored
ingested data into
this DB and explore
the relations between
different data entities.
Build Machine Offload the data into Amazon RedShift or

intelligence learning Amazon RedShift S3 setup, insights
layer engineer or S3. Draw insights drawn using Amazon
from the data using Athena
Amazon Athena.
Build Machine learning Enrich the insights AWS Lambda

activation engineer and from the intelligence and Amazon
layer data scientist layer with predictions Personalize setup,
or recommendations recommendations, or
from AI/ML to create predictions generated
the next best action
and personalize
customer experiences.
Use AWS Lambda and
Amazon Personalize
for this.
Report AI architect Document the Report with

preparation architecture necessary details and
built, problems screenshots
encountered,
solutions
implemented, and
insights drawn from
the customer data.
Include screenshots of
the AWS console.
Develop an Event-Driven Architecture Using IoT Data

The event-driven architecture shown in Figure 13.4 is focused on getting value by making near real-time decisions
from the data. Some of these use cases are as follows:
➤➤ Data from medical devices used to provide personalized healthcare
➤➤ Data from equipment to prevent equipment failures or to impact employee training
➤➤ Data from autonomous vehicles to improve safety
DATA DATA DATA

DATA TRANSFORM INSIGHTS
SOURCES COLLECTION INGESTION
• IoT Stream
Devices
• Data Ingestion • Data Streaming
• Medical
• Edge ML from Edge • Data Analytics
Equipment
Telemetry Inference • Asset Model • StreamEnrichment and
• Car Sensors
data View Anomaly Detection • Event
• Industrial
IoT Sensors Triggers
• AI/ML
Telemetry data Inferences
Log Analytics
STORAGE
• Reporting
Identity Access Management

Key Management Service • Batch
Secret Manager Processing
Logging • Data Lake
Auditing • Batch Analytics
CI/CD Stack
FIGURE 13.4: Event-driven near real-time predictive analytics using IoT data (AWS components
shown in bold)
DATA DATA DATA

DATA TRANSFORM INSIGHTS
SOURCES COLLECTION INGESTION
• IoT Stream • AWS Kinesis Data • Event
• Data Ingestion
Devices Streams Triggers
• Edge ML from Edge
• Medical • AWS Kinesis Data (AWS
Inference (AWS IoT Core)
Equipment Analytics Lambda)
Telemetry (AWS IoT • Asset Model
• Car Sensors • StreamEnrichment and • AI/ML
data Greengrass) View (AWS IoT
• Industrial Anomaly Detection Inferences
SiteWise)
IoT Sensors (Amazon Redshift) (Amazon
Telemetry data SageMaker)
• Log Analytics
STORAGE (AWS
OpenSearch)
• Reporting
Identity Access Management
Batch Processing (AWS
Key Management Service
• Data Lake (S3) OpenSearch
Secret Manager
• Batch Analytics Dashboard)
Logging
(AWS Kinesis
Auditing
Data Firehose)
CI/CD Stack
FIGURE 13.5: IoT-based event-driven predictive analytics using AWS components (AWS components
shown in bold)
Figure 13.5 shows the IoT-based event driven architecture using AWS components. The following table shows the
steps involved in setting up an Internet of Things (IoT)–based architecture that facilitates an end-to-end journey—
from data collection to insight generation.
STEPS DESCRIPTION AWS SERVICES
Data collection Data from IoT devices, such as medical devices, car AWS Greengrass
sensors, and industrial equipment is collected.
Data ingestion The collected data is then ingested into the cloud AWS IoT Core, AWS IoT
platform. SiteWise
Stream transformation The streamed data from these IoT devices is Amazon Kinesis Data
transformed using data from a data warehouse. Analytics
Data storage The transformed data is then stored for further Amazon Redshift
downstream analysis.
Model training and The model is trained and deployed to generate Amazon SageMaker
deployment inferences.
Event triggering Inferences from the models trigger certain events. AWS Lambda
Dashboard creation Dashboards are created to visualize insights. Amazon OpenSearch

Service Dashboards
TIP IoT architecture leverages streaming data from IoT devices to make real-time decisions
for use cases such as personalized healthcare, autonomous cars, and equipment monitoring.
EXERCISE: BUILDING AN EVENT-DRIVEN ARCHITECTURE USING AWS

AND IoT DATA
This is an experiential learning exercise that allows the team members to design and implement
an event-driven architecture that collects, processes, and analyzes data from IoT devices in
real time.
Prerequisites: You need a basic knowledge of AWS services, IoT devices, and some data sci-
ence concepts.
Tasks: Use the steps outlined in the previous table to build this architecture. The details of
building this architecture is outside the scope of the book, but it would be a good starting
point if your use case involves processing IoT data.
Personalized Recommendation Architecture

Figure 13.6 shows a personalized recommendation architecture, and Figure 13.7 shows the architecture with
AWS components.
In Figure 13.7, the first step is to gather click views and item view data from websites and upload them into
Amazon S3. The data is then cleaned using Amazon Glue DataBrew.
PERSONALIZE ENGINE
1 DATA 2 3 4 EVENT
Data Files DATASET Models
Upload LAKE Input Create Recommended TRIGGER
Data Preprocessed Models List
Output
Input
Preprocessed
Data
Output
Files
5
PREPROCESSOR CLIENT Get Real-Time
API GATEWAY
Recommendation
FIGURE 13.6: Personalized recommendation architecture
PERSONALIZE ENGINE
Amazon Personalize
EVENT
1 DATA LAKE 2 3 Model 4 TRIGGER
Data Files DATASET
Upload S3 Input Create (CAMPAIGNS) Recommended AWS
Data Preprocessed Models List Lambda
Output
Input
Preprocessed
Data
Output
Developer A Files
PREPROCESSOR
5 Amazon
AWS Glue CLIENT Get Real-Time API GATEWAY
DataBrew
Recommendation
FIGURE 13.7: Personalized recommendation architecture using AWS components (shown in bold)
The second step involves training the model in Amazon Personalize to make real-time personalization recommen-
dations. In this step, you use three types of data.
➤➤ Event data, such as the items the user clicks on, purchases, or browses
➤➤ Item data, such as the price, category, and other information in the catalog
➤➤ Data about the users, such as the user’s age, location, and so on
The next step is to feed this data into the Amazon Personalize service and get a personalized private model hosted
for you through the Amazon API Gateway. You can then use this API to feed the recommendations to your users
in your business application.
NOTE To train a personalized model, you must gather event data, item data, and user data.
EXERCISE: BUILDING A PERSONALIZED RECOMMENDATION

SYSTEM USING AWS
This is an experiential learning exercise that allows the team members to develop a personal-
ized recommendation system using Amazon Personalize.
Prerequisites: You need a basic knowledge of AWS services, data gathering techniques, data
cleaning processes, and some machine learning concepts.
TASK NAME TASK OWNER TASK DETAILS DELIVERABLE

(SKILL SET)
Data gathering Data engineer Gather click views and item Data stored in
view data from websites. Amazon S3
Store this data in Amazon
S3.
Data cleaning Data scientist/ Use Amazon Glue Cleaned data in

data engineer DataBrew to clean and Amazon S3
prepare the data for model
training.
Model training Machine Train a recommendation Trained

learning model in Amazon recommendation
engineer Personalize using event model
data, item data, and user
data.
Model hosting Machine Host the trained model via API endpoint for
learning Amazon API Gateway to model
engineer make it accessible for the
application.
Recommendation Software Use the API created in the Application

implementation engineer/ previous step to integrate with integrated
full-stack recommendations into the recommendation
developer business application. feature
Real-Time Customer Engagement

Figure 13.8 shows a real-time customer engagement reference architecture that implements use cases such as
churn predictions, recommendations using customer segmentation, and engaging with customers by sending them
messages based on real-time transactions.
DATA LAKE
Email
DATA COMMUNICATIONS STREAMING Engagement Data
CENTRALIZED
SOURCES SERVICE DATA Customer
DATA CATALOG
Email Email PIPELINE Demographic and
Subscription Data
S3
Train DATA CONSUMPTION

BATCH
TRANSFORMATION
MODEL SQL Query Analyzer
Jupyter BI Intelligence
Notebook
FIGURE 13.8: Real-time customer engagement architecture
The following steps implement the real-time customer engagement reference architecture using the AWS compo-
nents shown in Figure 13.9:
STEP NAME STEP DESCRIPTION AWS SERVICES INVOLVED
Upload user data First, you upload the users and their contact info, Amazon Pinpoint, Amazon S3
such as emails, into Amazon Pinpoint and collect
data about customer interactions into S3.
Ingest data You then ingest this data in Amazon S3 using Amazon S3, Amazon Kinesis Data
either Amazon Kinesis Data Firehose or Amazon Firehose, Amazon Kinesis Data
Kinesis data stream. Stream
Train model You then use this data in S3 to train a model in Amazon S3, Amazon SageMaker
Amazon SageMaker to predict the likelihood of
customer churn or gather customer segmentation
data. You can create a SageMaker endpoint once
the model is ready for production.
Run inference You then run the inference in a batch manner and Amazon S3, Amazon Pinpoint
and export export the results into S3 and Pinpoint.
results
Combine data You can combine the data in Pinpoint with other Amazon Pinpoint, Amazon
and gather data from your data lake and get insights using Athena
insights Amazon Athena.
Visualize data Finally, you can use Amazon QuickSight to Amazon QuickSight
visualize the data and share it with others.
DATA LAKE
STREAMING
COMMUNICATIONS Email
DATA
DATA SERVICE Engagement Data CENTRALIZED
PIPELINE
SOURCES Email Customer DATA CATALOG
(Amazon
Email (Amazon PinPoint Demographic and (Amazon Glue)
Kinesis Data
Amazon SNS) Subscription Data
Firehose)
S3
DATA CONSUMPTION
BATCH
Train SQL Query Analyzer
TRANSFORMATION
(Amazon
MODEL (Amazon Athena)
Jupyter BI Intelligence
SageMaker)
Notebook (Amazon QuickSight)
FIGURE 13.9: Real-time customer engagement architecture on AWS (AWS components shown in bold italics)
TIP By using Amazon SageMaker Pipelines, you can automate the steps to build, train, and
deploy machine learning models and save time and effort.
Figure 13.10 shows a customer engagement architecture that you can implement using Azure components.
User and Movie

User Clients
Interaction Data
Azure Blob Data Science Azure Machine Azure Container

Storage Virtual Machine Learning Service Instances – Model
Service
Azure
Cosmos DB
FIGURE 13.10: Real-time customer engagement architecture on Azure

EXERCISE: BUILDING A REAL-TIME CUSTOMER ENGAGEMENT SYSTEM
This is an experiential learning exercise that allows the team members to implement a real-time
customer engagement system focusing on use cases such as customer churn, and personalized
recommendations using customer segmentation.
Prerequisites: You need a basic knowledge of AWS services, IoT devices, and some data
science concepts.
Tasks: Use the steps outlined in the table to build this architecture. The detail of building this
architecture is outside the scope of the book, but it would be a good starting point if your use
case involves engaging with customers.
Data Anomaly and Fraud Detection

The fraud detection architecture shown in Figure 13.11 helps detect fraud in transactions such as fake accounts
and fraudulent payments.
NOTE Use Amazon SageMaker Autopilot to automatically build, train, and deploy machine
learning models for fraud detection.
EVENT
API ML DATA LAKE
MANAGEMENT
GATEWAY PLATFORM MODEL AND DATA
SERVICE
DATA CONSUMPTION
DATA LAKE
DATA PIPELINE SQL Query Analyzer
RESULTS
BI Intelligence
FIGURE 13.11: Fraud detection architecture
To build the fraud detection architecture using AWS components shown in Figure 13.12, you must first use the
Amazon CloudFormation template to instantiate an instance of Amazon SageMaker with the sample dataset and
train models using that dataset.
The solution also contains an Amazon Lambda function that invokes the two Amazon SageMaker endpoints for
classification and anomaly detection using the sample data.
The Amazon Kinesis Data Firehose loads the transactions into the Amazon S3, which is used to make predictions
and generate new insights.
ML
API PLATFORM
EVENT
GATEWAY (Amazon DATA LAKE
MANAGEMENT
Amazon SageMaker MODEL AND DATA
SERVICE
API - Random S3
(Amazon Lambda)
Gateway cut forest
- xgboost)
DATA CONSUMPTION
DATA PIPELINE DATA LAKE SQL Query Analyzer
(Amazon Kinesis RESULTS (Amazon Athena)
Data Firehose) S3 BI Intelligence
(Amazon QuickSight)
FIGURE 13.12: Fraud detection architecture on AWS (AWS components shown in bold italics)
EXERCISE: BUILDING A REAL-TIME CUSTOMER ENGAGEMENT SYSTEM
This is an experiential learning exercise that allows the team members to build a working
fraud detection system that can automatically detect fraudulent transactions in real time.
Prerequisites: You need a basic knowledge of AWS services and some data science concepts.
Initialize Data scientist Use the Amazon Amazon

Amazon CloudFormation template SageMaker
SageMaker to create an instance instance with
of Amazon SageMaker sample dataset
equipped with a sample
dataset for fraud detection.
Train models Data scientist Use Amazon SageMaker Trained machine

Autopilot to automatically learning models
build, train, and deploy
machine learning models for
fraud detection.
Set up Lambda AWS Create an Amazon Lambda Working AWS

Function architect function that will invoke Lambda function
your Amazon SageMaker
endpoints. This function
should be able to feed new
data to these endpoints and
receive their predictions in
return.
continues
continued
Set up Kinesis AWS Use Amazon Kinesis Real-time data

Data Firehose architect Data Firehose to create a pipeline with
real-time data streaming Amazon Kinesis
pipeline. This pipeline will Firehose
load new transactions into
Amazon S3.
Run inference Data scientist As new transactions are Working fraud

on new data loaded into Amazon S3, use detection system
your Amazon SageMaker
models to make predictions.
These predictions should
indicate whether each
transaction is likely to be
fraudulent.
FACTORS TO CONSIDER WHEN BUILDING AN ML PLATFORM

This section covers ML platform considerations such as building versus buying an ML platform, choosing a
suitable cloud provider, and the MLOps maturity model.
The Build vs. Buy Decision

Building a platform to meet the needs of a healthcare ML platform can be a daunting task. You would need to
customize the entire stack using best-of-breed, open-source, and out-of-the-box services and tools provided by
cloud providers. It is vital for you to embrace the automation, instrumentation, and scalability of MLOps to
achieve your platform goals. The logical question is whether to build or buy best-of-breed solutions. To answer
this question effectively, carry out the tasks in the following exercise.
HANDS-ON EXERCISE: BUILDING VS. BUYING AN ML PLATFORM
This is a hands-on experiential learning exercise that allows the team members to make an
informed decision on building versus buying an AI/ML platform based on the company’s needs.
Using the table, execute the steps to deliver a recommendation to your leadership team for a
build vs buy decision.
TASK NAME/ ROLE BUY BUILD DELIVERABLE/
DETAILS IMPORTANCE
Identify Organization’s Business Pay specific attention to the Consider factors like Comprehensive list
Needs analyst need for customization and in-house expertise, cost, of organizational
List the specific requirements configuration, integrate-ability, integration, scalability, requirements for an
and constraints of your vendor expertise and support. training needs, and the ML platform
organization with respect to Also factor in cost, scalability, and need for control of data.
an ML platform. security requirements.
Explore Available Options Data Factor in attributes to evaluate N/A Comparative analysis
Explore various ML platforms scientist such as features, cost, community report of different
both ready-made and open- support, security and compliance. ML platforms
source options and evaluate Also consider vendor support and
their pros and cons. skillsets, ease of use, scalability,
integration ability, interoperability,
training, and upgradability.
Assess In-house Talent HR You must consider the vendor’s You need AI/ML talent Report on the
Your success depends on manager AI/ML expertise and support. You to build, maintain, and current state of
your capacity to support need a smaller team to implement operate the platform in-house talent and
either scenario. and maintain the system. Evaluate and there is significant training needs
your inhouse skills and training investment in training
needs. and development.
Consider Interoperability IT architect Evaluate the cloud provider’s You can customize Assessment report
Ensure the chosen AI/ compatibility with other systems. to ensure maximum on interoperability
ML services and solutions Usually, it is easier to integrate compatibility with other of potential ML
can work seamlessly with within their ecosystem. Vendor systems, but it requires platforms
existing or planned systems is lock-in is possible. extra resources and time.
paramount.
DETAILS IMPORTANCE
Evaluate Control over IT manager Limited control over underlying AI/ You would have greater Report on the
Platform ML algorithms and tools, and you control over aspects such degree of control
For an AI/ML platform, may have to depend on vendor’s as adding or modifying provided by
the ability to tweak the upgrade/release cycle for new features as needed. different options
algorithms or change features.
the architecture can be
crucial depending on your
company’s specific needs.
Explore Open-Source Data Vendor may not offer full support You could incorporate Report on the open-
Support scientist for open-source AI/ML tools. as many open-source source support by
Open-source tools provide Some may work well only with components as necessary different options
greater community support their proprietary tools. providing you with
and have frequent updates flexibility and innovation
and hence is important to speed.
consider.
Estimate Time-to-Market Project You can deploy faster due to Custom development Estimated timeline
You have to factor in the time manager prebuilt AI/ML capabilities but can be time-consuming, for each option
urgency to choose between may still need customization and leading to slower
the two options. integration. deployment.
Assess Cost Financial You would have upfront costs, Potentially high Cost assessment
You must factor in both the analyst possibly with ongoing licensing or upfront development report for build
initial and ongoing costs subscription fees. costs, ongoing costs versus buy
when choosing to buy or for maintenance,
build. upgrades, and potential
modifications.
DETAILS IMPORTANCE
Review Feature Business May not be as customizable. Might It is usually possible Report on the
Requirements analyst not meet all specific requirements. to meet most specific alignment of feature
Your choice will depend on business requirements. requirements with
your specific business and available options
functional requirements.
Examine Scalability and Data The vendor should be able to You can build your Report on the
Flexibility engineer scale and adapt to your growing platform with scalability scalability and
As your business grows, data and model complexity. in mind from the start flexibility of each
your need to deal with large and provide greater option
volumes of data and complex flexibility.
models will also grow.
Assess Vendor Support and IT manager Vendors provide regular support You have to take greater Assessment report
Upgrades and updates, but you are still responsibility for support on vendor support
AI/ML field is constantly dependent on their update cycle. and updates, and this and upgrades
evolving, and this is a key may require additional
factor to consider from an resources.
innovation perspective.
Analyze Security and IT security AI/ML cloud providers should You have greater control Report on the
Compliance officer ensure that their solution over the compliance security and
Ensuring data security meets industry and regulatory process, but it also compliance of each
and meeting regulatory compliance standards such as means you have to option
compliance are critical, HIPAA or GDPR. design with security and
regardless of whether the regulatory compliance in
solution is bought or built. mind, potentially adding
complexity.
Make a Recommendation Project Make a recommendation on whether to build or buy the ML Recommendation
lead platform report on whether to
build or buy the ML
platform
TIP Consider factors such as talent, interoperability, control, open-source support, time-to-
market, cost, feature requirements, scalability, vendor support, and security compliance when
making the build versus buy decision.
Choosing Between Cloud Providers

When choosing between cloud providers, you need to factor in a few considerations. The following exercise helps
you to evaluate the appropriate criteria.
HANDS-ON EXERCISE: CHOOSING THE RIGHT CLOUD PROVIDER
The goal of this hands-on exercise is to help you gain a comprehensive understanding of the
different cloud providers, including their strengths and weaknesses, and choose the right cloud
provider or a combination of providers for different AI/ML services, platforms, and tools that
best suit your company’s needs.
Here are the steps needed to choose the best cloud provider:
TASK TASK EVALUATION CRITERIA DELIVERABLE

OWNER
Identify cloud Team Shortlist the cloud providers List of cloud

providers. lead you want to evaluate. providers
Evaluate AI/ Data Check for the availability of Report regarding

ML services and scientist AI and ML services for various AI/ML services
capabilities that aspects of the machine and capabilities
are crucial for your learning lifecycle, such as data available for each
AI and ML projects. processing, model training, cloud provider
testing, deployment, and
monitoring.
Tip: Look out for the
availability of prebuilt models,
frameworks, and development
tools.
Examine scalability Data Evaluate the scalability and Report on

and performance, engineer performance including support scalability and
which are essential for distributed processing, and performance
for growing compare their performances for each cloud
businesses. and latencies for large volumes provider
of data and high-intensity ML
workloads.
Factors to Consider When Building an ML Platform ❘ 205
TASK TASK EVALUATION CRITERIA DELIVERABLE

OWNER
Examine data Data Compare in terms of their Report on data

management and architect support for data processing, management
integration, which labeling, and data governance, and integration
are important as well as the ability to capabilities
when dealing with integrate with industry-specific
large or sensitive data.
datasets.
Evaluate platform Software Compare the availability of Report on

ecosystem and engineer developer community, vendor platform
community, which support, and the development ecosystem and
are useful for faster ecosystem such as APIs, SDKs, community
development and and development tools.
troubleshooting.
Assess security Security Evaluate the cloud providers Report on

and compliance, analyst in terms of their support for platform support
in particular for encryption access controls, for security and
dealing with data privacy, and other compliance
sensitive data. industry certifications that they
may have.
Analyze pricing Financial Evaluate their pricing model Cost

model and costs, analyst and compare their costs for comparison
which is important data storage, data processing, report
for budget data transfer costs, compute
management. costs, and ML service pricing.
Evaluate vendor Project Evaluate their documentation, Report on

support, which manager training, product support, and vendor support
is required the available developer and
for smoother partner ecosystem.
implementation
and operation.
Make a final Team Make a recommendation Recommendation

recommendation. lead based on your evaluations and report
choose a cloud provider or a
combination thereof that suits
your company.
TIP Consider the range of AI/ML services offered by the various cloud providers, including
prebuilt models, development tools, and support for the entire ML lifecycle when choosing a
cloud provider.
KEY COMPONENTS OF AN ML AND DL PLATFORM

An ML platform’s main components are enabling data ingestion, data preparation, model training, deployment,
governance, and monitoring. Many organizations want to deploy their machine learning models in the cloud for
the benefits discussed earlier. This is true for companies just beginning to deploy machine learning models and
those who are advanced in this journey.
Organizations trying to get value from business applications or using curated data are looking at the underlying
machine learning platform and infrastructure as a means to an end. By adopting AI services in the cloud, even
developers with little or no skills in machine learning can build, train, and deploy complex models in production.
As you keep trying to adopt ML widely across your organization and as the requirements become increasingly
complex, you must adopt a combination of cloud-based services, open-source tools, best-of-breed solutions, and
your own proprietary custom solutions to build an enterprise-wide AI platform. This will give you fine-grained
control over the entire platform but also take a lot of effort and cost. You can derive unique value from the
platform software, infrastructure, and expertise, but you have to decide, based on the cost-benefit analysis, which
approaches you want to take.
KEY COMPONENTS OF AN ENTERPRISE AI/ML HEALTHCARE

PLATFORM
This section considers the use case of a healthcare company trying to build an enterprise AI/ML imaging plat-
form. You can use this to guide your own AI/ML platform. The platform must have the following components:
COMPONENT DESCRIPTION
Data ingestion Ingest enormous amounts of medical data such as X-rays, CT scans, and MRIs
from multiple sources.
Preprocessing and data Automatically preprocess and clean the data such as reducing noise, adjusting
cleaning contrast, and carrying out normalization.
Labeling Label a subset of images to create a labeled dataset for subsequent model
training. This may involve identifying and annotating abnormalities and identifying
new feature sets.
Model training Train the ML models using the labeled data to carry out machine learning tasks,
such as anomaly detection classification and segmentation.
Model testing Test the machine learning models by integrating them into a simulation
environment that simulates real-life clinical scenarios. This helps with thorough
performance analysis and evaluation of the models before they can be deployed
in production.
Model deployment Deploy the models in a production scenario, such as a specific hospital or a
diagnostic center, to assist doctors and radiologists in accurately interpreting
medical images.
Performance monitoring Continuously monitor the performance of these models by measuring their
accuracy recall and precision metrics in real time.
Key Components of an Enterprise AI/ML Healthcare Platform ❘ 207
COMPONENT DESCRIPTION
Alerting, dashboarding, Generate alerts, dashboards, and feedback reports to assist healthcare
and feedback reports professionals so that they can make further improvements to these models,
as well as provide better patient care by suggesting alternative diagnoses,
identifying better treatments, and providing better care overall.
NOTE Handling large amounts of data requires a robust and automated data ingestion and
preprocessing pipeline.
Figure 13.13 shows some core platform architecture components, each of which is covered in the follow-
ing sections.
MODEL
DATA DATA ML MODEL
TRAINING AND
SOURCES MANAGEMENT EXPERIMENT DEPLOYMENT
EVALUATION
• ERP • Data
• CRM Transformation • Pre-
• Model Registry
• SCM • Versioning training • Model
and Deployment
• Third- • Shared Data • Automate Training and
• Model
party Lake Triggers Tuning
Operationalization
• Social • Processing & • Feature • Model
• Model Execution
media, Labeling Engineering Evaluation
• Model Serving
etc.
MODEL MONITORING & SECURITY AND MULTI-CLOUD

PERFORMANCE GOVERNANCE ARCHITECTURE
• Encryption
• Model Monitoring
• Security Controls
• Performance and Scalability of • Workload distribution
• Model Inventory
Models • Redundancy management
Management
• Distributed Computing • Disaster recovery
• Audit Trail
• Parallel Processing • Data synchronization
Management
• GPU Acceleration
• Tag Management
FIGURE 13.13: Basic components and their integrations in an AI/ML platform
Data Management Architecture

The data management components shown in Figure 13.14 are required for storing, ingesting, collecting, cleaning,
and preparing data for model training. You also need data orchestration, distribution, governance, and access
control tools.
Shared Data Lake

To use data for machine learning, you must store the data in one central location. It is recommended to set up a
data lake for this purpose. Consider factors such as data volumes, data access patterns, and data lifecycle manage-
ment requirements. You must choose data storage solutions such as Amazon S3, Google Cloud storage, and Azure
blob storage based on your specific requirements.
Once you store large datasets in S3, other data processing and ML services can access that data. Storing data in
one central location, such as S3, helps you manage data scientists’ workflow, facilitates automation, and enables
true collaboration and governance.
Creating the following folder structure with proper data classification will help you exercise access control and
facilitate machine learning workflows, tracking, traceability, and reproducibility during experiments:
FOLDER DESCRIPTION
Users and project teams Individual data scientists and teams can use these folders to store their
folders training, test, and validation datasets that they use for experimentation
purposes.
Shared features folders These folders can contain shared features such as customer data, product
data, time series, and location data, which can all be used for model training
across teams and models.
Training, validation, and These are used for formal model training, testing, and validation purposes.
test datasets folders Managing these folders is essential, as this can help you with the traceability
and reproducibility of models.
Automated ML pipelines These folders can be accessed only by automated pipelines, such as a WS
folders code pipeline for storing artifacts related to the pipeline.
Models in production These models should be actively monitored and managed through version
folders control and play a key role in ensuring traceability.
Data Sources
Your data can come from one or more sources located in various locations or systems. It helps to treat them
as enterprise-level assets rather than department-specific assets. The data sources can include databases, data
warehouses, data lakes, external APIs, files, and even streaming data from IoT devices and sensors. These sources
provide the raw data for the ML models during the training process.
Data Ingestion
Data ingestion is the process of loading the data from these data sources into the machine learning platform,
which can be an automated or manual process. It includes techniques such as data extraction, integration,
and loading.
Data Catalog
The data catalog acts as the central repository for all metadata about the datasets that are used in the ML system.
It contains information related to data sources, dataset description, data quality metrics, schema information,
data format, and usage statistics. Data scientists, analysts, and data engineers use this catalog to discover, under-
stand, and access data related to their projects.
TIP Establish a shared, centralized data lake to promote governance, collaboration, and
seamless data access across your company.
STRUCTURED DATA BATCH LOAD ANALYTICS
• Supported by • Extracts data

• BI Data
relational from various
Warehouses
databases sources
• Searchable
• For transactional • Data is moved
• Queryable
type of data into the data
DATA LAKE
• Have predefined lake at periodic Amazon Redshift
schema intervals Athena, QuickSight
• Enterprise-level
SEMI-STRUCTURED Amazon EMR data science
DATA AWS Glue • Scalable MACHINE
• Contain • Cloud scale LEARNING
identifiers STREAMING • Centralized
• Predictive
• No predefined Analytics
schema • Ingests data Amazon S3 • Prescriptive
from various
UNSTRUCTURED Analytics
sources
DATA • Ad-hoc Analysis
• Such as log
• AI/ML
• No predefined files,
Automation &
schema telemetry,
Efficiency
• Stored as files mobile, social
Amazon
Amazon Kinesis
SageMaker
Amazon MSK
FIGURE 13.14: Data management architecture
Data Transformation
Data transformation is the process of getting the raw data in a format and structure ready to use in ML models.
It involves cleaning the data, preparing it, feature transformation, data normalization, dimensionality reduction,
and so on.
Data Versioning
Data versioning is the process of assigning versions to datasets and tracking any changes to the data used in the
model. It helps you revert to the older version if necessary, and you can track the lineage of datasets used in dif-
ferent model training experiments. This can help you with the traceability and reproducibility of ML experiments.
Data Science Experimentation Platform

As shown in Figure 13.15, the data science or machine learning experimentation platform should provide com-
prehensive tools and services to help data scientists and teams conduct multiple individual or team-based experi-
ments for their projects. Consider these factors:
FEATURE DESCRIPTION EXAMPLE TOOLS/SERVICES
Data querying Teams must be able to explore data, analyze data, Amazon Athena
and analysis conduct SQL analysis or data transformations, and Google BigQuery ML
visualize results. For example, use Amazon Athena
Dataflow
to query data in a data lake using SQL.
Azure HDInsight, Azure Data
Lake Analytics
continues
(continued)
Code authoring Teams should be able to write and edit code Amazon SageMaker Notebook,
using IDEs and notebooks. For example, Amazon SageMaker Studio
SageMaker Notebook Instance and SageMaker AI Platform Notebook
Studio can be used for code authoring, data
Azure ML Studio
processing, model training, and experimentation.
Data processing This includes tools for data preprocessing and Amazon Data Wrangler,
transformation, such as data cleaning, data SageMaker Processing
wrangling, and feature engineering. For example, Google AI Hub
use Amazon Data Wrangler for data import,
Azure Machine Learning Studio
transformation, visualization, and analysis. Amazon
SageMaker processing can be used for large-scale
data processing, including integration with Apache
Spark and scikit-learn.
Pretraining Pretraining is the process of training a model on

a large, unlabeled corpus of data. It helps identify
general patterns and features from the data using
unsupervised or semi-supervised learning. This
model can then be used on a smaller subset of
labeled data to develop better predictions for
specific tasks.
Automate Events or triggers are set up in the machine

triggers learning experimentation project to automate
various steps. For example, when new data is
available, or a model has been updated, they can
trigger another workflow for deploying a model in
production or initiate the evaluation of a trained
model.
Feature store Enables sharing common features across teams for Amazon SageMaker Feature
model training and inference. Store
Vertex AI Feature Store
Azure Machine Learning
Model training Involves the ability to train and test using various Amazon SageMaker Training/
and tuning techniques and algorithms, evaluate models, and Tuning Service
conduct hyperparameter tuning. For example, use Google Cloud AutoML
Amazon SageMaker training/tuning service.
Vertex AI
Azure ML Studio
Azure Automated Machine
Learning
Model In a machine learning experimentation platform, Amazon SageMaker Training/

evaluation model evaluation is a crucial step that helps Tuning Service
you compare different experiments, conduct
Google Cloud AutoML
hyperparameter tuning, and choose the best
model available for deployment. Compare the Vertex AI
performance of different trained models using Azure ML Studio
metrics such as accuracy, precision, F1 score, and Azure Automated Machine
recall based on the type of task and the problem
Learning
domain. Model evaluation determines how well
the model generalizes against the unseen data and
solves the problem.
Containerization The ability to build, test, and deploy models and Amazon ECR
workflows as containerized applications or services. Kubernetes Engine
For example, use Amazon ECR to store training,
Azure Container Registry
processing, and inference containers.
Source code Managing code using source code control systems Git
control such as Git to enable tracking changes and
collaboration between team members and ensure
reproducibility. For example, Artifacts repositories
are private package repositories used to manage
library packages and mirror public library sites.
Data science Providing easy access to data science libraries NumPy, Pandas, scikit-learn,
library access such as NumPy, Pandas, scikit-learn, PyTorch, PyTorch, TensorFlow
and TensorFlow to allow easy development and
experimentation.
Self-service Using ML tools, services, and products to allow Self-service catalog

catalog data scientists and team members to quickly
access their ML needs.
Implement a Developing a self-service catalog of various ML Implement a self-service catalog

self-service tools, services, and products so that other team
catalog members, such as data scientists, can provision
tools such as SageMaker Studio or Canvas through
the self-service portal.
Hybrid and Edge Computing

Hybrid and edge computing are two different types of computing approaches in the world of machine learning
that can be used to embrace the best of both worlds (see Figure 13.16).
Hybrid Computing Scenarios

In the case of hybrid computing, the following are some typical practices:
CENTRALIZED ENVIRONMENT
Source Code Control - CodeCommit
Continuous Delivery Service - CodePipeline
Artifact Repository - CodeArtifact
CORPORATE DATA CENTER

Direct Transit
On-premise code repository
Connect Gateway
On-premise package repository
DEVELOPMENT ENVIRONMENT
Container repository – Amazon ECR
Data lake – S3
Internet Workflow orchestrator – AWS Step Functions
Event management – AWS Lambda
ETL software – AWS Glue
ML platform – Amazon SageMaker
FIGURE 13.15: Components of an ML experimentation platform
➤➤ Deploy models in the cloud for increased scalability and performance and then deploy some models in
the edge for increased security.
➤➤ You can carry out data storage, data preprocessing, and model training in your own on-prem infrastruc-
ture and offload distributed processing, large-scale training, and hosting the model to the cloud.
➤➤ Hybrid computing can help you with resource optimization, cost optimization, and data governance.
Edge Computing
Here are some considerations for edge computing:
➤➤ It brings data processing and model training to edge devices such as IoT devices, sensors, or edge servers
closer to the point of data generation or data source.
➤➤ It can be handy for use cases that require real-time processing of data with exceptionally low latencies,
reduced bandwidth, and increased security and privacy.
NOTE Hybrid and edge computing are especially useful in use cases such as remote health-
care monitoring, autonomous vehicles, and industrial IoT.
Designing Models for the Edge

Designing models for the edge, you need to factor in some specific considerations, such as the following:
➤➤ Model size: The model should be small enough to fit on the edge device.
➤➤ Model latency: The model latency should be low enough to suit the requirements of the application.
➤➤ Model accuracy: The model should be accurate enough to meet the requirements of the application.
To meet these requirements, you may be required to adopt model compression techniques such as quantization
and model pruning to reduce the model size. Moreover, the model should be designed to operate in offline or
intermittent connectivity scenarios.
Data Synchronization, Latency,

and Network Connectivity
Hybrid Edge Computing
Considerations
• Model deployment in • Data processing
the cloud and at the • Model training on IoT • Push or pull synchronization
edge devices and edge servers strategies
• On-prem infrastructure • Model size, model latency,
• Cloud distributed and model accuracy
• Techniques to improve latency,
processing • Remote healthcare such as edge caching
monitoring, autonomous • Preprocessing, or local inference
vehicles, and industrial IoT • Address network connectivity
issues
FIGURE 13.16: Hybrid and edge computing in machine learning
➤➤ Data synchronization, latency, and network connectivity: In the case of hybrid and edge computing sce-
narios, data synchronization latency and network connectivity are crucial factors for the effective opera-
tion of machine learning models.
➤➤ To keep the data consistent between the cloud and the edge devices: Adopt a push or pull synchroniza-
tion strategy. In the case of push synchronization, the cloud pushes the data to the edge device, or in the
pull synchronization strategy, the edge device pulls the data from the cloud.
➤➤ To improve latency: You can adopt techniques like edge caching, preprocessing, or local inference to
avoid round trips to the cloud.
➤➤ To address network connectivity issues: You can build fault-tolerant systems by using compression tech-
niques or by optimizing data transmission protocols.
The Multicloud Architecture

The multicloud architecture aims to improve the scalability and reliability of the overall enterprise platform by
leveraging the strength of the various cloud providers. However, it can be a complex endeavor, and you need to
manage various aspects such as data synchronization, workload distribution, system redundancy, disaster recov-
ery, and implement security protocols, as shown in the following table:
CATEGORY DETAILS
Workload You can implement workload distribution in such a manner that you can choose one
distribution cloud provider for model training based on their infrastructure capabilities, such as
specialized GPU instances, and choose another provider for real-time inference based
on their low latency services.
• You can also reduce the risk of service disruptions by distributing the workloads
across multiple cloud providers.
• Using geolocation services, you can distribute the workload to the users based on
their locations.
• You can also use load-balancing features to distribute the load across multiple cloud
providers so that no one cloud provider is overloaded.
continues
(continued)
CATEGORY DETAILS
Redundancy You can employ redundancy management to ensure no single point of failure. You
management can implement this by having data redundancy, infrastructure redundancy, application
redundancy, load balancing, and service failover techniques. You can have either active-
active or active-passive configurations set up to ensure continuity of services.
Disaster recovery By implementing the redundancy management techniques discussed earlier, you can
also ensure you have a sound disaster recovery system in place so that you can failover
should the system go down.
Data Keeping the data synchronized across multiple cloud environments is a critical step, and
synchronization you must employ different techniques, such as real-time or scheduled synchronization.
You must employ distributed databases, data synchronization, replication, and
streaming techniques.
Security You must employ a robust multicloud security strategy to protect data, applications,
and networks. Security measures such as IAM, access controls, data encryption,
network security, and monitoring are essential. Data flow between cloud environments
should be encrypted, and security protocols across the different cloud environments
should be synchronized.
NOTE Achieving a successful multicloud architecture helps you avoid vendor lock-in, but it
can be a complex endeavor.
WORKBOOK TEMPLATE: ENTERPRISE AI CLOUD PLATFORM SETUP

CHECKLIST
Download the “Enterprise AI Cloud Platform Setup Checklist” template from the download section of the book
(www.wiley.com/go/EnterpriseAIintheCloud). Use this template to ensure you not only have the needed
resources, but also have secured and optimized the environment. Track your progress with setting up your enter-
prise AI cloud platform.
SUMMARY
This chapter laid the foundations for the rest of your AI journey. The infrastructure serves as the technical bridge
between your initial strategic planning, team building, and the subsequent stages of data processing, model train-
ing, deployment, and scaling.
This chapter covered the complex components of an enterprise AI cloud platform and dived into five types of
reference architectures that are most common. You also reviewed the considerations for buy versus build deci-
sions and the criteria for choosing between cloud providers. The chapter delved into the specific components of
the ML and DL platforms, data management architecture, and the nuances of the hybrid and edge computing and
multicloud architecture.
Whether it is building an AI platform for healthcare or data synchronization in real time, this chapter provided
the necessary knowledge and tools to benefit IT professionals, cloud architects, system administrators, data scien-
tists, decision-makers, and technology strategists.
But your journey doesn’t stop here. With your infrastructure now set, you will be processing your data, choosing
the right algorithms, training your models, and ensuring they are deployed efficiently and ethically. Remember
that the infrastructure plays a key role in scaling your AI efforts, deploying models seamlessly, and evolving and
maturing your AI capabilities, all of which are covered in the remaining chapters of this book.
REVIEW QUESTIONS
1. Name the AWS service used to collect data from IoT devices in the described architecture.
A. AWS Greengrass
B. AWS IoT Core
C. Amazon Redshift
D. Amazon S3
2. The ______ service is used to train a personalized recommendation model.
A. Amazon S3
B. Amazon Glue DataBrew
C. Amazon Personalize
D. Amazon API Gateway
3. The ______ service is used to automate the steps of building, training, and deploying machine learn-
ing models.
A. Amazon Pinpoint
B. Amazon S3
C. Amazon SageMaker
D. Amazon Athena
4. The biggest advantage of adopting machine learning in the cloud is that developers with little or no
machine learning skills can:
A. Only deploy pretrained models
B. Only build simple models
C. Build, train, and deploy complex models in production
D. Only access curated data
5. ____________ should be considered when deciding whether to build or buy solutions for an enterprise
ML platform. (choose more than one answer)
A. Availability of in-house talent
B. Integration with third-party applications
C. Cost of hardware and software licenses
D. Time required for model training
6. True or False: It is more important to manage data than to improve models in machine learning.
7. _______ is NOT a technique for improving latency in hybrid and edge computing.
A. Edge caching
B. Preprocessing
C. Local inference
D. Round trips to the cloud
8. Which of the following statements are true regarding multicloud architecture?

A. Workload distribution in multicloud architecture allows choosing different cloud providers based
on their specialized infrastructure capabilities.
B. Disaster recovery in multicloud architecture is not necessary as the risk of service disruptions
is reduced.
C. Data synchronization in multicloud architecture can be achieved only through real-time syn-
chronization.
ANSWER KEY
1. B 4. C 7. D
2. C 5. A, B, C, D 8. A
3. C 6. True
14
Operating Your AI Platform with
MLOps Best Practices
The value of an idea lies in the using of it.
—Thomas Edison
Now that the AI platform has been built, your focus shifts toward operating the AI platform; ensuring the
efficiency, reliability, and repeatability of the ML operations workflow; and enabling scale with automation
(see Figure 14.1).
The target audience includes tech leads, operations managers, ML engineers, data scientists, systems admin-
istrators, and IT professionals looking to employ automation in daily operations.
Automation using MLOps isn’t just a buzzword. It’s the secret sauce to scaling rapidly, reducing manual
errors, speeding up processes, and enabling adaptability in a dynamic business landscape.
In this chapter, you review various key components, from model operationalization and deployment sce-
narios to automation pipelines, platform monitoring, performance optimization, security, and much more.
It includes actionable deliverables that will help you crystallize the MLOps ideas presented in this chapter.
This chapter serves as a pivotal transition point from setting up your AI infrastructure (Chapter 13) to
actual data processing and modeling (Chapters 15–17), highlighting the role of MLOps in ensuring this
transition is smooth, efficient, and automated.
CENTRAL ROLE OF MLOps IN BRIDGING INFRASTRUCTURE,

DATA, AND MODELS
This section discusses the role that the MLOps best practices in this chapter play in ensuring automation
and integration across the platform infrastructure, data, and model layers.
What Is MLOps?
MLOps stands for “machine learning operations,” which is a set of practices, principles, and tools that unify
machine learning (ML) development and operations (Ops). It focuses on automating the end-to-end machine
learning lifecycle, from data preparation to model training to deployment and monitoring. It enables faster,
more reliable, and scalable ML implementations, which is the focus of this chapter. MLOPs is similar to
DevOps in software development, but it is specifically geared towards automating ML workflows.
218 ❘ CHAPTER 14 Operating Your AI Platform with MLOps Best Practices
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Govern Your
Your AI
Team Setup Infrastructure and
journey
09 Manage Operations
ENTERPRISE AI
Setup 04 • Setup Enterprise AI Cloud
Evolve and JOURNEY MAP Infrastructure
Mature AI Platform Infrastructure
and Manage
Operations • Automate AI platform operations for
scale and speed
08 Process Data
Scale and Objective
AI
Deploy and 05 Efficiently manage and scale your AI platform
Monitor
Models by integrating MLOps best practices for
07 automation, model deployment, and tracking
throughout the AI lifecycle
06
FIGURE 14.1: SETUP INFRASTRUCTURE AND MANAGE OPERATIONS: Automate AI platform operations
for scale and speed
MLOps is about continuous integration and continuous delivery (CI/CD) of not just models, but also of data.
The goal of MLOps is to automate and scale the tasks of processing and engineering data. Your goal must be to
ensure that you employ the best practices in this chapter so that when processing data in Chapter 15, as new data
comes in, it’s processed and made available automatically for model training seamlessly.
In Chapters 16 and 17, you may try different algorithms and train and tune multiple models. Your goal must be
to deploy the best practices in this chapter to ensure that each model version is tracked, stored, and can be rolled
back or forward as needed. This also ties in with the concept of tagging and container image management, which
will facilitate this process.
Once you have trained your models, they need to be evaluated. The practices outlined in this chapter and
Chapter 19 can help automate this step. It ensures that as soon as the model is trained using the practices in
Chapter 17, it’s automatically evaluated and metrics are logged.
Finally, I want to emphasize the importance of having feedback loops. By that I mean the evaluations and insights
obtained from the training, tuning, and evaluation of the models should be fed back into the system to refine the
data processing methods in Chapter 15.
Automation Through MLOps Workflows

To achieve speed and scale, you must automate the different steps in the machine learning workflow, such as data
collection, ingestion, cleaning, feature transformation, model training, testing, tuning, deployment, and monitor-
ing. These steps should be done in an automated fashion in a specific sequence at predefined times. Automating
routine tasks helps the data science team focus on value-added tasks such as developing new models or insights.
You can use tools like Apache Airflow or Google Cloud Composer to schedule and orchestrate your data pipeline
tasks so they are executed in a given order and on time.
The following table shows various MLOps practices that can be adopted across the machine learning workflow
with a particular focus on data management. It details the tools or applications that can be used along with use
cases where applicable.
STEPS IN MACHINE DESCRIPTION TOOLS/APPLICATIONS USE CASES
LEARNING WORKFLOW
Data collection, ingestion, Automate these steps in an ordered, Apache Airflow, Google N/A.
cleaning, feature predefined sequence. Cloud Composer
transformation
MLOps automation: You can create machine learning Kubeflow, MLflow, AWS You can retrain the model if it doesn’t
model training, testing, workflows using the CI/CD pipeline SageMaker pipelines, meet the performance thresholds,
tuning, deployment, and and a discipline known as MLOps, Azure Machine Learning validate it, and redeploy it once it
monitoring which evolved from DevOps, a pipelines, Google Cloud meets the performance criteria.
common practice in software AI platform pipelines
development.
Using a feature store A feature store is a key feature in Online and offline It is helpful for training and batch
the MLOps lifecycle. It speeds up feature storage. Online inferences. It helps machine learning
machine learning tasks and eliminates feature storage is used engineers and data scientists create,
the need to create the same feature for real-time inferences, share, and manage features during ML
multiple times. while offline feature development and reduces the curation
storage is used to work required to create features out of
maintain a history of raw data.
feature values.
Continuous learning When the model was released AWS SageMaker, Azure Taking the cooking analogy, just like a
in production, it would have Machine Learning, restaurant’s menu needs to be updated
been trained on past data, but Google Cloud AI with the customers’ changing tastes,
its performance may decline as Platform. you need to keep your models up-to-
time goes on. This is because date.
customers’ tastes, product lines,
and many other variables change.
In other words, while the world
has changed, the model remains
unchanged. This means you need to
continuously monitor the model for
its performance and retrain it in an
automated manner. This process is
known as continuous learning.
STEPS IN MACHINE DESCRIPTION TOOLS/APPLICATIONS USE CASES
LEARNING WORKFLOW
Data version control Data version tracking is like a recipe You can use data version One advantage of DVC is that you
book, where you capture all the control tools like the can maintain a record of all the data,
ingredients, the process steps, and Data Version Control code, and models used and share that
the meal created in the end. This (DVC) to manage the data and models across other teams,
list helps you re-create or tweak the data versions and track reproduce experiments, and roll back
meal because you know what was the models with which changes, thus making your machine
done precisely. In the context of they were used. DVC learning workflows collaborative,
machine learning, you maintain the can integrate with other reproducible, and reliable.
list of datasets used, the models and version control tools,
their versions, and the parameters such as Git, and handle
used to tune the model. All this large amounts of data.
information helps you in the future
to either tweak the model for better
performance or to troubleshoot the
model to identify the root cause
of its poor performance. If you get
different results every time, you can
assess whether the model, the data,
or both contribute to the results.
Model Operationalization ❘ 221
TIP Use a machine learning workflow management tool that can automate the machine
learning process to scale your ML operations
MODEL OPERATIONALIZATION
Model operationalization is the process of taking a previously trained model and getting it ready for execution,
hosting, monitoring, and consumption in production. It ensures that the model can be executed on new data,
hosted on a scalable and reliable infrastructure, monitored for performance, and seamlessly integrated into appli-
cations and systems to draw business insights for further action. Remember that while training a model by itself is
a significant achievement, its real value is only realized when it’s operationalized. Let’s look at each step of model
operationalization in more detail:
STEP PURPOSE TOOLS/SERVICES
Model The process of taking a previously trained model Model execution can happen using two
execution and integrating it into an existing application or methods: one is online processing, and
system to take new data and make predictions the other is batch processing.
in real-world scenarios. Online processing: In the case of
online processing, one piece of data is
fed into the model, as is the case with
product recommendations.
Batch processing: In the case of batch
processing, a large batch of data is fed
into the model for prediction, which is
the case for fraud detection and risk
management.
Model The ability to host models in production in a Model hosting can be done using
hosting controlled environment to ensure functionality various approaches, such as deploying
and performance testing before deploying it it on serverless computing, containers
for widespread use. It involves setting up the such as Dockers, and dedicated servers.
necessary infrastructure and providing the In-house platforms can also host
necessary compute resources to run efficiently models but can be more expensive to
without a significant latency for high workloads. maintain. Amazon SageMaker Hosting
is an example of this.
Model The process of continuously monitoring the Metrics gathering tools, Amazon
monitoring deployed models for performance and behavior CloudWatch.
in a production environment.
Metrics: It involves gathering the metrics such
as model response time, prediction outputs,
and other data to assess the model’s accuracy,
stability, and adherence to performance goals.
It helps detect concept drift and model drift
and identify anomalies to ensure the model
performance does not degrade over time.
STEP PURPOSE TOOLS/SERVICES
Model The actual process of using the model to make Model consumption can happen in
consumption predictions. two ways: by using an API for real-time
You may need to integrate the model into prediction and a web service for batch
downstream applications, systems or user predictions for large datasets.
interfaces, and other business processes or
workflows so that users can make predictions
using the model.
TIP Implement continuous model monitoring to maintain quality and reliability of models
in real-world scenarios.
Automation Pipelines
Automation pipelines are the bedrock of enterprise AI. Through automation, you can enforce governance and
standardization such as tagging and naming conventions and repeatability of the entire machine learning lifecycle,
starting from data preparation model training, testing, deployment, and monitoring.
As shown in Figure 14.2, the automation pipeline consists of several components for model training and deploy-
ment, such as the following:
SHARED CENTRALIZED ENVIRONMENT MODEL REGISTRY

Source Code Control – CodeCommit Container Registry - ECR
Continuous Delivery Service – CodePipeline Data Storage – S3
Artifact Repository – CodeArtifact Model Registry – SageMaker Model Registry
CORPORATE DATA CENTER

Direct Transit Approver
On-premise code repository
Connect Gateway
On-premise package repository
TEST/UAT
ENVIRONMENT
DEVELOPMENT Data Storage
ENVIRONMENT (S3)
ML Workflow PROD ENVIRONMENT
Internet Data Storage API Gateway
(AWS Step Functions)
(S3) (SageMaker Endpoint)
Data Pipeline
ML Platform
(SageMaker Processing)
(Amazon SageMaker)
Model Training
(SageMaker Training)
FIGURE 14.2: CI/CD flow for model training and deployment

The table below describes the various components included in Figures 14.2 and 14.3 for model training and
deployment.
COMPONENT DESCRIPTION EXAMPLE SERVICES
Code repository The starting point for an automated pipeline Docker files, training scripts,
run, and it consists of several artifacts such dependency packages
as Docker files, training scripts, and other Azure OpenAI Code Repository,
dependency packages required to train Google Artifact Registry
models, build containers, and deploy models.
Code build service A code build service can be used to build AWS CodeBuild, Jenkins
custom artifacts like the Docker container Azure DevOps, Google Cloud Build,
and push them to a container repository like Kubernetes
Amazon ECR.
Data processing A data processing service can be used to Amazon SageMaker data processing,
service process raw data into training/validation/test Dataflow, Dataproc, Azure Databricks
datasets for model-training purposes.
Model training The model training service gets the data from Amazon SageMaker training service,
service the data lake, trains the model, and stores Google AutoML, Vertex AI, Azure
the model output back into the output folder Machine Learning Studio
of the data storage, typically S3.
Model registry A model registry is used to manage the AWS Model Registry, AI Hub, Vertex
model inventory, and it contains the AI Model Registry, Azure OpenAI
metadata for the model artifacts, the Code Repository, Azure DevOps,
location, the associated inference container, Azure Machine Learning Model
and the roles. Registry
Model hosting The model hosting service accepts the model AWS Model hosting service, Google
service artifacts from the model registry and deploys AI Platform Prediction, Azure
them into a serving endpoint. Kubernetes Service
Pipeline Orchestrates pipeline workflow using SageMaker pipeline, AWS

management SageMaker pipeline or AWS Step Functions. CodePipeline, AWS CodeCommit,
At the same time, the AWS CodePipeline AWS Lambda, AWS CloudFormation,
integrates with the AWS CodeCommit AWS Step Functions, Cloud Data
repository and AWS CodeBuild to build the Fusion, Azure Data Factory, Azure
code. AWS CodePipeline supports different Machine Learning Pipelines
deployment actions by integrating with AWS
Lambda, AWS CloudFormation, and AWS
Step Functions.
Figure 14.3 shows the implementation of a CI/CD pipeline using AWS components.
This pipeline architecture completely automates the machine learning workflow from code commit, pipeline start,
container build, model building, model registration, and production deployment. In the case of AWS, you can use
tools like CodeCommit, CodeBuild, CloudFormation, Step Functions, and SageMaker Pipelines.
Implementing automated pipelines is the secret sauce behind an enterprise AI platform, as it facilitates automa-
tion, standardization, and reproducibility in machine learning workflows.
TIP By automating the deployment of models in production it becomes possible for com-
panies to scale their AI implementations across the enterprise.
Shared Services Account
Model Approver
CodeStar CodePipeline Registry/Container CodePipeline
Corporate Data Center
ECR S3
CodeCommit
CodeBuild
On prem code repo
CodeBuild SageMaker
Registry
Data Science Account Test/UAT Account Prod Account
CloudFormation AWS Step Functions CloudFormation

SageMaker
Workflow
SageMaker Processing
S3 S3
SageMaker API
SageMaker Training Endpoint Gateway
Lambda
FIGURE 14.3: CI/CD pipeline for ML training and deployment on AWS
EXERCISE: DESIGNING AND IMPLEMENTING A DATA PIPELINE
The goal of this exercise is to give you an idea of the steps to design a data architecture and
implement a data pipeline to gather, clean, and store data for machine learning experiments. In
addition to automating the model training and deployment process, automating the data
pipeline can help you scale.

(JOB TITLES)
Identify data sources: List different data Data engineer A list of different
sources that will be utilized, including data sources
databases, APIs, filesystems, etc.
Design a data flow diagram: Create a Data engineer A diagram showing

visual representation of how data will the data flow
flow through the system, including data
gathering, transformation, and storage.

(JOB TITLES)
Implement an ETL process: Write a script Data engineer A script or tool

or use a data pipeline tool to extract data configuration for
from the identified sources, transform it the ETL process
(e.g., clean, aggregate), and load it into a
data warehouse or database.
Test the pipeline: Ensure that the data Data engineer A report
is correctly processed and stored in detailing the
the desired format and create a report implementation
summarizing the results. and test results
Deployment Scenarios
In the previous sections, we discussed automating the data pipeline, model training and deployment steps. In
this section, let us discuss creating different automation pipelines based on the deployment scenario. This section
reviews these different scenarios. The deployment scenarios can range anywhere from comprehensive end-to-
end pipelines that span from raw code production to production deployment, to more specialized scenarios like
Docker build, or model registration pipelines.
➤➤ End-to-end automated pipeline: This involves taking the source code and data and training the model
all the way to deploying it into production. It involves stages like building the Docker image, processing
data, training the model, getting approval for deployment, and finally deploying the model.
➤➤ Docker build pipeline: This involves building the Docker image from the Docker file and pushing it into
a container repository, such as Amazon ECR, along with associated metadata.
➤➤ Model training pipeline: This involves training the model using an existing Docker container and then
optionally registering it into a model registry.
➤➤ Model registration pipeline: This pipeline registers the model in the ECR registry and stores the
model in S3.
TIP Use a continuous integration/continuous delivery (CI/CD) pipeline to automate the pro-
cess of building, testing, and deploying your machine learning models.
➤➤ Model deployment pipeline: This pipeline deploys the model from the model registry into an endpoint.
Figure 14.4 shows a code deployment pipeline. These pipelines can be triggered by an event, or a source code
commit or through the command line.
In Figure 14.4, a Lambda function is used to trigger the pipeline. The Lambda function can look for a change in
an existing source code file or determine when a new file is deployed to trigger the pipeline accordingly.
Model Inventory Management

Another best practice of MLOps is implementing model inventory management for effective model risk manage-
ment. As the number of models increases in your organization, it becomes important to manage them to ensure
transparency, efficiency, and control over these assets. You can keep track of deployed models including their
versions, metadata, and so on, and monitor and control their performance to ensure they are aligned with busi-
ness objectives. The table shows the different model inventory management tools available from the major cloud
providers.
Stage: Source Stage: Build
1 - Change
committed to
source SOURCE CONTROL CD PIPELINE
CI PIPELINE
FILE REPOSITORY (Amazon
(AWS CodeBuild)
(AWS CodeCommit) CodePipeline)
2 - Change to 4 - Kicks off a

source triggers pipeline
monitoring event deployment
3 - Triggers
MONITORING
a function COMPUTE
SERVICE
PLATFORM
(Amazon
(AWS Lambda)
CloudWatch)
FIGURE 14.4: Code deployment pipeline
CLOUD PROVIDER MODEL INVENTORY FEATURES AND USES

MANAGEMENT TOOLS
AWS SageMaker Model Registry Whenever a new model is developed, you can
register it there, and every time it is trained, the
latest version can be created. It supports
➤➤ Model auditing
➤➤ Model lineage tracking management of asso-

ciated metadata
➤➤ Creation of new model versions
Azure Azure Machine Learning Model ➤➤ Attach metadata to models (model name,
Management Service owner name, business unit, version number,
approval status, and other custom metadata)
➤➤ Manage the deployment of various versions

in the organization across accounts and
environments
CLOUD PROVIDER MODEL INVENTORY FEATURES AND USES

MANAGEMENT TOOLS
Google AI Platform Model Registry ➤➤ Register models and their versions
➤➤ Deploy models to various environments and

deployment targets
➤➤ Create prediction services, deploy custom

containers, or integrate with other Google
services, such as the Kubernetes engine
SHARED CENTRALIZED ENVIRONMENT MODEL REGISTRY

Central Model Store/Container Repository
Model Registry –
Container Registry – Store Model File –
SageMaker Model
Amazon ECR Amazon S3
Registry
2. Create model from

central model
package repo
Data Pipeline PROD ENVIRONMENT

(SageMaker Processing)
Create Model
TEST/UAT (AWS Lambda)
ENVIRONMENT Model Training
(SageMaker Training)
3. Create endpoint
1. Create model package API Gateway

Register Model
and register model (SageMaker Endpoint)
(AWS Lambda)
version
FIGURE 14.5: Centralized model inventory management
There are two main ways to implement model inventory management:

➤➤ Distributed model management: Models developed manually or generated by various automation pipe-
lines can be stored in the respective accounts. For example, different organization units can have their
models registered to their own ML UAT/test accounts.
➤➤ Centralized model management: As shown in Figure 14.5, the models developed by the automation
pipelines can be stored in a shared services account. You can create a model package group to store the
different versions associated with the model. At the time of deployment, a particular version is used to
deploy from the package group.
NOTE Implement model inventory management to track model lineage, manage model
metadata, and audit model use.
Logging and Auditing

Tracking the operations carried out by various platform components for auditing is essential. By tracking the ML
operations, you can assist with model debugging, performance monitoring, audits, and health monitoring of the
ML platform. Figure 14.6 shows a logging and auditing reference architecture.
The following table lists the logging and auditing tool features provided by AWS, Azure, and Google:
CLOUD LOGGING AND FEATURES AND USES

PROVIDER AUDITING TOOLS
AWS AWS CloudTrail ➤➤ Tracks operations for AWS services.
➤➤ Stores logs in S3.
➤➤ Logs contain event name, user identity, event time, event

source, and source IP address.
➤➤ Users can access logs directly from the AWS console.
➤➤ Integration with Splunk for downstream analysis.
➤➤ Tracks activities within the SageMaker instance.
➤➤ Good practice to create roles for each user for attribution

during auditing.
Azure Azure Activity Logs ➤➤ Tracks operations performed on Azure resources.

Azure Monitor ➤➤ Logs can be sent to either Azure storage or Azure monitor
Azure Storage for further analysis.
Azure Sentinel ➤➤ Azure storage provides options such as Azure
Blob Storage.
➤➤ Azure Sentinel provides advanced security monitoring, inci-

dent response, and reporting.
Google Google Cloud Audit Logs ➤➤ Captures logs that are sent to other downstream systems
Google Cloud Storage, for analysis.
Google Cloud Logging ➤➤ Stores logs in Google Cloud storage buckets.
Google’s Security ➤➤ Logs can be analyzed using Google’s centralist monitoring
Command Center
system, Google Cloud Logging.
➤➤ Security command center offers security and compliance
monitoring, vulnerability scanning, and threat detection
capabilities.
TIP Implement a robust auditing system to track all activity and monitor suspicious activity.
Create roles for each user so that proper attribution can be made for these activities during
auditing.
Data and Artifacts Lineage Tracking

Tracking the data and artifacts lineage is essential, as it will help you reproduce the entire data processing, model
building, and deployment processes to meet regulatory and compliance requirements. This section reviews the
different artifacts that you need to track, as shown in Figure 14.7.
PRODUCTION
ACCOUNT
CENTRAL OPS
Experimentation Auditing and Logging ENVIRONMENT
• ML Platform - SageMaker • Audit – CloudTrail
• Data Lake - S3 • Notification – Automation Pipeline Components
• Container Store - ECR SNS
• CodePipeline
• CodeBuild
QA/UAT
• Data Lake
ACCOUNT
• Container Registry
Experimentation Auditing and Logging

• Data Lake - S3 • Notification –
Central Logging Log Analysis
DEVELOPMENT • SQL Query
ACCOUNT Analyzer – Athena
• Audit – CloudTrail
• Log Analysis UI –
• Notification – SNS
Experimentation Auditing and Logging CloudWatch
• Central Data Lake – S3
• Big Data Analysis
- Splunk
• Data Lake - S3 • Notification –
FIGURE 14.6: Logging and auditing architecture
This table lists the tools provided by AWS to track code, datasets, metadata, container, training job, model, and
endpoint versions:
ARTIFACT TRACKING TOOLS DESCRIPTION
Code Versioning GitLab, Bitbucket, Code repositories help you track the versions of the code
CodeCommit artifacts. You can check in and check out using a commit ID,
which you can use to track the versions of the artifact.
Dataset Versioning S3, DVC You can track the versions of your dataset by using a proper
S3 data partitioning scheme. When you upload a new dataset,
you can add a unique S3 bucket/prefix to identify the dataset.
DVC is another open-source dataset versioning system that
can be integrated with source code control systems such as
GitLab and Bitbucket when creating new datasets, and S3 can
remain your backend store.
Metadata Tracking SageMaker Lineage Metadata associated with datasets can be tracked by
Tracking Service providing details such as the dataset properties and the
lineage.
Container Amazon ECR Amazon ECR tracks the container versions using the image
Versioning (Elastic Container URL containing the repository URI and image digest ID.
Repository)
Training Job SageMaker SageMaker’s training job contains its ARN name and other
Versioning metadata, such as URIs for hyperparameters, containers,
datasets, and model output.
Model Versioning Model Packages Model versions can be tracked using model packages that
contain their own ARN names and details such as the model
URI and the container URI.
Endpoint SageMaker As part of creating the SageMaker endpoint, you can add
Versioning other metadata, such as the model used to the endpoint
configuration data
NOTE Implement a data and artifact lineage tracking system to understand how models
work and identify potential problems.
This table lists the tracking tools provided by Azure:
Code Versioning Azure DevOps Tracks code versions and integrates well with Git
repositories.
Dataset Versioning Azure Machine Learning Used to track dataset versions.

Datasets
Metadata Tracking Azure Machine Learning Helps track metadata such as model training metrics,
Services dataset properties, and experiment configurations.
Container Azure Container Registry Keeps track of container images and their versions.
Versioning
Training Job Azure Machine Learning Manages training job versions and other parameters
Versioning Services such as hyperparameters, training data, and outputs.
Model Versioning Azure Machine Learning Tracks model versions. Models can be deployed as
Services web services and metadata such as model properties,
deployment configurations, and version information
can be tracked.
Artifact Versioning Azure DevOps, Azure Data Used to track the versions of artifacts throughout the
Factory, Azure Data Catalog machine learning lifecycle.
This table lists the tracking tools provided by Google:
Code Versioning Git, Google Cloud Source Supports integrating Git and Google Cloud source
Repositories repositories to help track code versions.
Dataset Versioning Google Cloud Storage Datasets can be stored in Google Cloud Storage in
versioned buckets, which allows you to track different
versions of the datasets.
Metadata Tracking Cloud Metadata Catalog Can track different versions of model datasets and
other artifacts by associating metadata such as
properties, descriptions, and lineage.
Container Google Cloud Container Can track different versions of container images.
Versioning Registry
Training Job Google Cloud AI Can track different versions of training jobs, models,
Versioning Platform and endpoints.
Artifact Versioning Kubeflow, Data Catalog, Provides services to track artifacts throughout the
Dataflow machine learning lifecycle.
Code Versioning Git, Google Cloud Source Supports integrating Git and Google Cloud source
Repositories repositories to help track code versions.
Code Repository Data Repository

Experiments Data Files
• Python notebooks • Raw data
and code • Processed data
ML Training & Deployment • Labeled data
• Python code • Training data
• Docker files • Validation data
• BuildSpec files • Inference data
• CloudFormation • Inference feedback
templates
Manifest Repository
Release
Model Registry • Container IDs
Images • Model name
• Data processing • Model binary path
• Model training • Instance type and count
• Model deployment • Monitoring config
Model Binaries Build
• model.tar.gz • Container IDs
• Input data
• Data processing
• Compute
• Training
• Artifact registry
• Execution outputs
FIGURE 14.7: Data and artifacts lineage tracking

Google Cloud also provides services such as Kubeflow, Data Catalog, and Dataflow to track artifacts throughout
the machine learning lifecycle.
Container Image Management

Container image management is a vital step to enable MLOps, because it ensures that the models are deployed in
consistent environments. Container orchestration tools such as Kubernetes play a critical role in automating the
deployment of ML models and platforms like AWS, Azure, and Google Cloud support this.
The following table lists several types of images that need to be managed:
TYPE OF IMAGE STORAGE LOCATION/OWNERSHIP DESCRIPTION
Public images Docker Hub, GitHub, ECR Public Publicly available images stored in public
Gallery repositories.
Base images Customer ECR Instance Foundational images made from public
images. They contain additional OS patches
and other hardening.
Framework images Central Team Built on top of base images and represent
stable versions of the environment. They may
include specific versions of frameworks such
as TensorFlow Serving.
Application Application Teams Created for specific environments and built

frameworks on top of framework images. They contain
additional sources, binaries, or model
artifacts.
NOTE A secure container registry should allow you to store container images and provide
image signing, vulnerability scanning, role-based access control, and image auditing features.
The following are container management-related best practices using AWS to meet security and regulatory
requirements:
BEST PRACTICE TOOL/METHOD DESCRIPTION
Security controls Security patches Implement security controls, consistent patching, and
vulnerability scanning to meet security and regulatory
requirements.
Organization Namespace Organize repositories using namespace conventions to

conventions group repositories together.
Versioning Image tags Use image tags to indicate specific versions of containers,
and unique tags can be used to track GitHub repository
branches.
Access control Identity access Implement granular access control for containers by
management adopting resource-based and identity-based policies.
Container security Regular scanning Implement container security by regularly scanning

vulnerabilities and patching.
Cross-account IAM policies Enable cross-account access for repositories stored in

access shared accounts to users from other AWS accounts.
This table lists Azure-related security management best practices for containers:
Image building Azure Pipelines Use Azure Pipelines, a CI/CD service, to build your images in
an automated fashion.
Image storage Azure Container Store your containers in Azure Container Registry, a secure
Registry and scalable repository.
Policy Azure Policy Use Azure Policy to enforce organizational security policies
enforcement on your images.
This table lists Google-related security management best practices for containers:
Image building Cloud Build Use Cloud Build to build your images.
Image storage Google Container Repository Store your images in GCR.

(GCR)
Policy Cloud Policy Use Cloud Policy to implement security and

enforcement compliance policies.
Image Google Kubernetes Engine Use GKE to deploy your images.

deployment (GKE)
Tag Management
Tag management is another critical component of MLOps as it helps to track ML experiments and model ver-
sions. In addition to all the things discussed about using tags to manage resource usage, you can also use tags to
manage other tasks, covered in this section.
Control SageMaker Endpoints and Studio Domain Creation

Using tags, you can control the creation of SageMaker endpoints and SageMaker studio domains (see
Figure 14.8).
➤➤ SageMaker endpoints: The SageMaker endpoint functions as an API exposed to other endpoints to make
the calls with the required input data to obtain predictions.
➤➤ SageMaker studio domains: These platforms help data scientists, machine learning engineers, and devel-
opers collaborate. They provide capabilities such as Jupyter notebooks, integrated debuggers, experi-
ment environments, model registries, and Git integration.
Track resource usage,

cost management,
billing, and access
control
Identify different ML
environments
Identify the Create a tag key
Assign the tags to
resources that and value for
the resources
you want to tag each resource
Automatically delete
resources
Control access to AWS

resources
FIGURE 14.8: Using tags to track resource usage, cost management, billing, and access control
Cost Management and Billing

The following table reviews the tagging and cost management strategies of Azure and Google:
PROVIDER TAGGING STRATEGY COST MANAGEMENT STRATEGY
Azure Azure resource manager tags help you attach The Azure cost management and billing
metadata to various resources using tags to service helps you control the usage of
control their cost, usage, organization, and ML resources and manage its budget
access control. You can assign these tags to accordingly. The Azure machine learning
virtual machines, storage systems, and ML service controls costs by setting up
platform components. These tags can be compute targets, for instance, types and
assigned based on organizational units, cost sizes.
centers, environments, or projects. Azure also provides cost management
You can use the Azure Policy to mandate that APIs that can help you retrieve costs and
the tag should be assigned to all resources, usage data related to various resources,
which can help you ensure governance, cost which can be used in custom reporting for
control, and enforce resource usage limitations. tracking purposes.
Google Google provides similar resource labels to tag You can also control costs using Google’s
virtual machines, storage buckets, models, and budget and billing alerts, AI platform
other ML components. auto-scaling, BigQuery cost control, and
Using the Google Cloud policy, you can cloud billing API.
mandate that tags be associated with various
resources, and these tags can be maintained at
the organization, folder, or project level.
Tagging to Identify Different ML Environments

Tagging can differentiate between experimentation, automation, and production environment resources like
SageMaker endpoints and model artifacts. Business tags can track ownership, monitor AWS spending, and create
financial reports. Automation-related tags can be used to shut down specific resources during certain weeks while
keeping other resources running.
TIP You can use tags to track resource ownership, automate tasks, and comply with regula-
tions.
Automatic Deletion of Resources

Tagging can be used to identify resources that need to be removed automatically. You can use this capability to
delete resources as part of a workflow, put specific resources in quarantine, or delete noncompliant resources,
such as those that are not tagged. You can automate the allocation and deallocation of resources using MLOps
thus controlling costs optimally.
Controlling Access to AWS Resources

You can also enforce access to AWS resources by mandating that the resources have a specific tag to allow access.
Security-related tags can be used to apply various levels of security controls based on their sensitivity. Role-based
access is also critical in MLOps to ensure only authorized individuals can access and modify models and data.
HANDS-ON EXERCISE: SETTING UP AI PLATFORM WITH MLOps
This hands-on exercise aims to set up an enterprise AI platform with MLOps best practices
that contains the major architectural components such as a data pipeline, model experimenta-
tion platform, MLOps environment, workflow automation and CI/CD pipeline, performance
optimization, platform monitoring, security, and governance control practices.
Also note that the roles used here are for general guidance, and the actual roles used in our
company will depend on your org structure, operating culture, and other considerations.
Finally, note that each of the following tasks depends on other tasks, and they have to come
together holistically to contribute to the effective functioning of the platform as a whole:

(SKILLS)
Design Data Pipeline: Design a Data engineer A list of different data

data pipeline to ensure that data sources; diagram
is properly gathered, cleaned, and showing data flow
stored for use in machine learning
experiments.
continues
continued
(SKILLS)
Design your ML Experimentation Data scientist Detailed report on

Environment: Setting up an setup; diagram of
experimentation platform to conduct interaction with
machine learning experiments, data sources
such as model training, testing, and
validation.
Note: MLOps solutions can streamline
experimentation, making it easier to
compare different model versions and
experiments.
Document MLOps setup: ML engineer Detailed report on

Documenting the steps to deploy model execution,
trained models in a production serving, and monitoring;
environment, serving the models, and diagram illustrating
monitoring their performance. model flow
Note: The MLOPs setup is a
critical step for automating model
deployment, monitoring, and
retraining.
Design Workflow Automation DevOps engineer Report describing the

and CI/CD Pipeline Architecture: automation process;
Designing the architecture to diagram of automation
automate repetitive tasks and integration
establish a continuous integration/
continuous delivery pipeline.
Design Security Architecture: Security officer Security architecture

Designing the security architecture to and plan outlining
ensure all data and models are secure, measures; diagram
manage access controls, and meet of integration of
regulatory requirements. security measures
Configure Performance Performance Report on performance

Optimization: Ensure the system engineer optimization measures;
works at optimal levels and makes the diagram of how
most of available resources. measures integrate into
the system
Configure Platform Monitoring: System Report on monitoring

Monitoring system performance and administrator system; diagram of how
health, and document methods to this system integrates
send alerts if any issues arise. into the system

(SKILLS)
Document Governance and Control Compliance Report on governance

Practices: Ensuring all operations are manager and control measures;
controlled, auditable, and traceable, diagram illustrating
and maintaining compliance with how these measures
various regulations. are integrated into
the system
WORKBOOK TEMPLATE: ML OPERATIONS AUTOMATION GUIDE

Download the “ML Operations Automation Guide” template from the download section of the book (www
.wiley.com/go/EnterpriseAIintheCloud). Use this template to put together an operations automation
guide to help scale AI in your enterprise. If done right, it can help you to reduce operational costs, increase
efficiency, and increase productivity to allow employees to devote more time to innovation rather than
maintenance.
SUMMARY
In the world of AI/ML production environments, three pillars stand out: efficiency, scalability, and reliability. To
achieve those goals, automating the ML lifecycle is essential. MLOps is the secret sauce to achieve that automa-
tion, covered in this deep, technically intensive chapter. By enabling this automation, MLOps serves to bridge ML
development and operations, thus bringing speed to deployment. This chapter focused mainly on cloud engineers
and other IT professionals responsible for operating and automating as much of the ML lifecycle as possible.
Moreover, business and technology strategists stand to benefit, given the accelerated speed to market and scale
that MLOps brings.
The chapter covered a number of key concepts, such as container images and their management, using tagging
for access control, tracking resource usage, and cost control. In addition, it explored platform monitoring, model
operationalization, automated pipelines, model inventory management, data, and artifact lineage tracking. It
discussed best practices using AWS, Azure, and Google Cloud platforms, covering topics such as security controls,
organization, versioning, and access control.
The chapter provided hands-on exercises for setting up a modern enterprise AI platform incorporating the latest
MLOps practices, encapsulating crucial architectural components like data pipeline, MLOps environment, CI/
CD pipeline, and governance practices. These exercises structured around real-world roles like data engineer, data
scientist, and ML engineer not only teach the MLOps methodology but also increase awareness around the roles
and responsibilities of these roles.
This chapter is tailored toward professionals working in fields such as data engineering, machine learning,
DevOps, and cloud architecture, providing both theoretical knowledge and practical tools to implement robust AI
platforms. Whether a beginner or an experienced practitioner, you are likely to gain insights that will further your
understanding of the complex landscape of modern AI infrastructure.
In the next chapter, let’s discuss how to process data and engineer features.
REVIEW QUESTIONS
1. The components involved in model operationalization are

A. Model execution, model monitoring, and model deployment
B. Model hosting, model evaluation, and model consumption
C. Model training, data preprocessing, and feature selection
D. Data visualization, model tuning, and model evaluation
2. Which of the following components is not part of an automated pipeline for machine learning?
A. Code repository
B. Data processing service
C. Model training service
D. Data visualization tool
3. The Lambda function is used in the pipeline architecture for which of the following?
A. Building the Docker image and pushing it to a container repository
B. Training the model using an existing Docker container
C. Registering the model in the ECR registry and storing it in S3
D. Triggering the pipeline based on a change in source code or deployment of a new file
4. Which of the following is NOT a benefit of platform monitoring?
A. Improved accuracy
B. Improved reliability
C. Improved speed
D. Reduced costs
5. _______ is NOT a benefit of using a model inventory management system.
A. Improved model lineage tracking
B. Improved model metadata management
C. Improved model auditing
D. Reduced model development costs
6. Which of the following is NOT a method or tool used for tracking data and artifacts lineage in
machine learning?
A. Code repositories such as GitLab and Bitbucket
B. Dataset versioning using S3 data partitioning
C. Azure DevOps for tracking code versions
D. Kubeflow, Data Catalog, and Dataflow in Google Cloud
7. Which of the following tasks can be managed using tags in cloud platforms like AWS, Azure, and
Google Cloud?
A. Controlling access to AWS resources
B. Deploying containers using Azure Pipelines
C. Managing DNS settings in Google Cloud
D. Monitoring network traffic in AWS
8. _______ is NOT a technique for improving latency in hybrid and edge computing.
A. Edge caching
B. Preprocessing
C. Local inference
D. Round-trips to the cloud
Answer Key ❘ 239
ANSWER KEY
1. A 4. A 7. A
2. D 5. D 8. D
3. D 6. D
PART VI
Processing Data and Modeling
Data is the lifeblood of AI. Part VI is where you get your hands dirty with data and modeling. I teach you
how to process data in the cloud, choose the right AI/ML algorithm based on your use case, and get your
models trained, tuned, and evaluated. It is where science meets art.
15
Process Data and Engineer
Features in the Cloud
The world is one big data problem.
—Andrew McAfee
In the world of AI, the true essence lies not just in sophisticated algorithms, but in the quality and struc-
ture of data. As most AI practitioners would attest, raw data is rarely ready to be used as is. This is where
cloud-based data processing comes in. It refers to the process of collecting, ingesting, storing, preprocessing,
engineering, and processing data for use in machine learning models through cloud-based data technologies
from AWS, Azure, and Google.
So why is data so vital? Simply put, the most sophisticated model is only as good as the data on which it is
trained. Incorrectly processed data or poorly engineered data can lead to inaccurate predictions and deci-
sions costly to your business.
As companies try to embrace the scalability and power of the cloud, learning how to process data in the
cloud becomes paramount. Cloud provides several benefits, including scalability, performance, security,
flexibility, and cost-effectiveness.
In this chapter, you learn about exploring your data needs and the benefits and challenges of cloud-based
data processing, and you dive into hands-on exercises, including feature engineering and transformation
techniques.
This chapter underscores the importance of data augmentation, showcases methods to handle missing data
and inconsistencies, and presents the art and science of feature engineering, all happening in the cloud. See
Figure 15.1.
This chapter addresses many topics, including data storage and exploration, data storage architectures, and
distributed data processing. Always remember, in the end, data is king—it is not just the starting point, but
the foundation for everything else when it comes to AI.
244 ❘ CHAPTER 15 Process Data and Engineer Features in the Cloud
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Govern Your
Your AI
journey Team Process Data and Modeling
09 • Process Data and Engineer
ENTERPRISE AI Features in the Cloud
Evolve and
Setup 04 • Choose your AI/ML Algorithm
and Manage • Train, Tune, and Evaluate Models
Operations
08 Objective
Master cloud-based data processing by
AI
Deploy and 05 understanding, collecting, and refining data,
Monitor
Models coupled with advanced feature engineering
07 techniques, to ensure optimized input for
AI/ML model training
06
FIGURE 15.1: PROCESS DATA AND MODELING: Process data and engineer features in the cloud
UNDERSTANDING YOUR DATA NEEDS

You are now at a critical stage in your ML project, where you have a good idea of the business problem you need
to solve. For example, consider a use case of a retail company. Assume you are working for them and trying to
predict the next quarter’s sales so they can plan their inventory, staffing, and other resources. To predict the sales,
you need to develop a machine learning model. Armed with this information, you can now carry out these steps
in the following hands-on exercise.
HANDS-ON EXERCISE: UNDERSTANDING YOUR DATA NEEDS
Problem Statement: The management of a retail chain store is facing problems in predicting the
sales for the upcoming quarter. In the past, inaccuracies in the forecasts have resulted in either
overstocking or understocking resulting in inefficient resource allocation or missed revenue
opportunities. The management wants to leverage the company’s past data to develop a
machine learning model that can predict sales more accurately for the next quarter across their
retail stores. They need this prediction a few weeks before the start of the quarter to allow
effective inventory planning, staffing, and other resource planning. The goal is to minimize
waste and increase profitability.
Here are the steps your team can follow:
STEP TASK EXAMPLE TASK OWNER DELIVERABLES
(SKILL SETS)
Understanding First you need to consult stakeholders to get Stakeholders want to predict sales Project A clear
the business a clear understanding of what they want to a few weeks before the start of manager, statement of
problem predict their time frame. the next quarter. business the problem,
analyst goals, and
objectives of
the project
Identifying You can now start understanding the data Past sales data is essential, along Data A
data needs needs, such as what kinds of data are with data related to promotional scientist, data comprehensive
required and the potential sources for the offers, seasonal trends, price analyst list of required
same. List what types of data you might changes, customer reviews, data and
need for such a prediction and possible competitor data, economic potential data
sources. indicators, and so on. sources
Gathering You then try to identify the data sources Collect data from company’s Data Collected
data and decide to gather the past sales data databases, marketing team, engineer, raw data
from the company’s databases, promotional publicly available economic data, data analyst from different
offers, and seasonal data from the marketing industry reports, etc. sources
team, customer reviews from the website,
economic indicators from the public
databases, and competitor data from
industry reports. Do not underestimate the
amount of effort needed to clean this data.
STEP TASK EXAMPLE TASK OWNER DELIVERABLES
(SKILL SETS)
Data cleaning Once you gather the data, the next step is Handle missing sales figures, Data Cleaned and
to clean the data, which may have missing correct inconsistencies in scientist, data standardized
values, outliers, and other inconsistencies. promotion categorization, remove analyst dataset
outlier values in sales figures.
Exploratory Explore the data for seasonal trends, and the Identify patterns in sales based on Data Report on the
Data Analysis relationships between various data elements. promotions, holidays, competitor scientist, data findings of the
(EDA) This will eventually help you choose the activities, economic indicators, analyst EDA, including
suitable ML model and guide you in etc. visualizations
preparing the data for your model. and initial
insights
Feature Based on EDA, brainstorm potential new Create features like frequency of Data New features
engineering features that could improve your model’s promotions, duration since last scientist, added to
and model performance. promotion, average spending per machine the dataset,
building Prepare the data for consumption in machine visit, etc. learning ML model
learning models. engineer prototypes
Benefits and Challenges of Cloud-Based Data Processing ❘ 247
NOTE While data may be available internally, one point that is often overlooked is the suit-
ability of the data for consumption by the required algorithms.
BENEFITS AND CHALLENGES OF CLOUD-BASED DATA PROCESSING

Processing data in the cloud has its own benefits and challenges, each discussed next.
Benefits of Cloud-Based Data Processing

Some of the benefits are large computing capacity, data integration, and preparation capabilities:
➤➤ Computing capacity: Regarding processing large amounts of data, it is crucial to have a large computing
capacity to process data efficiently and quickly. Unlike the on-premises servers, which may be limited in
capacity, cloud-based services provide scalable and robust cloud servers.
➤➤ Data integration: Note that your data sources may also be distributed in disparate sources. For example,
sales data may be stored in S3, customer reviews in Google Cloud storage, and economic indicators in
a public database hosted in Azure. Each cloud service has several tools that allow these data to be inte-
grated and processed quickly. For example, you could use AWS Glue for extract, transform, and load
(ETL) and Amazon Athena to query data directly from S3. Google Cloud’s BigQuery can run SQL que-
ries on the customer review data.
➤➤ Data preparation: Cloud-based tools such as AWS Glue can be used to clean the data for missing values,
re-engineer features, and prepare the data to be used in the ML models.
TIP Cloud-based tools like AWS Glue can clean the data, handle missing values, engineer
features, and prepare data for the ML model.
Challenges of Cloud-Based Data Processing

Handling data in the cloud can pose security challenges and raise cost concerns. Consider these issues:
➤➤ Security and compliance: When the data is stored in the
cloud, it is essential to ensure it is securely stored and pro-
cessed and complies with privacy regulations like GDPR
or CCPA. While cloud providers may provide many tools,
such as access control and data encryption, you must have
in-house skills to implement them appropriately. Structured
Data
➤➤ Cost control: Another factor to consider is the aspect of 01
02
cost control, because even though the cloud providers pro-
vide a pay-as-you-go model, the cost can pile up as large Semi-
volumes of data are being processed. Managing costs is structured
especially important when dealing with Gen AI models. Data
Unstructured
Data
Handling Different Types of Data 03
Another challenge when preparing the data is to handle different
types of data, as shown in Figure 15.2. You can think of the predic-
tions as the meal, the different AI, ML, and deep learning models are
the cooks, and the data is the ingredients. No matter how good the
models are, the predictions will be good only if the data is correct. FIGURE 15.2: Data types
As noticed in the previous case study about the retail company, there are different data types, such as structured,
unstructured, and semi-structured data. Each data type has a different purpose, format, and suitable models, as
shown in the following table:
DATA TYPE EXAMPLE PURPOSE FORMAT AND PROCESSING SUITABLE MODEL
Structured data Historical Helps identify Comes as CSV file or a SQL Regression
sales data sales trends database models for
and seasonality predicting future
sales
Unstructured Customer Provides Can be messy, requires Deep learning

data reviews insights into natural language processing models to
customer for feature extraction identify patterns
preferences or in customer
sentiments sentiment
Semi-structured Economic Helps assess Comes as XML or JSON Decision trees to

data indicators how economic files, predefined structure evaluate impact
indicators but not as rigid as in a table of economic
impact sales indicators on sales
Different ML models can use this data, such as a regression model for predicting future sales that uses structured
sales data well. However, a deep learning model can be used on unstructured customer review data to identify
patterns in customer sentiment. On the other hand, a decision tree may work well with semi-structured data, such
as economic indicators, to assess how economic indicators impact sales.
TIP Match the data types with the most suitable models and processing techniques.
Going back to the meals analogy, the better the quality and variety of the ingredients used, the better the meal—in
this case—the predictions.
HANDS-ON EXERCISE: HANDLING DIFFERENT DATA TYPES
This hands-on exercise can help make this concept of handling different data types clearer.
Problem Statement: A retail company is facing challenges dealing with handling multiple data
types such as structured, unstructured, and semi-structured data to use in the ML models. Your
task is to guide them to build a clean, consolidated dataset to use in their ML models.
STEP NAME DESCRIPTION TASK OWNER DELIVERABLES

(SKILL SETS)
Process First you can import the Data analyst, An organized

structured data historical sales data from data scientist CSV file or SQL
(such as historical the company’s database database of
sales data). into a CSV file or SQL historical sales
database. data
Benefits and Challenges of Cloud-Based Data Processing ❘ 249

(SKILL SETS)
Process For unstructured data, you Natural A processed

unstructured data can use NLP techniques language dataset of
(such as customer to extract features from processing customer reviews
reviews). customer reviews data. expert, data with relevant
scientist features extracted
Process semi- Then for semi-structured Data analyst, A processed

structured data data such as economic data scientist dataset of
(such as economic indicators, you can economic
indicators). clean and process them. indicators
This data is crucial
to understand the
macroeconomic trends that
might influence sales.
Align datasets. Now ensure that all the Data analyst, Datasets that are
structured, unstructured, data scientist aligned based on
and semi-structured time, e.g., daily,
datasets are aligned weekly, or monthly
properly and aggregate
them as needed.
Match datasets. Identify differences in Data analyst, Matched datasets

the datasets and try to data scientist that are correlated
match them as needed. to each other by
For example, if you have the right entities,
sales data that is based on such as store,
stores but the economic region, etc.
indicators are based on
region, then try to match
the store to the correct
region.
Consolidate Once the data has been Data analyst, A single,

datasets. aligned and matched, you data scientist consolidated
can consolidate them into dataset ready to
a single table or a single be cleaned
database with multiple
tables depending on the
complexity and data size.
continues
continued

(SKILL SETS)
Check and clean Perform a quality check for Data analyst, A high-quality
datasets. missing values, outliers, data scientist dataset free of
and other inconsistencies missing values,
and correct them. outliers and other
inconsistencies
and is ready for
machine learning
model training
THE DATA PROCESSING PHASES OF THE ML LIFECYCLE

This section discusses the data processing phases of the ML lifecycle. The data processing workflow involves data
collection and ingestion and then data preparation.
Data Collection and Ingestion

Once you have identified the data needs, the next step is to collect and ingest the data from various sources.
Figure 13.3 shows the various activities involved in the collection phase. These activities are labeling, ingesting,
and aggregating.
The following table explains some of the data collection methods:
STAGE DESCRIPTION
Labeled data Data about which you already know the target answer. If the label is missing, the activity
involves labeling the missing data either manually or using automated means.
Data collection Depending on the use case, data can be collected from various sources, such as sensors,
time-series data, event data, IoT data, and social media.
In the retail company case study, you may need to use an API or a SQL query to store
the past sales data, pricing data, or promotional offers data in a cloud storage such as
S3.
Web scraping Use web scraping to gather the review data and handle real-time ingestion for customer
reviews, as review data is continuously trickling in. The marketing data can come as
semi-structured data such as an XML, JSON, or CSV file, which you need to ingest into
the cloud environment.
Direct downloads In the case of economic indicators data, you can use the APIs offered by the financial
institutions or the government bodies or a direct download to extract data and use
different methods to ingest data into the cloud.
Data ingestion Collected data must be ingested into storage facilities, through either real-time or batch
methods processing.
The Data Processing Phases of the ML Lifecycle ❘ 251
Figure 15.3 shows the activities involved in a data collection process.
Data Collection
Label
Ingest (Streaming, Batch)
Aggregate
FIGURE 15.3: Data collection process
NOTE Data can be collected from various sources such as sensors, time-series data, event
data, IoT data, and social media using API or SQL query, web scraping, or direct down-
loads.
Data that’s collected must be ingested into various storage facilities using real-time or batch processing. The fol-
lowing table compares the pros and cons of the two methods:
DATA INGESTION DESCRIPTION EXAMPLE USE CASE PROS CONS

METHOD
Real-time data For real-time Customer reviews can Handles Requires specific
ingestion data ingestion, be smaller in volume but high tools and systems
you must use high in velocity as new velocity and is not suitable
a real-time reviews keep coming in. data; for large volume
processing For this scenario, a real- ensures data.
system such as time ingestion method is data is
Apache Kafka appropriate. current
or AWS Kinesis.
Batch data Processes large The velocity with which Handles Data may not
processing volumes of the data comes in can large always be current
data at specific vary; for example, sales volume and is not suitable
time intervals. data can come in large data; for high velocity
quantities but be updated doesn’t data.
once a day or weekly. In require
such situations, you may immediate
use a batch ingestion processing
process.
NOTE Choose real-time or batch ingestion methods and tools based on the velocity and
volume of data. Unless you are positive that you are doing one-time pulls of data from a
source, consider automation as an important part of your AI/ML journey.
Data Storage Options

After collecting your data, you need to store the data for further processing and analysis. You have many options
to store such data depending on the data type. Data storage options vary and can be any of the following:
➤➤ Transactional SQL databases
➤➤ Data lakes
➤➤ Data warehouses
The following table lists the pros and cons of these options:
STORAGE OPTION PROS CONS
Transactional SQL Ideal for structured data and supports Not ideal for handling large
databases real-time processing volumes of data.
Data lakes Useful for storing a vast amount of raw Data may need further processing
data in its native format before it can be used.
Data warehouses Optimized for analyzing structured data Might not be the best choice for
and performing complex queries unstructured or semi-structured
data.
The following table lists the storage options from various cloud providers:
STORAGE DESCRIPTION PROS CONS

OPTION
Amazon S3 In the retail company • It is scalable. • The trade-off with S3 is its

example, you can • It can also cost because even though it is
store the sales and seamlessly reasonably priced, it can rise if
promotional events integrate with you have to handle large volumes
data in Amazon S3 other AWS services of data more frequently.
(Simple Storage that you use to • Moreover, it may need to
Service), which is carry out further integrate better with other non-
highly scalable and processing. AWS services.
grows with your
needs.
Google Cloud Google Cloud • It is scalable. • It can also become expensive.

Storage Storage is another • It has good Google • It may need to integrate better
option, as it’s also Cloud services with non-Google Cloud services.
scalable, like S3, integration. • It is also not considered as
and integrates well comprehensive or mature as S3,
with other Google although it’s been catching up in
Cloud services such recent years.
as Natural Language
Processing.
Understanding the Data Exploration and Preprocessing Stage ❘ 253
STORAGE DESCRIPTION PROS CONS

OPTION
Azure Blob Azure Blob Storage • It is scalable. • It can also become more
Storage can also be a good • It has good Azure expensive.
choice due to its Cloud services • Integration with non-Azure
scalability and integration. services can be challenging.
integration ability with • It is also less user intuitive than
Azure Cloud services. S3 or Google Cloud Storage.
TIP When choosing a data storage option, it’s crucial to consider factors such as cost,
user-friendliness, integration with other services, in-house skill sets, tiered storage, disaster
recovery, and data type and volume.
When choosing these storage options, you should evaluate them against the following factors:
➤➤ Cost
➤➤ User-friendliness
➤➤ Integration ability with other services
➤➤ In-house and easily acquirable skills
➤➤ The ability to have tiered storage where you can store infrequently accessed data to lower cost storage
➤➤ Multiregional storage for disaster recovery
UNDERSTANDING THE DATA EXPLORATION AND PREPROCESSING

STAGE
Once your data is stored in the cloud, the next step is to explore and clean it and then ensure it is ready for model
building. This is similar to how you clean your ingredients before they go into the pot for cooking.
According to one of the data science surveys, more than 66 percent of a data scientist’s time is spent cleansing and
loading data and doing the visualization activities, as shown in Figure 15.4.
Data Preparation
Data preparation is the process of preparing the collected data for further analysis and processing. Data prepara-
tion includes data preprocessing, exploratory data analysis, data visualization, and feature engineering.
Figure 15.5 shows a data processing workflow. Note that your project may have different workflows with dif-
ferent steps.
➤➤ Data preprocessing: Data preprocessing is the stage where data is cleaned to remove any errors or incon-
sistencies.
➤➤ Exploratory data analysis: Data exploration or exploratory data analysis is about looking at what’s in
your dataset. It is part of data preparation and can be used to understand data, carry out sanity checks,
and validate the quality of the data.
➤➤ Methods used: It involves generating summary statistics, visualizing the distribution of data and
variables, and identifying relationships between variables. You can use out-of-the-box features
provided by various cloud ML platforms, data visualization tools such as Tableau and Google
Data Studio, and software libraries such as Pandas and Matplotlib in Python.
➤➤ Example: In this case study example, you can explore summary statistics such as mean, median,
and range of sales per day or product. You can create histograms to visualize this data or create
scatterplots to identify relationships between sales and other variables such as price and promo-
tional events.
➤➤ Feature engineering: Feature engineering is the phase where new features are created from existing data
to improve the performance of the models.
➤➤ Data visualization: Data visualization can help with exploratory data analysis by providing visual repre-
sentations of data to better understand patterns, trends, and correlations.
Approx. Time Spent in Tasks
Deploying models
10%
Data loading
20%
Model training and

scoring
12%
Model selection
10%
Data cleansing
26%
Data visualization
22%
Data loading Data cleansing

Data visualization Model selection
Model training and scoring Deploying models
FIGURE 15.4: Average time spent in ML tasks
TIP When preparing data, ensure thorough data preprocessing for cleaning, exploratory
data analysis for understanding, feature engineering for model performance enhancement,
and data visualization to discern patterns and correlations.
Data Preprocessing
Data preprocessing is about preparing the data to be in the right shape and quality for model training.
Figure 15.6 shows the different strategies available for data preprocessing.
Data Wrangling
Data
Preprocessing
Data Data
Data Collection EDA
Preparation Visualization
Feature
Engineering
FIGURE 15.5: Data processing workflow
Data Preprocessing
Clean (replace, impute, remove outliers, duplicates)
Partition (train, validate, and test)
Scale (normalize, standardize)
Unbias, balance (detection and mitigation)
Augment
FIGURE 15.6: Data preprocessing strategies
Data wrangling is part of the data preparation process. It involves the process of transforming and mapping data
from one “raw” format into another to allow subsequent analysis and consumption.
Using tools such as SageMaker Data Wrangler will help you prepare data using its visual interface to access data,
perform EDA, perform feature engineering, and then seamlessly import the data into the SageMaker pipeline for
model training.
SageMaker Pipeline comes with about 300 built-in transforms, custom transforms using Python and PySpark,
built-in analysis tools such as scatterplots, histograms, and model analysis capabilities such as feature importance,
target leakage, and model explainability.
Data Cleaning Techniques

Your next step is to clean up the data. You look for missing values, outliers, and inconsistent formatting.
Dealing with Missing Values

Missing values, such as missing sales data for some days, can result in your algorithms producing inaccurate
results or failing. For missing values, either you can replace them with the imputation techniques, or you can
simply remove the record. If you’re using imputation techniques, you can fill the missing data with mean or
median values, use the regression method, or use multiple imputation by chained equations (MICE) techniques.
DATA ISSUE DESCRIPTION TECHNIQUES RETAIL CASE STUDY

USED
Missing Missing values in the data can Imputation Problem: A retail store finds missing
values lead to inaccurate results or sales data for some days in its sales
algorithm failure. tracking system.
These gaps in the dataset Regression Action: They decide to fill the missing
need to be dealt with method values to maintain the continuity and
effectively to maintain the reliability of the data.
integrity of the data analysis.
MICE Technique Used: They use a regression

model based on the historical pricing
or promotional events data to fill in the
missing sales data.
Mean/median Alternatively, they replace missing sales

Imputation price data with the average sales price to
impute the missing data.
Handling Outliers
In the case of outliers, you have to identify if there’s an error or an extreme value. If it is an error, you try to
resolve the error, and if it is an extreme value, you can decide to keep it but try to reduce the impact of it by using
techniques such as normalization. Data normalization helps keep the data on a similar scale. See the following
table for more details:
DATA ISSUE DESCRIPTION TECHNIQUES USED RETAIL CASE STUDY
Handling Outliers are Normalization Problem: A retail store notices that its sales
outliers data points quantity data can vary from hundreds to
that deviate thousands, while product price ranges from $1
significantly to $50.
from other
Min-max scaling Action: To avoid giving more importance to
observations.
sales quantity data due to its higher value, they
They can normalize the data to bring it to a similar scale.
be errors or
extreme values Standardization Technique Used: They apply min-max scaling,
and can lead where they transform the data to fit within a
to skewed or specified range, typically 0 to 1. This reduces
misleading the impact of outliers.
results in data Z-score They also apply standardization (Z-score
analysis. normalization normalization) to handle outliers as it’s less
sensitive to them. This process makes the mean
of the data 0 and standard deviation 1.
TIP Be sure to clean up data by handling missing values using imputation techniques like
mean or median values, regression, or MICE, and take care of outliers using normalization
techniques like min-max scaling or standardization.
Data Partitioning
It is important to split data into training, validate, and test datasets so that the models do not overfit, and you can
evaluate the models accurately. You need to ensure that there is no leakage of test datasets into training datasets.
One way to achieve this is to remove duplicates before splitting.
Data Scaling (Normalize, Standardize)

Consider the example of predicting the sales of various products in the store. Say you have collected a dataset
with the following features:
➤➤ Average customer footfall per day, ranging from 100 to 400
➤➤ Average income level, ranging from $50,000 to $200,000
➤➤ Average price of the product, ranging from $1 to $100
Normalization
To calculate the normalized value, you take each value, subtract the minimum value from it, and then divide it by
the range. So, for a footfall value of 200, your normalized value is 200 – 100/300 = 0.33.
You do the same for the income level and the price. By normalizing all the values, the vast difference in the abso-
lute values is minimized, and therefore your algorithm will give equal weightage to all these features.
FEATURE ORIGINAL RANGE NORMALIZATION NORMALIZED EXAMPLE

METHOD
Average customer 100–400 (value-min)/range For a customer footfall value of 200,

footfall/day normalized value is (200–100)/300
= 0.33.
Average income 50–200 (value-min)/range For an income of 100, normalized

level ($K) value is (100–50)/150 = 0.33.
Average price of 1–100 (value-min)/range For a product price of 50, normalized

product ($) value is (50–1)/99 = 0.49.
Standardization
If you didn’t scale these values, income levels would have an unfair advantage in influencing the predictions,
making your results inaccurate. Note that standardization is preferred for handling outliers, as it’s less sensi-
tive to them.
Unbias and Balance (Detection and Mitigation)

This section discusses detecting and mitigating bias in data. Say you are trying to predict your customers’ spend-
ing habits and using features such as age, income, purchasing habits, and location. When you train the model
and discover that it has different accuracies for different age groups, that is an example of bias. For example, this
might happen if most of your data belongs to middle-income, middle-aged customers, so your model may be very
good at predicting those segments. However, it might not be suitable for lower-income, younger groups, or high-
income, older groups.
Sometimes the bias can come from the algorithm itself. For example, if it has the predisposition to predict favora-
bly for larger group segments, and if you’re running a discount program using this model, it might give more
discounts to the middle-aged, middle-income groups.
You can mitigate bias by adopting data level, algorithm level, post-deployment level including continuous moni-
toring techniques, as shown in the following table:
BIAS MITIGATION DESCRIPTION EXAMPLE

TECHNIQUES
Data level You can build a balanced You could attempt to balance your
dataset by undersampling dataset by including more data from
the overrepresented groups lower-income, younger groups, or high-
and oversampling the income, older groups.
underrepresented groups.
Algorithm level 1. Try different algorithms. You can test different algorithms or
2. Change the parameters of the adjust the parameters of your chosen
chosen algorithm. algorithm to see if they perform better for
underrepresented groups. You could also
3. Use an ensemble method that
use ensemble methods to combine the
combines the predictions of
predictions of several models.
multiple models.
Post-processing level 1. Conduct a fairness audit of First you must check if the model is
the model. predicting accurately across multiple
2. Do a fairness correction to groups. Then, you can experiment by
adjust the model’s output until adjusting the model’s output to reduce
the bias is reduced. any identified bias.
Continuous monitoring 1. Monitor the model’s To begin with, regularly monitor the
performance across all groups. model’s performance across all customer
2. Make corrections to the model groups and adjust as necessary. This can
as needed. help ensure that the model continues to
make accurate predictions for all customer
groups.
NOTE You can detect and mitigate bias in machine learning models by using tools like Sage-
Maker Clarify, AWS Machine Learning Interpretability Toolkit, Azure Fairlearn, and GCP
Explainable AI.
Using SageMaker Clarify, you can monitor for bias at the data preparation stage before the training starts and
during each stage of the ML lifecycle. This can also detect bias after deployment and provides detailed reports
that quantify different types of bias and provide feature importance scores to explain how the model made its
predictions. It explains the data, model, and monitoring used to assess predictions.
Feature Engineering ❘ 259
Azure provides the Azure Machine Learning Interpretability Toolkit and Azure Fairlearn, and Google provides
GCP Explainable AI to detect and mitigate bias and provide explanations for predictions.
Data Augmentation
Another technique you can use is data augmentation. This involves creating new data from existing data. It is
often used in deep learning. See the following table for more details:
STEPS TECHNIQUES AND METHODS RETAIL CASE STUDY EXAMPLE
Creation of new data This involves creating new In a retail setting, you could create synthetic
data from the existing data. new customer reviews to augment your existing
dataset.
Adding noise to data Adding noise to data can help You could add noise to the sales data. This could
improve the robustness of the allow your model to handle a broader range of
model. data and possibly improve its performance.
Handling Inconsistent Formats

In the case of inconsistent formats, you can use either text processing or parsing techniques. For example, you can
rectify the incorrect date formats using dateUtil libraries or regular expressions.
FEATURE ENGINEERING
Each unique attribute of data is a feature. Feature engineering is about selecting the right attributes for the
model to become more accurate and generalizable for new data. You choose different features based on the
type of model.
Feature engineering consists of the components shown in Figure 15.7.
Feature Engineering
Feature selection
Feature transformation
Feature creation (encoding, binning)
Feature extraction (automated through deep learning)
FIGURE 15.7: Feature engineering components
Feature Types
This section explains the various feature types that you can create for your models. The following table lists these
feature types, their suitable models, and some examples:
FEATURE TYPES DESCRIPTION USEFUL MODELS CASE STUDY EXAMPLE
Time-series Time-series features capture ARIMA, Prophet The fluctuating sales

features temporal information such throughout the week
as day of the week, month, or the sales promotions
holiday season, or if it’s a during the holiday
weekend. season.
Ratio-based Ratio-based features Decision trees (XGBoost Comparison of the price

features represent relationships and random forests), of an item among its
between different numerical regression models competitors.
variables. For instance, the
relationship between the
current price and the average
price.
Categorical Categorical features Classification models Promotion type can be a

features represent categories or such as logistic key feature, for instance,
discrete values such as the regression, support “Buy one get one”
type of promotion. vector machines versus “20% off.”
Binary features Binary features take on two Various models Whether or not a sales
possible values like yes/no promotion was running.
or 1/0.
Sentiment These features are derived Deep learning models Sentiment scores or
scores or topic from text data like customer such as RNNs, topics obtained from
features reviews. Sentiment scores or transformers the analysis of customer
topics extracted using NLP reviews.
techniques.
Interaction Interaction features represent Useful for any model For instance, a feature
features dependencies between that might miss these representing the effect
other features, such as the relationships of a promotion on a
interaction between price high-priced item.
and promotions.
TIP Use domain knowledge to create features that are relevant to the problem you are try-
ing to solve.
Feature Selection
Feature selection is the process of selecting the attributes that are more relevant to predicting the results. Con-
sider the use case of predicting a customer’s spending habits. Assume the dataset contains features such as the
customer’s age, income, location, purchase habits, number of children, job title, education level, and frequency of
store visits.
When selecting the features to predict the customer’s spending habits, you have to decide which feature is going
to be more relevant to predict this spending habit, and you have to decide which one is irrelevant, a duplicate, or
even counterproductive.
Feature Importance Scores

One method used to determine the relevancy of the features to the target is to train the model and examine the
feature importance scores. Algorithms like decision trees and their ensembles, such as random forest and gradient
boosting, natively provide these feature importance scores. By training these models, you may find that occupa-
tion titles and education levels do not impact spending habits as much as income levels, purchasing history, and
store visits do.
Correlation Matrix
Another metric that can help you select features is the correlation matrix, a table showing the relationship
between the target variable and the feature. A high value implies that it plays a greater role in predicting the
target, in this case, the spending habits. You can also use this correlation metric to eliminate features that have a
high correlation. For example, suppose income and occupation title are highly correlated. In that case, you may
retain the income and remove the occupation title from the dataset, because the income may correlate more to the
target variable, namely, the spending habit.
The advantage of feature selection is that you can reduce the size of the datasets by removing redundant and irrel-
evant features. The smaller dataset will result in less overfitting of the model in addition to higher performance
and simplicity, which makes it easier to interpret.
NOTE Use feature selection techniques like examining feature importance scores and cor-
relation matrices to select relevant features, reduce overfitting, improve accuracy, and reduce
training time. This results in higher performance and simplicity, leading to better interpret-
ability.
Feature Extraction
Feature extraction involves techniques that combine the existing attributes to create new ones. This is done to
reduce the dimensionality of the dataset and retain the original information. Figure 15.8 shows some examples of
those techniques.
Principal Component Analysis

The principal component analysis (PCA) technique creates new features, which are linear combinations of the
original features. They try to capture as much of the variance in the original data as possible.
In the case study, this technique may combine products that are brought together, such as paints and paint
brushes, and combine them under a new category called Painting Essentials.
Independent Component Analysis

Unlike the principal component analysis, which tries to maximize the variance in the dataset, independent compo-
nent analysis (ICA) chooses attributes that are independent.
In the retail case study, this technique may choose attributes such as customers who buy gardening tools along
with those who buy baby products because they’re usually independent of each other.
NOTE Use feature extraction techniques like PCA, ICA, and linear discriminant analysis
(LDA) to create new features. This reduces the dimensionality of the dataset while retaining
the original information.
Independent
Component
Analysis (ICA)
Principal
Linear
Component
Discriminant
Analysis (PCA)
Analysis (LDA)
FEATURE
EXTRACTION
FIGURE 15.8: Feature extraction techniques
Linear Discriminant Analysis

Linear discriminant analysis uses a supervised learning technique using the class information to select a linear
combination of the features.
For example, if you choose to have different classes for high and low spenders, this technique may find the selec-
tion of products that separate the low spenders from the high spenders.
Feature Creation
Feature creation helps you add new relevant variables to your dataset to improve your model’s performance. Still,
you must be careful to ensure that it does not add noise or complexity to the model. The following table explains
the different feature creation techniques with examples:
FEATURE CREATION DESCRIPTION EXAMPLE

TECHNIQUES
One-hot encoding One-hot encoding involves You can convert the preferred shopping
converting the categorical variables day category into two binary categories:
into a form that can be fed into the the prefers_weekday or prefers_
model for better predictions. weekend feature. Each feature takes a
value of 0 or 1, depending on whether
the customer prefers to shop during the
weekend or weekday.
Binning Binning is a technique that combines You can create groups of ages such as
similar items to create numerical or 16 to 25, 26 to 35, 36 to 45, and so on. It
categorical bins. increases the accuracy of the prediction by
reducing the impact of minor differences
between ages.
FEATURE CREATION DESCRIPTION EXAMPLE

TECHNIQUES
Splitting Splitting involves creating two or You can split an address into different
more features from one feature. features, such as city, state, and street,
which can help you discover spending
patterns based on these new features.
Calculated features Calculated features are new features You can create a new feature, such as
that you can create by performing average spending per visit, by adding all
calculations on existing features. the spending over the year and dividing
it by the number of customer visits in that
year.
TIP Evaluate the performance of your model with different sets of features.
Feature Transformation
Feature transformation helps ensure the data is complete, clean, and ready for use in a model. The following table
explains the different feature transformation techniques available with examples:
FEATURE DESCRIPTION EXAMPLE

TRANSFORMATION
TECHNIQUE
Cartesian products of By calculating the cartesian Suppose you have two features: average
features products of two features, you spending per visit and total number of visits per
can create a new feature with month. In that case, multiplying these two will
greater meaning to the end give you the monthly spending, which may have
result. more relevance to solving your problem.
Nonlinear The nonlinear transformation A good example is binning numerical values into
transformations will help you understand the categories. For example, you can create a group
nonlinear effects of data. of categories using the age feature, such as age
groups between 16 to 25, 26 to 35, 36 to 45,
and so on. This categorization will help you assist
with the impact of age on spending habits from a
nonlinear perspective.
Domain-specific These are new features that If you know that some customers buy organic-
features you can create based on your only products, you know these are high spenders,
domain knowledge. and you can create a new feature named organic
buyers.
TIP Use a variety of transformation techniques, such as scaling, normalizing, and transform-
ing. Start with a small number of transformation techniques and add more as needed.
Feature Imputation
Feature imputation is the process of handling missing data using different imputation methods depending on the
feature type, as shown in Figure 15.9.
Predictive
Imputation
Simple
Multivariate
Imputation
Imputation
FEATURE
IMPUTATION
FIGURE 15.9: Feature imputation techniques
Simple Imputation
Simple imputation involves replacing the missing values with mean, median, or mode values. For example, if you
have a feature named “number of visits per month,” you can fill in the missing values with the median number of
visits for all customers per month.
Predictive Imputation
Predictive imputation involves predicting the missing values using other features. For example, you can predict
the missing value for income by training a model using other attributes such as occupation and education levels.
Multivariate Imputation
Multivariate imputation is used when there are relationships between features. Missing values are filled by run-
ning a model using other features in the round-robin fashion. You can use methods like MICE to fill in missing
values in different features by predicting them from other features in a round-robin fashion. This results in multi-
ple datasets, which can then be combined to account for uncertainty in the missing data.
For example, in a dataset that contains three variables such as Math score, English score, and Attendance, and if
the values for Math and English scores are missing, instead of imputing them separately, you can use the Attend-
ance record to predict the Math score and then use the predicted Math score and the Attendance record to predict
the English score.
NOTE Choose the right imputation method based on the type of missing data—such as
simple imputation, predictive imputation, or multivariate imputation—to reduce bias based
on available resources.
WORKBOOK TEMPLATE: DATA PROCESSING & FEATURE

ENGINEERING WORKFLOW
Download the “Data Processing & Feature Engineering” template from the download section of the book (www
.wiley.com/go/EnterpriseAIintheCloud). Use this template to put together a workflow chart for data
processing and feature engineering. You can use this as a roadmap for the various stages that the data goes
through from collection to being ready for model training.
SUMMARY
This chapter covered a wide range of topics related to cloud-based data processing and feature engineering and
essential steps in the machine learning lifecycle that are crucial to professionals in cloud engineering, data engi-
neering, and machine learning. The detailed insights into the data collection techniques, preprocessing methods,
and feature engineering combined with the hands-on exercises should give you the skills and knowledge to start
data processing. You also learned how cloud providers such as AWS, Microsoft, and Google play a vital role in
making data scalable, secure, and cost-effective.
Beyond the technology factor, you also learned how the human element comes into play for making decisions
around data partitioning, augmentation, balancing, and handling inconsistent values.
REVIEW QUESTIONS
1. Which of the following is true about data processing in the cloud?
A. Local servers are more scalable than cloud servers.
B. Data from multiple sources cannot be processed easily using cloud tools.
C. AWS Glue can be used for ETL and Amazon Athena for data query in S3.
D. Data stored in the cloud does not have rigorous security requirements.
2. Which of the following is an example of structured data?
A. Customer reviews
B. Economic indicators in XML or JSON files
C. Sales trends or seasonality
D. None of the above
3. Which of the following is true regarding the data processing workflow?
A. Data processing workflow includes only data collection and does not include data preparation.
B. Labeled data indicates information that has no target.
C. Exploratory data analysis is part of data preparation and can be used to understand data.
D. Data cannot be collected from sources such as sensors, time-series data, event data, or IoT data.
4. Which of the following is true about real-time and batch data ingestion?
A. Real-time data ingestion cannot handle high-velocity data.
B. Apache Kafka and AWS Kinesis are used for real-time data ingestion.
C. Batch data ingestion is the preferred method to keep data current.
D. The choice between the real-time and batch ingestion methods does not depend on the velocity
and volume of data.
5. Select the technique used to handle missing values.

A. Normalization
B. Standardization
C. Regression
D. Outlier removal
6. Which technique minimizes the difference in absolute values of features in a dataset?
A. Regression
B. Normalization
C. Outlier removal
D. Standardization
7. What is true about feature engineering?
A. It is about selecting random attributes from the dataset.
B. Ratio-based features are better suited to ARIMA models.
C. Sentiment scores from NLP are useful for RNNs and Transformers.
D. Interaction features identify the differences between variables.
8. Which of the following is true about feature selection?
A. Feature selection involves choosing all the variables for maximum accuracy.
B. Feature selection involves selecting the most appropriate variables.
C. Feature selection has no impact on model accuracy.
D. Feature selection depends on feature importance scores and less on correlation matrices.
9. Which of the following is NOT a feature extraction technique commonly used in machine learning?
A. Component analysis (PCA)
B. Independent component analysis (ICA)
C. Linear discriminant analysis (LDA)
D. Random component analysis (RCA)
10. Which of the following techniques is NOT related to feature creation in a machine learning model?
A. One-hot encoding
B. Binning
C. Splitting
D. Principal component analysis
11. Which of the following is NOT a feature transformation technique?
A. Cartesian products of features
B. Nonlinear transformations
C. Domain-specific features
D. One hot encoding
12. What does the acronym ETL stand for in the context of data pipelines?
A. Export, transform, load
B. Extract, test, load
C. Extract, transform, load
D. Export, test, load
Answer Key ❘ 267
13. Which of the following statements is true about data warehouses and data lakes?
A. Data warehouses hold highly structured data, while data lakes hold structured and unstruc-
tured data.
B. Data warehouses are like a farmer’s market, while data lakes are like well-organized gro-
cery stores.
C. Data warehouses provide scalability and flexibility, while data lakes allow for higher performance
and easy data analysis.
D. Data warehouses and data lakes are interchangeable terms for storing big data.
14. What is the main purpose of a feature store in MLOps?
A. To maintain a history of model performances
B. To provide a platform for hosting machine learning models
C. To create, share, and manage features during machine learning development
D. To orchestrate and schedule data pipeline tasks
15. What is the main purpose of data version control in machine learning?
A. To store multiple versions of data, track the models with which they were used, and reproduce
experiments
B. To keep track of all the models in production and monitor their performance
C. To orchestrate and schedule data pipeline tasks
D. To provide a platform for hosting machine learning models
16. What is the purpose of data lineage tracking in machine learning?
A. To create versions of your data and model artifacts to protect from deletion
B. To track the changes the data has undergone throughout the machine learning lifecycle and iden-
tify any unintended changes
C. To enforce security by developing a plan to reduce the exposure of data and the sprawl of data
D. To protect data privacy and ensure compliance with data protection regulations
ANSWER KEY
1. C 7. C 13. A
2. C 8. B 14. C
3. C 9. D 15. A
4. B 10. D 16. B
5. C 11. D
6. B 12. C
Choosing Your AI/ML
16
Algorithms
The only source of knowledge is experience.
—Albert Einstein
Data, as pivotal as it is, still requires algorithms to become actionable and give life to your AI initiatives.
Choosing the right AI/ML algorithm is like choosing the right tool for a job. This chapter covers different
machine learning algorithms and explores aspects such as how they work, when to use them, and what use
cases they can be employed for (see Figure 16.1).
02
Plan and
01 Strategize
and Prepare Launch Pilot
03
Continue Build and
Your AI Govern Your
journey Team
09
ENTERPRISE AI
Build 04
Evolve and
Mature AI
and Manage
Operations
Process Data and Modeling
08 Scale and Process Data
• Process Data and Engineer Features in the Cloud
AI
Deploy and 05 • Choose your AI/ML Algorithms
Monitor
• Train, Tune, and Evaluate Models
Models
07
06
FIGURE 16.1: PROCESS DATA AND MODELING: Choose your AI/ML algorithm
Back to the Basics: What Is Artificial Intelligence? ❘ 269
Note that the choice of a machine learning algorithm is not a straightforward process, and the relationship
between the algorithms and models is many to many; the same use case can be served by more than one algo-
rithm, and the same algorithm can solve many use cases and business problems.
This chapter explores different categories of machine learning, such as supervised learning, unsupervised learn-
ing, and deep learning. You’ll discover a plethora of algorithms such as linear regression, decision trees, neural
networks, and more.
Along the way, I discuss factors to consider when choosing an algorithm, ensuring that you have a robust meth-
odology to align your choice with the problem at hand. By the end of this chapter, you’ll have a firm grasp on
how to pick the right algorithm.
BACK TO THE BASICS: WHAT IS ARTIFICIAL INTELLIGENCE?

Artificial intelligence is a broad, umbrella term that refers to the collection of technologies enabling computers to
mimic human intelligence. The dictionary definition of intelligence is “the ability to acquire knowledge and skills
and apply them.”
Artificial intelligence refers to a machine’s ability to perform the human mind’s cognitive functions, such as per-
ceiving, reasoning, learning, interacting, solving problems, and even showing creativity. Anything that a machine
does, for example, listening to someone and responding in human words, such as Alexa, is artificial intelligence.
Artificial intelligence is also when a machine looks at the surroundings using computer vision and helps a car
navigate. The technologies involve the following:
➤➤ Radars and sensors
➤➤ Devices that take in all data and feed it to the car so that it can take appropriate action
Artificial intelligence includes those collections of technologies, such as natural language processing, speech rec-
ognition, and computer vision. Figure 16.2 shows the range of technologies and applications that fall under the
broad umbrella of artificial intelligence.
Here’s Gartner’s definition of AI:
Artificial intelligence (AI) applies advanced analysis and logic-based techniques, including
machine learning, to interpret events, support and automate decisions, and take action.
(Source: www.gartner.com/en/topics/artificial-intelligence)
Machine Learning: The Brain Behind Artificial Intelligence

The next question is, then, what is machine learning? Machine learning is often used interchangeably with artifi-
cial intelligence. An excellent way to answer that question is to take the driverless car example. Again, the sensors,
radars, and other systems in the car take in all the data from the environment and move the car around. But, how
is the car moving autonomously?
There is a machine learning model that takes in the data and predicts the outcome. The model’s output tells the
car to go left, right, stop, move fast, and accelerate.
Machine learning is about crunching all the data, processing it, identifying the patterns, and, based on that,
returning a prediction.
Artificial intelligence is an umbrella term for all the technologies that try to mimic human intelligence.
But the actual process by which large amounts of data are crunched to come up with insights is through
machine learning.
270 ❘ CHAPTER 16 Choosing Your AI/ML Algorithms
Supervised Learning e.g., Spam

Detection, Credit Scoring,
Image Classification
Text Generation: e.g., OpenAI’s GPT-
4 writing essays or poetry Reinforcement Learning e.g.,
Game Playing, Robot
Image Creation: e.g., DeepArt’s Navigation
algorithm to create new artistic
Machine
images Learning Deep Learning e.g., Facial
Generative Recognition, Image
Music Composition: e.g., OpenAI’s AI Colorization
MuseNet composing new music
pieces Unsupervised Learning e.g.,
Market Segmentation,
Deepfakes: e.g., Creating synthetic Anomaly Detection
media that look like someone else’s
likeness
Speech Recognition: e.g.,

Alexa
Medical Diagnosis: e.g., IBM’s Watson for
oncology Text Analysis: e.g., Grammarly
Natural
Financial Analysis: e.g., Robo-advisors Expert Artificial Language Sentiment Analysis: Social
for personal finance Systems media monitoring
Intelligence Processing
Fraud Detection: e.g., Credit card Machine Translation: e.g.,
transaction monitoring systems Google Translate
Chatbots: e.g., Cleverbot
Industrial Robots: e.g., Robots on the

assembly line
Image Recognition: e.g.,
Personal Service Robots: e.g., Facebook’s auto-tagging
Roomba vacuum cleaner feature
Robotics
Surgical Robots: e.g., Da Vinci Object Detection: e.g., Amazon
Surgical System Go’s cashier-less shopping
Computer system
Autonomous Vehicles: e.g., Drones
delivering packages Vision Optical Character Recognition:
e.g., Adobe’s PDF to text
conversion
Autonomous Vehicles: e.g.,

Tesla’s self-driving cars
FIGURE 16.2: Umbrella of AI technologies
Features and Weights of Predictive Algorithms

An algorithm has two key components—features and weights. Features are parts of a dataset identified as essen-
tial to determine an outcome. The features may also be known as columns or attributes in a dataset.
You can think of features as essential attributes to predicting an outcome. For example, if you are trying to pre-
dict the price of a home, the features could include the number of bedrooms, square footage, location, and so on.
However, you also need to know the context of the attribute impacting the outcome, in other words, how
important that attribute is to predict an outcome. That’s where the weights come in. These weights are assigned
as part of training, and if, say, it is highly influential to the outcome, the system might assign a higher weight, say
0.8. The training process uses the dataset to assign the weights to different factors to predict the outcome with a
certain level of accuracy.
FACTORS TO CONSIDER WHEN CHOOSING A MACHINE LEARNING

ALGORITHM
This section walks through an example scenario to understand the steps in choosing a machine learning algo-
rithm. Suppose your company has asked you to use machine learning to increase sales. You have a large amount
of data such as customer purchase history, demographic information, website behavior information, and so on.
Factors to Consider When Choosing a Machine Learning Algorithm ❘ 271
FACTOR DESCRIPTION
Identify the type The first thing to do is identify the problem you need to solve. In this case, you’re
of problem. trying to increase sales, which means you’re trying to identify customers who are
likely to purchase from your site. This is the classification problem because you’re
trying to identify who will buy and who will not.
Assess the size of The next step is to assess the size of the dataset. If you have a large dataset, using
the dataset. complex algorithms such as deep learning is possible. However, if the dataset is
smaller, you can use simpler algorithms such as linear regression.
Decide on the The algorithm choice also depends on whether the model needs to be interpretable,
interpretability of meaning is it required to explain how the model came up with its prediction? If
the model. you have to explain how the model predicts the outcome, you cannot use complex
algorithms such as deep learning.
Decide on the It is also vital to decide on the metrics you’ll use to evaluate the model’s
performance metrics. performance. In this case, you can use accuracy, precision, and recall, thereby
evaluating how well the model predicts which customer will likely make a purchase.
Plan for scalability. It would help if you also planned for the scalability of the model. In this case, given
that you have a large dataset, it may be a good idea to plan for a distributed
machine learning algorithm such as Apache Spark.
TIP When choosing a machine learning algorithm, carefully consider the problem type,
dataset size, interpretability requirements, appropriate performance metrics, and the need for
scalability to ensure an effective and efficient solution.
HANDS-ON EXERCISE: SELECTING A MACHINE LEARNING ALGORITHM
This is a strategic hands-on exercise focused on choosing the right machine learning algorithm
for your project’s use case. It is a critical step because choosing the right algorithm will decide
the accuracy of the model and therefore drives the outcome of your overall project.
It will help you to think critically about various phases of machine learning such as identifying
a problem, assessing data, interpreting models, measuring performance, and considering scal-
ability requirements. This exercise mirrors the thought process that goes into planning a data
analysis or modeling project.
Goal: Assume that you have been hired as a data scientist for an e-commerce company that
has trusted you with the responsibility of leveraging the customer data to gain a competitive
advantage to address the stiff competition the company has been facing. The marketing team
has told you that they would like to predict which customers are likely to purchase next month
so they can plan their marketing campaigns accordingly. They have given you a dataset that
contains the customers’ past purchases as well as other demographic information such as age,
location, browsing behavior, and so on. Now your task is to identify a suitable algorithm to
solve the problem.
continues
continued
STEP DESCRIPTION
Identify the type of problem. Decide what type of machine learning problem this
is; in this case, this a classification problem as this is
trying to identify the customers who are likely to buy
vs those who may not.
Assess the size of the dataset. If the dataset is very large, you may need a deep
learning algorithm, else a simpler algorithm would
suffice.
Check whether the model needs If the model needs to be interpretable, you will need
to be interpretable. a simpler model.
Choose appropriate Choose the appropriate performance metrics

performance metrics. based on the model; for example, in this case it’s a
classification problem, so you can choose accuracy,
precision, and recall.
Check out the scalability If the dataset is large and likely to grow, you
requirements. should explore machine learning algorithms that
are compatible with distributed processing such as
those offered by Apache Spark.
Write a report regarding the Prepare a report with the type of algorithm, how it
finalized algorithm. meets the requirements, as well as the challenges,
limitations with this approach, and other alternatives
if available.
DATA-DRIVEN PREDICTIONS USING MACHINE LEARNING

Machine learning is a branch of artificial intelligence that allows computers to learn from data without explicit
programming. They achieve this by identifying common patterns and then using probabilistic distributions to
make predictions. In other words, they develop the conditional probability of an output variable based on a set of
input variables.
You feed in a dataset containing input-output pairs to the model, and the model tries to discern patterns. Using
this insight, the model makes predictions on a new, unseen dataset of input variables. Machine learning adopts
the probabilistic reasoning approach, which aims to determine the probability of a particular outcome based
on the evidence available.
The data can be in the form of text numbers, images, or any data. Machine learning uses different types of
algorithms to make these predictions, for example, supervised learning, unsupervised learning, and so on. These
algorithms are based on statistical methods.
NOTE An essential characteristic of machine learning is that these models keep improving
over time because they learn from the existing data and can make better predictions over
time.
Data-Driven Predictions Using Machine Learning ❘ 273
Different Categories of Machine Learning

You can classify machine learning problems into three categories: supervised learning, unsupervised learning, and
reinforced learning. Figure 16.3 shows how different tasks can employ supervised or unsupervised learning and
how certain tasks can be implemented using supervised or unsupervised learning techniques.
UNSUPERVISED LEARNING SUPERVISED LEARNING
Image Generation Image Recognition
Video Generation Question and Answer

Image Classification
Language Modeling
Language Translation
Text Generation Sentiment Analysis
Speech Synthesis Speech Recognition
FIGURE 16.3: Use cases of supervised and unsupervised learning tasks
The next few sections discuss the different algorithms shown in Table 16.1.
TABLE 16.1: Supervised and Unsupervised Machine Learning Algorithms
SUPERVISED LEARNING UNSUPERVISED LEARNING
Linear regression Autoencoder

Logistic regression K-means clustering
Linear/quadratic discriminant analysis Gaussian mixture model
Decision tree Hierarchical clustering
Naïve Bayes Recommender system
Support vector machine Manifold learning
Random forest
AdaBoost
Gradient boosting trees
Simple neural network
Note that the main point of this discussion of different applications is for you to be able to understand these dif-
ferent types and then go back to your organization and apply these to your different problems.
Using Supervised Learning

This section discusses supervised learning, including when to use it, how it works, and typical use cases.
Supervised learning is a type of machine learning where the model is trained on labeled data. Labeled data implies
that every training data sample should have a predefined output. It is also known as ground truth labels.
One good example is the case of email spam applications, whereby you have a set of training data in which you
know which emails are spam and which ones are not. For each spam email, you also know the corresponding
features that identify that email as spam. Machine learning uses an algorithm that trains itself based on this data
to create a model. The model uses these features to predict the outcome, which in this case, is to predict whether
an email is spam.
TIP When using supervised learning, ensure you have a large, labeled, well-formatted, and
clean dataset for accurate training and prediction.
Let’s discuss the supervised learning process shown in Figure 16.4.
A human labels
the input data
Labeling
01
02 The algorithm is
Training then trained on
the labeled input
Prediction data
The algorithm is
03 applied to new
data to predict
the outcome
FIGURE 16.4: How supervised learning works
➤➤ Labeling: A human labels the input data by identifying the input and output variables. In the case of
identifying spam, they would identify the input fields such as subject, to, and from fields and define the
output variable as spam or not spam.
➤➤ Training: The algorithm is then trained on the labeled input data to understand the patterns between the
input variables and the output.
➤➤ Prediction: Once the training is complete and when you determine that the algorithm is pretty accurate,
the algorithm is applied to new data to predict the outcome. See Figure 16.5.
Classification
Within supervised learning, you have two types of algorithms, classification and regression (see Figure 16.6). In
the case of classification, you are trying to predict whether the outcome belongs to a particular type of category.
In contrast, you’re trying to predict a continuous value in the case of regression.
System learns by identifying patterns in the data previously labeled by humans
(a)
Model
(b) (c) (d)

New data Prediction
(e)
Pictures of animals
labeled by humans
FIGURE 16.5: How supervised learning works

Source: (a) Kate / Adobe Systems Incorporated, (b) anankkml / Adobe Systems Incorporated, (c) Quality Stock
Arts / Adobe Systems Incorporated, (d) anankkml / Adobe Systems Incorporated, (e) khumthong / Adobe Systems
Incorporated
Binary Classification
Multiclass Classification
Multilabel Classification
Classification Logistic Regression
Imbalanced Classification
Support Vector Machines
KNN
Supervised Learning
Linear Regression
Multiple Linear Regression
Regression Polynomial Regression
Ridge Regression
Lasso Regression
FIGURE 16.6: Types of supervised learning

Within classification are binary classification or multiclass classification:

➤➤ Binary classification: The outcome can be one of two values, such as whether the email is spam.
➤➤ Multiclass classification: You can have more than two outcomes, such as when you’re trying to classify a
set of animal images. For example, it could be a cat, dog, horse, and so on.
Another example of a classification problem involves an Amazon fulfillment center. There are millions of boxes.
They have hundreds of items in each box, and some items may be missing or damaged. You can use machine learn-
ing to identify which boxes are damaged and then send warehouse representatives to address the issue. Note that
classification can also be further subdivided into multilabel classification, imbalanced classification, ensemble meth-
ods, support vector machines, KNN, and so on. Logistic regression, SVM, and decision tress are types of binary clas-
sification, and neural networks, random forests, and naïve Bayes multivariate are types of multiclass classification.
Regression Problem
This section looks at the example of a regression problem that focuses on predicting a continuous value. One
such example is forecasting the demand for a product based on past historical time-series data. In this case, the
output of the regression problem is the number of products the customers will buy. You can even use regression to
predict a product’s price based on its historical data and other factors. Regression can be further subdivided into
linear regression, multiple linear regression, polynomial regression, ridge regression, lasso regression, and so on.
NOTE Regression is different from classification because it is focused on predicting contin-

uous numerical values based on the input variables. Classification is focused on categorizing
data into specific classes or categories based on input variables.
Types of Supervised Learning Algorithms

This section looks at the following ML algorithms that are available in the category of supervised learning.
Predicting the Value Using Linear Regression

Linear regression is a statistical method that can identify the relationship between one or more independent vari-
ables and a dependent variable. There can be an infinite number of values for the dependent output variable.
Say you want to predict a car’s sales price based on age. First, you need to collect the data for the last few months
or years, containing the sales price and the car’s age. Typically, you use a software algorithm to carry out the
analysis, but for the sake of simplicity, say you are using an Excel sheet for this analysis. You first start by captur-
ing this data on a scatter plot by plotting the car’s age on the x-axis and the price of the car on the y-axis.
You then fit a line so that the distance between the line and the data points is minimal, as shown in Figure 16.7.
This line should minimize the sum of the squared distances between the actual and predicted prices denoted by
the line. This line is the regression line and represents the best-fit line through the data. Using this line, you can
now predict the price of the car based on the age, by using the slope and intercept on the line.
Table 16.2 shows a list of potential use cases for linear regression algorithms.
TABLE 16.2: Use Cases for Linear Regression Algorithms
USE CASE DESCRIPTION INDUSTRY DATA TYPE PREDICTIVE

OUTPUT
Sales Assist with demand Retail, Sales data Sales forecast

forecasting planning, inventory e-commerce Marketing
management, and data
resource allocation

OUTPUT
Customer Improve customer Retail Customer Customer

analytics retention and increase Finance data lifetime value
sales Behavior Churn
Insurance
data purchase
Telecommunications
propensity
Financial For data-driven Finance Market data Stock prices

analysis investment decisions, Investment Financial Asset values
market analysis, and risk data
management
Demand To optimize inventory Retail Sales data Demand

forecasting management and supply Supply chain Market data forecast
chain operations
External
factors
Performance For process optimization, Manufacturing Supply Production Process

optimization resource allocation, and chain Logistics data optimization
cost reduction Supply data
Price For maximizing Retail Pricing data Pricing strategy

optimization revenues and market Manufacturing Sales data
competitiveness
Risk Make better lending Insurance Risk data Defaulting on

assessment decisions and reduce the Finance Historical a loan
risk of financial losses data
Fraud Prevent fraud and protect Financial services Transaction Fraud

detection losses Retail data
Energy Energy
To implement energy Energy Energy data
consumption consumption
saving measures, optimize Sustainability Building
analysis analysis
resources data
Other factors can impact the price of the car, such as the car’s condition, make and model, local market condi-
tions, competition, advertisement, and so on. Although this is a trivial example, it serves to help you understand
how linear regression works. Note that you will be using linear regression algorithms provided by different
frameworks in the real world.
Linear regression is a commonly used tool in various fields and industries, including finance, manufacturing,
statistics, economics, psychology, and so on. It can be used to optimize price points and estimate product-price
elasticities.
You can use linear regression in the following situations:
USE CASE DESCRIPTION SUITABLE ALTERNATIVE(S)
Linear relationship Linear regression is most effective when there is a

linear relationship between the input and output
variables.
No multicollinearity Linear regression is appropriate when there is no Ridge regression, lasso

multicollinearity (linear relationship) between the regression
independent variables. Multicollinearity can result
in unstable results.
Small number of Linear regression is a good choice when Support vector

independent variables the number of independent variables decomposition, random
involved is small. forest, gradient boosting
Simplicity of Linear regression may not be suited for capturing Neural networks or other
relationships complex relationships between the independent complex algorithms
variables.
15
10
–20 –10 10 20 30 40 50 60
FIGURE 16.7: Example of linear regression with one independent variable
TIP When using linear regression, ensure a linear relationship between variables, avoid mul-
ticollinearity, and limit the number of independent variables. For complex relationships, con-
sider advanced models such as neural networks.
Predicting Categorical Outcomes Using Logistic Regression

Logistic regression is helpful when you want to predict the outcome of an event; in other words, the outcome
variable is categorical. The value of the output variable is binary (e.g., yes or no, black or white) rather than con-
tinuous as in the case of linear regression.
Categorical values represent different categories or groups and are represented by labels. They can be further
divided into nominal or ordinal types.
VARIABLE TYPE DESCRIPTION INHERENT EXAMPLES

ORDER
Nominal Represent different No Gender (male, female), color

categories or groups and are (red, green, blue)
represented by labels. These
categories do not have any
inherent numerical meaning
and are mutually exclusive.
Ordinal Have a specific order Yes Education levels (high school,

or ranking. bachelor’s degree, master’s
degree), customer satisfaction
levels (poor, fair, good, excellent)
Consider an example where you want to predict whether a customer will churn. In this case, you collect all the
data that will contribute to this event, such as how many products they bought, how often they bought, how
much they spent, and how often they canceled their subscription. You should be able to get all this information
from a customer loyalty program.
Once you train the logistic regression model using this data, this software will draw a logistic regression line fit-
ting this data, and you can use this line to understand the relationship between the various independent variables
and their probability of whether the customer will churn, as shown in Figure 16.8.
1.00
Probability of Customer Churn
0.75
0.50
0.25
0.00
1 2 3 4 5
Number of Days of Inactive Usage
FIGURE 16.8: Example logistic regression graph showing the probability of customers churning based on the
number of days not using the service
This algorithm is useful in many use cases in healthcare, finance, and marketing, as shown in Table 16.3. Some of
these use cases are whether a customer will default on a loan, whether a prospect will buy a car, and whether a
customer will churn.
TABLE 16.3: Use Cases for Logistic Regression

OUTPUT
Customer churn Reduce customer Retail Customer Customer churn

prediction churn and increase Finance data probability
retention Behavior
Insurance
data
Telecommunications
Loan default Improve lending Financial services Customer Default probability
prediction decisions and prevent data Creditworthiness
losses Behavior
data
Fraud detection Prevent fraud and Retail Transaction Fraud probability

avoid financial losses Finance data
Insurance
Telecommunications
Email spam Reduce spam and Technology Email data Spam/not spam
filtering increase email Email Metadata
deliverability
Security
Medical diagnosis Improve patient care Healthcare Medical Disease prediction

and diagnosis data
Product Improve product Retail Customer Propensity to buy

recommendation recommendation and data a product
increase sales
Sentiment analysis Make data driven Social media Text data Sentiment label
decisions Marketing
Image Image label

Object and facial Computer vision Image data
classification
recognition apps
You can use logistic regression in the following situations:
USE CASE REQUIREMENT DESCRIPTION
Classification problems When the outcome variable is binary, the outcome can be yes or no, true or
false, and so on. In other words, it’s a classification problem.
Large datasets When the dataset is small, the model is likely to overfit. It is ideal for large
datasets, as it can lend itself to computational efficiency and more statistical
power to make more accurate predictions.
USE CASE REQUIREMENT DESCRIPTION
Interpretability When you want the model to be interpretable, meaning you want to understand
how the outcome is predicted based on the inputs.
Efficiency When you want an efficient algorithm that can perform fast with fewer
computational resources.
TIP Overfitting means the model has almost memorized the outcomes and cannot be gen-
eralized for new data. It performs poorly on new data even though it works perfectly on
training data.
Using Decision Trees

Decision trees are machine-learning algorithms that can be used for classification and regression. In other words,
it is a supervised learning type of algorithm. They work by breaking down the dataset recursively into smaller
subsets based on specific criteria until you reach a point where the outcome can be determined.
Say you want to identify the customer who is likely to default on a loan. You collect the dataset of past loan
applicants and their information, such as their income, loan amount, credit rating, age, employment, and so on.
Then you start by identifying the feature on which to break this dataset. That depends on which feature gives you
the highest information gain. You may break the dataset based on the credit score rating in this case.
You then analyze if the customer will default based on their credit rating. If you find that to be true, you create a
leaf node that predicts that the customer will default when their credit rating is below that level.
On the other hand, if there is no clear pattern, you choose a different feature, such as their income. You continue
with this process until you reach a point where all the data points under a specific subset belong to a particular
class, meaning either they have or have not defaulted. Or they stop at some predetermined point, such as a spe-
cific tree depth or any other criteria you agree on.
You can then use the prediction on new data once you finalize the decision tree, as shown in Figure 16.9.
The benefits of the decision tree are that it can operate on numerical and categorical data, handle outliers well,
and is also visual, making it easy for people to understand. On the other hand, it is susceptible to bias and inac-
curacies if it’s not developed properly. Therefore, alternative approaches should be explored if necessary.
Table 16.4 shows various use cases for decision trees.
TABLE 16.4: Use Cases for Decision Tree Algorithms
USE CASE BENEFITS INDUSTRY DATA TYPE PREDICTIVE

OUTPUT
Customer Personalized marketing Retail Customer data Customer

segmentation campaigns and Finance Behavior data segments
strategies, improved
Insurance
customer satisfaction
Telecommunications
continues
(continued)
USE CASE BENEFITS INDUSTRY DATA TYPE PREDICTIVE

OUTPUT
Fraud Identifying and Finance Transaction Fraudulent/

detection preventing fraudulent Insurance data nonfraudulent
activities, minimizing User data
financial losses,
enhancing security
Credit risk Informed lending Banking Credit data Credit risk

assessment decisions, assessing Finance Applicant data assessment
creditworthiness,
minimizing default risk
Fault diagnosis Identifying system Manufacturing Sensor data, Fault diagnosis

faults, troubleshooting IoT system
guidance, reducing parameters
downtime
Customer Efficient customer Customer service Customer Service routing

service routing support, improved queries,
response times, service data
enhanced customer
satisfaction
Demand Optimized inventory Retail Sales data, Demand

forecasting management, production Supply chain market data forecast
planning, and supply
chain operations
Medical Enhanced decision Healthcare Patient Potential

diagnosis support for healthcare symptoms, medical
professionals, improved medical diagnoses
accuracy in diagnoses records
Equipment Equipment
Proactive maintenance Manufacturing Sensor data,
maintenance maintenance
scheduling, reduced IoT maintenance
needs
downtime, optimized records
equipment performance
This section covers the situations when decision trees can be adopted.
CONDITION DESCRIPTION EXAMPLE
Interpretability Decision trees provide interpretability, Understanding the decision-making

which means you can explain how the process of the model
model predicts based on given inputs.
Both numerical and Decision trees can handle both Dataset that contains both numerical
categorical data numerical and categorical data types. values like age, income, and categorical
values like gender, nationality
Handling missing data Decision trees can handle missing data Imputing missing data based on the
by assigning probabilities based on the relationships in available data
available data.
Large datasets Decision trees are efficient for Processing datasets with high volume
large datasets.
Ensemble models Decision trees can be combined Building a random forest or gradient
with other algorithms to build boosting model
ensemble models.
Better performance Decision trees can be tuned to develop Improving the prediction accuracy or
high performing models. reducing error rates
Survival of passengers on the Titanic
gender
male female
age survived
0.73; 36%
9.5 < age age <= 9.5
died
siblings
0.17; 61%
3 <= siblings siblings < 3
died survived
0.02; 2% 0.89; 2%
FIGURE 16.9: Decision tree that shows the survival of passengers on the Titanic. Figures under the leaves
show the probability of survival and the percentage of observations in the leaf.
Source: Wikipedia
TIP Consider decision trees for large, mixed-type datasets that need to be well-understood
by stakeholders and for handling missing data and use ensemble models for enhanced
performance.
HANDS-ON EXERCISE: BUILDING A DECISION TREE
The goal of this exercise is to build a decision tree model to predict whether a loan applicant
will default.
Prerequisites: This is a high-level overview, but to actually execute this exercise, you should be
familiar with Python, pandas, and the scikit-learn library.
Access to a loan dataset that contains data such as applicant name, age, income, and so on,
and whether they defaulted.
STEP NAME TASK OWNER INSTRUCTIONS DELIVERABLES
Dataset Data analyst/ Load the loan dataset into A cleaned and
preparation scientist a pandas dataframe. Clean formatted dataset
and format the data as
needed.
Exploratory Data analyst/ Examine the dataset to Data insights

data analysis scientist understand the distribution
(EDA) of variables and evaluate the
ratio of defaulters to non-
defaulters.
Preprocessing Data analyst/ Encode categorical variables, A preprocessed

scientist handle missing values, scale dataset ready for
numerical variables, etc. model training
Splitting the Data analyst/ Divide the dataset into a A training set and a
data scientist training set and a testing set. testing set
Model Data analyst/ Choose a decision tree Selected model

selection scientist algorithm for the task. algorithm
Training the Data analyst/ Train the selected model A trained decision
model scientist using the training dataset. tree model
Model Data analyst/ Evaluate the model using the Evaluation results
evaluation scientist test dataset. Use relevant
metrics such as accuracy,
precision, recall, F1 score, or
AUC-ROC.
Model Data analyst/ Visualize the decision tree Model visualization

interpretation scientist and interpret its decision and interpretation
rules.
Model tuning Data analyst/ If the model performance A fine-tuned model

scientist is not satisfactory, consider with improved
tuning the parameters or try performance
using a more sophisticated
algorithm.
Using Random Forests for Classification and Regression

Random forests are helpful for both classification and regression types of problems. You can think of a random
forest as an improvisation of a decision tree because it is a collection of multiple decision trees built based on
random subsets of data and features. Figure 16.10 shows a simplified view of a random forest.
Instance
Random Forest
Tree-1 Tree-2 Tree-n
Class-A Class-B Class-B
Majority-Voting
Final-Class
FIGURE 16.10: A simplified view of a random forest

Source: https://en.wikipedia.org/wiki/Random_forest#/media/File:Random_forest_diagram_
complete.png
The methodology and benefits of using Random Forest are outlined in the following table:
STAGE DESCRIPTION IMPORTANCE
Training You break down your dataset into a training By doing this, you can eliminate some of
and test dataset and then divide the dataset the drawbacks of your decision tree, such
to build multiple decision trees, each using a as overfitting and avoiding bias. The final
random subset of data and features. decision is arrived at by combining the
outcomes of the individual trees.
Testing To evaluate the model’s performance, you Helps evaluate the performance of the
then test the model against the test dataset model and improve its accuracy.
and compare the predictions to the actual
outcomes using metrics, such as accuracy,
precision, recall, and F1 score.
Accuracy When you need accuracy in the predictions, Essential for making reliable predictions
this method is good because you are and minimizing errors.
using multiple decision trees, which is
more accurate.
continues
(continued)
STAGE DESCRIPTION IMPORTANCE
Robustness When you need robustness in the model Enhances the stability of the model
predictions, this method is good because it and its ability to handle a variety of
uses multiple decision trees and issues such data inputs.
as overfitting are avoided.
Interpretability When interpretability is required, even Important for understanding how

though these models are not as interpretable the model works and explaining its
as linear regression models, they still provide predictions.
visibility into how they came up with the
predictions.
Random forests are especially suited for customer churn prediction, fraud detection, image classification, recom-
mendation systems, predictive maintenance, demand forecasting, and credit risk assessment.
TIP Random forests are helpful for both classification and regression types of problems
because they use multiple decision trees to enhance the stability of the model and its ability
to handle a variety of data inputs.
Using Naïve Bayes

Naïve Bayes is a machine learning algorithm used for classification purposes. It uses the Bayes theorem to calcu-
late the probability of an outcome based on given evidence. The following table illustrates the steps involved in
implementing a naïve Bayes classifier:
STEP DESCRIPTION PURPOSE
Building the You’re building a classifier that identifies Helps in categorizing data based on
classifier different types of fruits, and you have independent features. Note: This is called
several features such as color, texture, size, a naïve theorem because in the real world
and shape. features are usually not independent.
Creating a Calculate the probability of the different The base model serves as the initial
base model labels based on the training dataset. representation of the problem, which can
then be improved through further training
and adjustments.
Prediction You then use the new data against this Applying the model to new data allows you
model to make predictions. to predict outcomes based on the learned
relationships from the training data.
The Naïve Bayes theorem is a powerful tool that can be used to identify spam, classify text and images, categorize
documents, analyze customer reviews, classify news articles, detect fraud, diagnose medical issues, perform senti-
ment analysis, and perform facial recognition.
The following table captures the conditions under which the Naïve Bayes theorem is recommended:
Text data Naïve Bayes works well when the Useful for sentiment analysis, topic
data is text type. classification, and spam detection
Categorical data Naïve Bayes works well when the Useful when dealing with colors, demographic
data is categorical. information, product classification, etc.
High Performs well on large datasets with Applicable in scenarios with many input
dimensionality many features. variables or features
Real-time Naïve Bayes is a simple and fast Beneficial for applications requiring real-time
inference algorithm, suitable for real-time predictions or decisions
applications.
Limited Naïve Bayes works well with limited Useful in situations where there’s limited
labeled data labeling in the dataset. labeled data for training
TIP Consider using Naïve Bayes for classification tasks involving text or categorical data,
especially in high-dimensional spaces. Its simplicity and efficiency make it a strong candidate
for real-time prediction tasks.
Using Support Vector Machines

Support vector machines (SVMs) can be used
for classification and regression analysis. They H2 H3
X2 H1
create a hyperplane separating the two out-
comes with the most significant margin possible.
Say you want to identify spam based on the
length of the emails and the number of punctua-
tions used in an email. You then plot the email
length on the x-axis and the punctuation on
the y-axis. SVM would then find the line that
separates the spam versus nonspam classifica-
tion types in the two-dimensional space. You
can then use this model to predict the outcomes,
as shown in Figure 16.11. This figure shows a
support vector machine that chooses a separat-
ing hyperplane for two classes of points in 2D.
H1 does not separate the classes. H2 does, but
only by a small margin. H3 separates them with
the maximum margin.
In the cases where the features are not linearly
separable, the SVM algorithm chooses a kernel X1
to transform the data into a higher dimensional
space where it is linearly separable. For example, FIGURE 16.11: Support vector machine trained from
two samples
the model would add a third feature, such as the frequency of certain words in the email, and transform the
two-dimensional data into three dimensions.
SVM can also choose several other types of kernel functions, such as linear, polynomial, and radial base functions.
The following table captures the conditions under which support vector machines can be adopted:
Linear and Support vector machines are helpful for classification Suitable for various
nonlinear data use cases, but more important, they are useful for classification problems,
both linear and nonlinear data due to the use of regardless of linearity
kernel functions.
High- SVMs can be helpful for high-dimensional data (many Useful when dealing with
dimensional data features) use cases with large amounts of data. many features or large
amounts of data
Limited data SVMs work well with limited data, avoiding overfitting Ideal for scenarios with
and improving generalization. limited training data
Complex Because SVMs can handle nonlinear use cases better, Suitable for complex or
classifications they are helpful for complex classification scenarios. intricate classification tasks
Outliers and SVMs are efficient at handling outliers and noise Good fit for datasets that may
noise handling in the data. contain noise or outliers
Table 16.5 lists some of the use cases suited to support vector machines.
TABLE 16.5: Use Cases for Support Vector Machines
USE CASE DESCRIPTION INDUSTRY DATA PREDICTIVE

TYPE OUTPUT
Image Identify objects in images, Retail Image Image labels

classification categorize images into predefined Healthcare data
classes, and computer vision apps
Manufacturing
Text Identify the topic of a document, Finance Text Document

classification sentiment of a text, detect spam Customer service data labels
Marketing
Anomaly Detection of outliers or unusual Cybersecurity Sensor Anomaly

detection patterns in data, identify potential IoT data labels
security breaches System
logs
SVMs can also be used for handwriting recognition, fraud detection, customer sentiment analysis, spam email
detection, medical diagnosis, and market trend prediction.
TIP Consider using SVMs for classification tasks involving both linear and nonlinear data.
They are especially useful when dealing with high-dimensional or limited datasets and
capable of handling complex classifications and outliers effectively.
Using Gradient Boosting

Gradient boosting is helpful for both classification and regression tasks. The name implies it’s focused on
iteratively improving a weak learner by adding new models to the ensemble and correcting the previous mod-
el’s errors.
Say you are trying to predict the price of houses using a certain number of features, such as the square foot area,
the number of bedrooms, and so on.
To build the tree ensemble, you would train a decision tree and then calculate the errors between actual and
predicted prices. Then, you would run another decision tree that addresses the errors in the previous model. The
new model will be added to the ensemble, and you will continue the process by training another decision tree to
correct the errors in the second decision tree.
You continue this process until you reach a point where you have high accuracy. The ensemble model would
perform better than any of the individual decision trees and would be less prone to overfitting. The final model
would be a weighted combination of the individual models based on their performance on the validation set.
Gradient boosting is a robust algorithm that provides highly accurate models for large datasets and is helpful in
click-through rate prediction, customer churn prediction, demand forecasting, credit risk assessment, financial
modeling, fraud detection, and image recognition use cases.
Here are the situations in which a gradient boosting model is a good candidate:
Large and Particularly useful for large and complex Useful when dealing with
complex datasets enterprise-level datasets. a complex dataset with
many variables
High accuracy Useful for high accuracy requirements. Ideal when high precision
is essential
Complex feature Can handle various input features, including Applicable in scenarios that
engineering categorical and numerical features. require intricate feature
engineering
Noise and outliers Since this is an ensemble model, it can handle Good fit for datasets that may
complex scenarios, making it more robust to contain noise or outliers
noise and outliers.
Real-time inferences Ensemble models can be trained very quickly and Beneficial in scenarios that
therefore are ideal for time-sensitive applications require real-time predictions
and real-time inferences.
TIP Consider using gradient boosting for complex, large datasets requiring high accuracy
because it corrects previous model errors, handles intricate features, noise and outliers, and
helps with real-time inferences.
Using K-Nearest Neighbors (KNN)

K-nearest neighbors (KNN) is helpful for both clas-
sification and regression problems. It can be used on
both categorical and numerical outcomes. It works
by identifying the K closest data points for a given
input data point and then using the outcomes of
those data points to come up with the outcome
(see Figure 16.12). The dark circle corresponds to
K=3, and the outcome is red triangles because you
have more triangles than squares. The dotted circle ?
corresponds to K=5, and the outcome would be blue
squares since they outnumber the red triangles.
Say you want to predict the price of a house based on
its features, such as number of bedrooms, square feet
area, and so on.
You start by assuming a value for K. Usually, this
value for K is obtained through trial and error.
You need to learn some metrics, such as utility and
distance, to measure the distance between the actual
houses in the dataset and the new house for which FIGURE 16.12: Example of K-NN classification
you need to make a prediction. Source: Wikipedia / CC by 3.0 / Public Domain
Once you identify the five closest houses, you can take
the average price to come up with a new prediction. Of course, this prediction assumes that houses with similar
features will have similar pricing.
When the value of K is small, it’s possible that your price is susceptible to the outliers and noise in the dataset. If
you choose a higher value for K, the outcome may be influenced by houses dissimilar to your new house.
In practice, KNN is an excellent algorithm to be chosen for applications such as product recommendations, image
classification, and anomaly detection.
KNN is a good candidate under the following conditions:
CONDITION DESCRIPTION EXAMPLE ALTERNATIVES
Labeled, The dataset should be Ideal for datasets where Unsupervised

structured data labeled and ideally have each instance is tagged learning algorithms
structured data. with a label for unlabeled data
Small and Works best with small to Not recommended for very Decision tree,
medium medium-sized datasets; large datasets due to high random forest, or
sized datasets larger datasets can become computational cost gradient boosting
computationally intensive. for large datasets
Classification and Suitable for problems that Not suitable for Deep learning
regression only involve classification or problems that don’t or reinforcement
regression. involve classification or learning for
regression tasks complex tasks
CONDITION DESCRIPTION EXAMPLE ALTERNATIVES
Not for real- Suitable for situations where Not ideal for situations Random forests or
time inference you can train offline and requiring real-time gradient boosting
deploy it online. If real- predictions machines (GBMs) for
time online performance real-time predictions
is essential, other models
should be explored.
Explainability A good model when Helpful when it’s necessary Decision tree,
explainability is required, to explain the decision- logistic regression
as predictions can be making process for explainability
easily traced.
TIP Use KNN in classification or regression tasks for small to medium-sized, structured, and
labeled datasets. It is essential to select an appropriate value for K to balance between noise
sensitivity and the included dissimilar data points.
Data Analysis and Predictive Modeling Using Neural Networks

Here are the conditions under which neural networks are recommended:
Large data Suitable for very large amounts of data Data-intensive fields like genomic
research or social network analysis
Complex Particularly well suited to analyze Image or speech recognition tasks

pattern analysis complex patterns in data
Predictive modeling Well suited for predictive modeling to Predictive maintenance or

make accurate predictions based on forecasting sales
historical data
Real-time Can be used for convolutional neural Autonomous vehicles or real-time

CNN use cases network (CNN) use cases that require financial trading
real-time decisions
Large-scale cloud AI Well suited for large-scale enterprise Allows workload distribution across
cloud AI implementations multiple servers and GPUs for
speedier training and inference
Refer to the section “Deep Learning” in this chapter for more details.
TIP Consider using neural networks for large datasets, complex pattern analysis, predictive
modeling, real-time CNN tasks, or large-scale cloud AI deployments.
Using Unsupervised Learning to Discover Patterns in Unlabeled

Data
Next is unsupervised learning, where the data is not Machine discovers and creates the labels itself.
labeled, and therefore, it is up to the algorithm to
find common patterns, as shown in Figure 16.13.
There are two types of unsupervised learning appli-
cations: clustering and anomaly detection.
Clustering
An excellent example of a clustering application
is when you have a large amount of customer
data. You use this algorithm to break the data into
multiple segments based on buying patterns and
other attributes. You could then analyze those dif- Original Data Clustered Data
ferent segments and identify one of those as college
students or moms, and so on, and then tailor your FIGURE 16.13: How unsupervised learning works
products and services based on which customer seg-
ments you are marketing to.
Another example of clustering applications is topic modeling. For example, you can feed the content from multi-
ple books into this algorithm, generate topics, and classify those books into different categories.
Anomaly Detection
Anomaly detection is used to detect anomalies in data. In other words, look for patterns that do not conform
to expected behavior. It is of great use in the case of fraud detection in customer transactions, such as an unusu-
ally large number of purchases in a foreign country or abnormal patient readings indicating a potentially serious
health condition. The following table summarizes the different types of unsupervised learning algorithms:
UNSUPERVISED LEARNING DESCRIPTION EXAMPLE

APPLICATION
Clustering Clustering algorithms divide a dataset into Customer segmentation

groups or clusters, based on similarity in the or topic modeling
data features.
Anomaly detection Anomaly detection algorithms identify Fraud detection or

outliers in the data that do not conform to the abnormal patient
expected patterns. readings in healthcare
TIP Use unsupervised learning when dealing with unlabeled data to discover hidden pat-
terns. Clustering helps with grouping similar data, while anomaly detection uncovers outliers
useful in fraud detection or health diagnostics.
Segmentation Using K-Means Clustering

K-means clustering is an unsupervised algorithm that groups data points into clusters, as shown in Figure 16.14.
Dim. 2
k-Means Clusters
4.4
4.3
4.2
4.1
4
3.9
3.8
3.7
3.6
3.5
3.4
3.3
3.2
3.1
3
2.9
2.8
2.7 Dim. 3
7
2.6
6
2.5
2.4
5 Cluster 1
4
2.3
2.2
3 Cluster 2
2.1 2
21 Cluster 3
4 5 6 7 Dim. 1
8
FIGURE 16.14: K-means clusters created from an Iris flower dataset

Source: Wikipedia
The algorithm starts by randomly assigning K cluster centers and then attaching the remaining data points to one
of the clusters closer to it. Then the cluster centers are recomputed by calculating the mean of all the data points
attached to that cluster center. The process is then repeated until the cluster centers do not change anymore.
K-means clustering is helpful for image segmentation, customer segmentation, fraud detection, and anomaly
detection.
These are the conditions under which means clustering can be adopted:
SITUATION DESCRIPTION USE CASE EXAMPLE
Large data You have a large amount of data to Analyzing massive user datasets in social
be processed. media platforms
Unlabeled data The data is not labeled. Segmenting a database of consumer

details without pre-established
categories
Pattern analysis When you want to identify patterns Identifying common buying patterns in
in the data. retail consumer data
Segmentation use cases There is a need for segmentation Segmenting customers into distinct
of customers, products, or any groups based on purchasing behavior
other entity.
continues
(continued)
SITUATION DESCRIPTION USE CASE EXAMPLE
Process optimization You want to optimize processes by Identifying inefficient processes in a

identifying inefficiencies or areas of manufacturing line
improvement.
Uncorrelated data Useful when the data is not highly Clustering unrelated blog posts
correlated. based on topics
Unknown number Ideally, the number of clusters is Organizing news articles into categories
of clusters not known. without pre-established groups
TIP Use K-means clustering for large, unlabeled datasets where the number of clus-
ters is unknown and the data is not highly correlated. It’s useful for tasks like customer
segmentation and process optimization.
Reducing Dimensions Using Principal Component Analysis

Principal component analysis reduces the dimensionality of a large dataset by identifying the underlying patterns
and reducing the number of features. It does this by transforming the original features into a new set of uncor-
related variables called principal components. And it does this without losing much of the original information in
the dataset. Since it deals with data that does not have labeled examples or explicit target variables, it is appropri-
ate for exploratory data analysis and feature extraction.
NOTE Given that PCA doesn’t use labels, it is considered an unsupervised learning tech-
nique.
Say you have product data served with a lot of information, such as sales history and specifications of the prod-
uct, such as its color, size, shape, and so on. Each row represents a product, and each column is an attribute of the
product. However, the dataset is so large that making any sense of that data is becoming difficult. Follow these
steps to use PCA:
STEP DESCRIPTION USE CASE EXAMPLE
Calculate the The first step is to use PCA to calculate For a dataset with variables like age,
covariance matrix. the covariance matrix of the data. This income, and purchasing history,
matrix tells you how correlated the the covariance matrix will reflect
different variables are. If two variables are how these variables correlate with
highly correlated, they will have a high each other.
value for covariance value, and if they are
negatively correlated, they will show a
negative covariance value.
STEP DESCRIPTION USE CASE EXAMPLE
Identify the Next, you use the covariance matrix to In a customer dataset, the principal
principal identify the principal components in the components might be combinations
components. data. These components are the linear of variables that explain the most
combination of variables that capture variability, such as a combination of
the significant patterns in the data. The age and income.
first principal component is the one that
explains the most variance in the data,
followed by the second.
Retain the top N After identifying all the principal If the top three principal components
components. components, you can choose to retain account for 90 percent of the data
the top N components and discard the variance, these can be retained while
rest and thus reduce the number of the rest are discarded, simplifying the
dimensions. After doing the principal data from 10 dimensions to 3.
component analysis, you discover that the
top three principal components account
for 90 percent of the data variance, as
shown in Figure 16.15. You can therefore
retain those three dimensions instead of
the ten or more you had.
Scree Plot
3.5
3
Point of Inflextion
2.5
Eigenvalues
1.5
0.5
0
0 0.5 1 1.5 2 2.5 3
Component
FIGURE 16.15: Interpreting the PCA: The start of the bend indicates that three factors should be retained.
NOTE A screen plot is a line plot of the eigenvalues of factors or principal components used
in PCA to identify the significant factors.
Here are some use cases of the principal component analysis:
Dimensionality reduction Used to reduce the dimensions in the data, thus helping to simplify complex
datasets with numerous variables
Pattern analysis Helps to reveal underlying patterns in the data by focusing on the principal
components that explain the most variance
Dataset compression Used to compress a dataset by retaining only the most informative dimensions,
which can be beneficial for saving storage and improving computational
efficiency
Noise reduction Removes noise from a dataset by focusing on the components that explain the
most variance and ignoring components associated with noise
Clustering Useful in a clustering scenario where it can be used to visualize clusters in a

lower-dimensional space after reducing the dimensionality of the data
TIP Use PCA for high-dimensional data to simplify the dataset and is beneficial for noise
reduction, data compression, and improving cluster visualization.
Feature Extraction and Dimensionality Reduction Using Singular Value Decomposition

Singular value decomposition (SVD) is a mathematical technique used in dimensionality reduction and feature
extraction tasks. Its usage in unsupervised tasks depends on the algorithm in which it is included. SVD is a fac-
torization method that breaks the matrix into three separate matrices: U, Σ, and V. Breaking it into its constituent
parts, you can understand the matrix’s structure and properties intuitively.
Say you have a dataset of all the customer purchases in a grocery store. You want to conduct a customer insight
analysis, but it is too complex to draw that insight. You can apply SVD to create three smaller components: a left
singular matrix, a diagonal matrix, and a suitable singular matrix.
The left singular matrix represents the customer and their purchases, and the suitable singular matrix represents
the purchases and the items bought. The diagonal matrix represents the strength between those relationships.
Armed with this kind of data, you can identify the group of customers based on their purchases or identify which
products are purchased together.
SVD can be used under the following conditions:
Large dimensionality When you have large dimensionality in the data and you want to reduce the
computational resources reducing the dimensions.
Identifying SVD can be used to identify the most essential features in a dataset.
essential features
Collaborative filtering SVD can be used for product recommendations, such as collaborative filtering,
to identify which products the users are likely to purchase based on their
historical behavior.
Image and signal In image and signal processing tasks, SVD can be employed to manipulate
processing and analyze data effectively. This includes tasks like image compression, noise
reduction, and signal enhancement.
TIP Utilize SVD for efficient dimension reduction and feature extraction in large datasets,
making it useful for recommendation systems and image processing tasks.
Using Autoencoders
An autoencoder is a type of artificial neural network that is primarily used in unsupervised learning and for
dimensionality reduction. It solves many problems, such as facial recognition, feature detection, and anomaly
detection, and it can generate meanings of words and in generative models to create new data that is like the
input training data.
NOTE Autoencoders are versatile tools that can be used for both supervised and unsuper-
vised learning.
The autoencoder is focused on creating the most compact representation of the input data, effectively reducing
the dimensionality of the data.
The autoencoder has an input layer that’s connected to an encoder, which is a series of hidden layers that
reduce the dimensionality of the input data. It contains two functions. It contains an encoding function that
transforms the input data into a lower-dimensional latent space representation, also known as an encoding.
Figure 16.16 shows how an autoencoder works.
ENCODER DECODER
Input Output
X Code X’
Layer Layer
FIGURE 16.16: Schema of a basic autoencoder
The encoded data is then passed to a decoder, which is a series of hidden layers. The decoder layer uses a decoder
function that re-creates the input data from the encoded representation. The output layer is compared to the input
layer and the error is used to train the autoencoder.
Autoencoders can be used for the use cases listed in Table 16.6.
TABLE 16.6: Use Cases for Autoencoders
USE CASE DESCRIPTION INDUSTRY DATA TYPE PREDICTIVE OUTPUT
Dimensionality Autoencoders learn Data Numerical Lower-dimensional

reduction lower-dimensional compression, data encoded
representations of high- image Categorical representation
dimensional data. It results processing data Image compression
in efficient storage and
Image data Feature extraction
computation.
Audio data
Video data
Feature Autoencoders can Natural Text data Salient features

extraction extract important language Image data Embeddings
features from input data processing
Sensor data Classification
learning meaningful Computer
representations for Clustering
vision
downstream tasks.
Anomaly Autoencoders reconstruct Network Numerical Anomaly score

detection normal data and in the security data Outlier detection
process identify outliers. Fraud detection Categorical
data
Data denoising Autoencoders re-create Speech Noisy Clean data

clean data by removing recognition Numerical Image denoising
unwanted variations. Image data Audio restoration
denoising Image data
Data New data similar to

Autoencoders can Image Image
generation input data
generate new data. generation Text
Text generation
TIP Leverage autoencoders in unsupervised learning tasks for efficient dimensionality

reduction, feature extraction, and anomaly detection.
Using Hierarchical Clustering

Hierarchical clustering is a clustering algorithm used in unsupervised learning to group similar data points into
clusters based on their similarities or differences. It offers several advantages, such as the ability to identify clus-
ters at different scales, and it preserves the full hierarchy of the clustering solutions.
Hierarchical clustering is a bottoms-up approach where it treats each data point as a cluster and then groups
them together to form clusters that form a hierarchical structure. It uses different linking techniques, such as
single linkage, complete linkage, and average linkage.
Hierarchical clustering is used in various domains, including biology, social sciences, image analysis, market
segmentation, customer segmentation, and so on. It can be used to cluster customers into micro-segmented groups
using various criteria, such as customer loyalty and social media data.
Personalized Recommendations Using Recommender Systems

Recommender systems are used in e-commerce, music, books, social media, and other domains to help users find
items that are relevant or interesting to them. These systems try to predict the preferences or ratings that the user
would give to an item based on the user’s interests, preferences, or past behavior.
There are two different types of recommender systems: content-based systems and collaborative filtering systems.
Collaborative systems work by recommending items to users based on the ratings or preferences of other users
who have similar interests. For example, if you liked a movie, the system may recommend movies that other users
who liked that movie also liked. Figure 16.17 shows how a collaborative filtering system works.
TV Binoculars Book Laptop
Tom Like Dislike Like Like
Nancy Like Dislike Dislike
Alfred Like Like Dislike
John Dislike Like
Laura Like Like ? Dislike
FIGURE 16.17: Collaborative filtering based on a rating system
Content-based filtering is based on the content of the item itself. For example, if you liked a book, the system may
recommend books that are like the book that you liked.
Recommender systems are used in the following areas:
Product To suggest products to users based on their past purchase and Amazon
recommendation browsing history
Movie To suggest movies to users based on their viewing history Netflix

recommendation and ratings
Social media To suggest friends, groups, or posts based on a user’s social Facebook
interactions
News personalization To recommend news articles based on a user’s reading history and Google News
topic preferences
Travel To suggest hotels, flights, and other travel arrangements based Expedia
recommendations on a user’s past search and travel history
continues
(continued)
Course To suggest courses, books, and other educational resources Coursera

recommendations based on a user’s search and educational history
Personalized To suggest treatments, medications, and other healthcare services IBM Watson
patient care based on a patient’s medical history
TIP To deliver relevant recommendations at scale, leverage recommender systems that apply
collaborative filtering or content-based algorithms.
Nonlinear Dimensionality Reduction Using Manifold Learning

Manifold learning is also known as nonlinearity dimension reduction and is used to find low-dimensional repre-
sentations of high-dimensional data while preserving its essential characteristics.
Some popular manifold learning algorithms are t-Distributed Stochastic Neighbor Embedding (t-SNE), Isomap,
Locally Linear Embedding (LLE), and Laplacian Eigenmaps.
Here are some good use cases for nonlinearity dimension reduction:
Customer To segment customers based on their purchase history, demographics, and other
segmentation factors to target their marketing campaigns effectively
Fraud detection To identify fraudulent transactions
Risk management To assess the risk of a customer defaulting on a loan or failing to meet a contract
Product To make product recommendations to customers based on their past purchases

recommendation
Enterprise AI From an enterprise AI perspective, you can use manifold learning for dimensionality
reduction, data visualization, feature extraction, and anomaly detection
Reinforced Learning: Learning by Trial and Error

Reinforced learning is a type of machine learning that allows an agent to learn by interacting with its environ-
ment via trial and error.
The agent is given a reward for taking actions that lead to favorable outcomes and a penalty for taking actions
that lead to negative outcomes. Over time, the agent learns to maximize its rewards. See Figure 16.18.
Reinforcement learning is used under the following situations:
➤➤ Dearth of data: When there is not a lot of training data available.
➤➤ The end state is not known: The desired end state is not known.
Environment
REWARD ACTION
Interpreter
STATE
Agent
FIGURE 16.18: Reinforcement learning is when an agent takes action in an environment, which leads to a
reward and is fed back to the agent.
Here are some good use cases for reinforcement learning:
Robotics RL can be used in robotics applications, such as picking and placing objects in a
warehouse, navigating through cluttered environments, and even playing sports.
Finance Can be used in finance to develop algorithms to predict stock prices and to automatically
place trades.
Customer service Used to create chatbots to place orders and troubleshoot problems.
Self-driving cars Optimizes the behavior of self-driving cars.
TIP For problems with limited training data or undefined end states, consider a reinforce-
ment learning approach where algorithms learn through trial-and-error interactions.
EXERCISE: REINFORCEMENT LEARNING
Let’s assume you have a robot that needs to be programmed to pick and place objects in
various locations.
Here are the high-level steps you can follow. Actual details are beyond the scope of the book,
but you can find additional resources from Open AI Gym, GitHub, and so on.
Step 1 Create a reinforcement learning environment using a software library such as

Open AI Gym.
Step 2 Define the actions the agent can take.
Step 3 Define the rewards the agent will receive.
Step 4 Train the reinforcement learning algorithm using one of the methods available.
Step 5 Test the reinforcement learning algorithm by giving the agent a new set of
tasks and check its performance.
Deep Learning
Deep learning is a subset of machine learning, and it uses the concept of neural networks, which mimic how the
human brain works. It’s a way of training computers to learn and improve independently without being explicitly
programmed. Just like our brains learn from experience, these deep learning models learn by training on large
amounts of data. Figure 16.19 shows how deep learning works and how it’s different from traditional and other
machine learning algorithms.
Traditional
Traditional programming Input Data Software New Data Predictions
Code
Manually
Machine learning Input Data Developed Model New Data Predictions
Features
Machine
Deep learning Input Data Developed Model New Data Predictions
Features
Artificial Neural Network
FIGURE 16.19: Deep learning is different from machine learning and traditional programming.
Neural networks consist of layers of interconnected nodes, where each node processes the data and sends it to
the next layer. This helps the neural network represent the data and generalize the patterns in that data. These
multiple layers allow these deep learning models to learn more complex data representations.
This ability to learn complex data makes these deep learning models good candidates for speech and image recog-
nition, natural language processing, and gaming use cases.
TIP One advantage of deep learning is that you do not have to do manual feature extraction
because the system can do that on its own.
From an enterprise AI perspective, deep learning can improve the quality of customer service interactions because
of its ability to improve the accuracy of speech and image recognition. Deep learning can also be used to detect
anomalies in financial transactions and fraud detections and improve supply chain operations by predicting
demand and optimizing inventory levels.
Deep learning can process a wider range of data resources, requires less preprocessing from humans, and pro-
duces more accurate results than traditional machine learning approaches.
Implementing deep learning in an enterprise does come up with its own nuances.
TIP Because of the intensive nature of the neural networks, you need a robust data storage,
processing, analytics infrastructure, and specialized hardware such as GPUs for proper
training and inference.
You also need strong data scientists and engineers who can design, build, and train deep learning models and
deploy them in production. They need to understand the different types of deep learning models available to use
the appropriate model to solve the business problem at hand. A proper understanding of the business require-
ments to develop technical specifications to adopt deep learning is also required.
Convolutional Neural Networks

Convolutional neural networks (CNNs) are a type of artificial neural network designed to handle structured grid-
like data such as images, videos, and time series. They are well suited for image classification and object detection
use cases due to their ability to learn spatial relationships in data.
As shown in Figure 16.20, CNNs operate by applying a series of convolution operations to the input data. A
convolution operation is a mathematical operation that takes in two functions as inputs and produces a third
function. The two inputs are the data and the filter, which is a small matrix of weights used to extract features
from the input data.
Fully
Connected
Convolution
Pooling Output
Input
Feature Extraction Classification
FIGURE 16.20: Typical CNN architecture
The convolution operation is then applied to the input data in a sliding window fashion. They filter across the
input data, one pixel at a time. The output of a convolution operation is a feature map. Feature maps can then be
used to classify the data or detect objects in the input data.
CNNs are useful under the following conditions:
SITUATION EXPLANATION
Grid-like When the data has a grid-like structure such as images.

data structure
Spatial When the data has spatial relationships where the position of the features is important
relationships for classification.
Large data When the amount of data is large. Since CNNs learn in a hierarchical manner, they can
learn from large amounts of data without overfitting.
Noisy data When the data has a lot of noise. Since CNNs learn in a hierarchical manner, they can
learn features despite noise in the data.
CNNs are used in several applications, including the following:
APPLICATION USE CASE
Object detection Utilized in self-driving cars to detect objects such as other cars, pedestrians,
and cyclists
Image classification Employed in Google Photos to automatically label the photos based on what’s
in the image
Natural language Adopted by Amazon to categorize customer reviews as positive or negative

processing
Medical imaging Leveraged by doctors to detect cancer in mammograms or identify Alzheimer’s

disease in brain scans
Financial trading Used by investment firms to predict stock prices or identify patterns in
market trends
TIP For machine learning problems involving image, video, audio, or time-series data,
leverage CNNs to automatically learn spatial relationships and features.
EXERCISE: BUILDING A CONVOLUTIONAL NEURAL NETWORK FOR IMAGE

CLASSIFICATION
The goal is to develop a CNN model that can accurately classify images in a predefined
dataset. But the larger intention is for the team to gain insights into the end-to-end process of
developing, training, and deploying a deep learning model.
Dataset Data scientist First, you must gather a labeled A labeled

collection dataset of images for the CNN to dataset of
classify. images
Preprocess Data scientist Clean and preprocess the data. As A

data part of cleaning, you might need preprocessed
to resize images, normalize pixel dataset ready
values, and split the dataset into for model
training and testing subsets. training
Model Data scientist You should then choose a suitable Selected

selection CNN architecture. Start first with model
and design a simple architecture and then architecture
progressively add complexity.
You can use transfer learning with
pretrained models if needed.
Model Data scientist Train the selected model using the A trained CNN
training training dataset. During this step, model
choose an appropriate loss function
and optimizer. Use techniques such
as data augmentation, dropout,
or batch normalization to improve
performance and reduce overfitting.
Model Data scientist Test the model’s performance using Evaluation

evaluation the test subset of the dataset. results
Measure accuracy and review the
classification report. You should plot
and review the confusion matrix.
Parameter Data scientist If the model’s performance is not A fine-tuned

tuning satisfactory, continue with tuning model with
the parameters of the model or the improved
training process. You can adjust performance
the learning rate, batch size, or the
architecture of the CNN.
Model Data scientist/ Once the model performance is Deployed

deployment ML engineer satisfactory, go ahead and deploy CNN model
the model for real-world image for image
classification tasks. classification
Recurrent Neural Networks

Recurrent neural networks (RNNs) are artificial neural networks designed to process sequential or time-series
data. RNNs can learn long-term dependencies in data, which makes them suitable for natural language process-
ing and speech recognition tasks.
RNNs use a feedback loop to connect the output from one layer to the input of the next layer. This allows the
network to learn relationships between data points that are not immediately next to each other.
Some of the key features of RNNs are sequential processing to analyze patterns in the data, recurrent connections
to allow data to be passed from one step to the next, hidden state that acts as a memory, and backpropagation to
help them learn and adjust the weights based on the sequence of inputs.
RNNs are suitable under the following situations:

➤➤ Sequential data: When the data is sequential, such as audio, speech, and text
➤➤ Long-term dependencies: When the data has long-term dependencies, such as when the meaning of a
word depends on the context
➤➤ Noisy data: When the data is noisy
RNN can be used for the following use cases:
Translation To translate between languages such as in Google Translate.
Speech recognition To recognize speech as in Alexa.
Forecasting To forecast. For example, a bank may use it to predict the price of gold in the
next month
Object detection To identify objects such as in self-driving cars.
TIP For sequential or time-series data like text, audio, or sensor streams, RNNs are well-
suited to learn temporal relationships.
Transformer Models
A transformer is an advanced neural network used for natural language processing tasks such as machine transla-
tion, text generation, text summarization, natural language understanding, speech recognition, and question
answering.
Transformers work based on the attention mechanism, which allows them to learn long-term dependencies in
text. They calculate similarity scores between each input token and output token. Similarity scores are then used
to compare the weights of the input tokens to the output token, which the transformer then uses to cater to dif-
ferent parts of the input to generate the output text.
The transformer model can learn long-range dependencies because of the use of the attention mechanism and is
also more efficient because it is not as computationally intensive as the recurrent neural network.
Transformer models are suitable under the following conditions:
Translation By capturing the relationship between words in different languages, they helped
translate between languages.
Contextual meaning By capturing the contextual meaning of the words, they have been effective
in sentimental analysis, text classification, named entity recognition, and text
summarization.
Question answering They have been able to help answer questions based on context.
Chatbots Useful for building conversational chatbots.
Recommendation systems Useful for building recommendations systems by analyzing the user preferences,
historical data, and item features.
Image processing Helpful for image classification, object detection, image generation, and image
captioning.
Speech recognition Used in speech to text and text to speech conversions.
Information extraction Useful in assisting with extracting key information from documents, document
classification, summarization, and information retrieval.
Text generation Useful in text generation for chatbots, story generation, content creation for
marketing, and so on.
Here are some examples of use cases for transformer models:
Translation By capturing the relationship between words in different languages, they helped
translate between languages.
Contextual meaning By capturing the contextual meaning of the words, they have been effective
in sentimental analysis, text classification, named entity recognition, and text
summarization.
Question answering They have been able to help answer questions based on context.
Chatbots Useful for building conversational chatbots.
Recommendation systems Useful for building recommendation systems by analyzing the user preferences,
historical data, and item features.
Image processing Helpful for image classification, object detection, image generation, and image
captioning.
Speech recognition Used in speech-to-text and text-to-speech conversions.
Information extraction Useful in assisting with extracting key information from documents, document
classification, summarization, and information retrieval.
Text generation Useful in text generation for chatbots, story generation, content creation for
marketing, and so on.
TIP Use transformers to provide state-of-the-art performance on problems such as transla-

tion, question answering, sentiment analysis, and dialogue.
Generative Adversarial Networks

Generative adversarial networks (GANs) are a type of machine learning model that is used to generate new data
like the data that it is trained on. GANs have been used to create audio, text, and images that look realistic but do
not exist in real life.
GANs consist of two neural networks—one is the generator and the other is the discriminator—as shown in
Figure 16.21.
➤➤ Generator: The generator is trained to create new data that is as realistic as possible.
➤➤ Discriminator: The discriminator is trained to distinguish between real and fake data.
As you train them, both get better at their respective tasks.
Real images Sample
Realistic
Discriminator
Images
Random input
Generator Sample
FIGURE 16.21: GAN model
GANs are suitable in the following scenarios:

➤➤ Large data: When there is a large amount of data
➤➤ Multimodal models: When the data is complex with multiple modals such as audio, text, and images
➤➤ Limited labeled data: When the data is not labeled or has limited labeling
➤➤ Realistic content: When the requirement is to create new realistic content
Here are some use cases for GANs:
Image and The image and video editing capabilities of GAN can be used to generate realistic
video editing images and videos in the entertainment, marketing, and advertising industries.
Text generation The text generation capabilities of GAN can be used to create content for websites,
social media, as well as new forms of art and literature.
Music composition The music composition capabilities of GAN can be used to create songs, soundtracks,
entire music albums, as well as for music in TV and movies.
The AI/ML Framework ❘ 309
Drug discovery GANs can be used to create new drug molecules with the potential to treat diseases.
Financial data GANs can be used to create realistic financial data, such as stock prices and
generation economic indicators, which can be used to make financial decisions and forecast
economic trends.
Climate control GANs can be used in climate control by generating realistic climate models to gain a
better understanding of climate change.
Artificial GANs can be used in artificial intelligence to create training data for other models to
intelligence improve their accuracy.
Facial GANs can be used in reconstructing a face from incomplete or degraded images that
reconstruction can be used in forensics, entertainment, and virtual character creation.
TIP For generating realistic synthetic data like images, audio, and text, consider using
GANs.
THE AI/ML FRAMEWORK

The AI/ML framework is a set of tools and resources that developers can use to build and train machine learn-
ing models.
Say you want to build a machine learning model to identify customers likely to default on a loan. You must
collect data, clean it before using it in a model, train the model, test the model, and deploy the model. All of this
requires code, and as a developer, instead of building this code from scratch, you can use prebuilt libraries from
frameworks such as TensorFlow, PyTorch cafe, and so on.
TIP AI/ML frameworks provide prebuilt components for data preprocessing, model
architecture, building, and deployment, allowing you to focus more on the functionality and
business problem.
TensorFlow and PyTorch

While TensorFlow is an open-source framework developed by Google, PyTorch is an open-source framework
that was developed by Facebook. These frameworks are powerful tools used to create new products and services,
transform industries, and build new markets. Table 16.7 compares these two frameworks.
TABLE 16.7: TensorFlow vs. PyTorch ML Frameworks
ATTRIBUTE TENSORFLOW PYTORCH
Released on November 9, 2015 October 7, 2017
Developed by Google Facebook
continues
(continued)
ATTRIBUTE TENSORFLOW PYTORCH
License Apache 2.0 Apache 2.0
Supported Python, C++, Java Python, C++, Lua

programming
languages
Focused on Large-scale machine learning Deep learning
Community support Large and active Growing and active
Available Extensive and well-written Good, but not as extensive as TensorFlow’s

documentation
Ease of use Learning curve involved User-friendly
Performance Very fast Fast
GPU supported Yes Yes
Deployment Supports a variety of deployment Supports a variety of deployment options,

options, including TensorFlow Lite including PyTorch Lightning and PyTorch
Mobile
Distributed Yes Yes

computing
Visualization Good visualization tools, including Good visualization tools, including the
TensorBoard PyTorch Debugger, Visdom
Community support Large and active community Growing community
Overall A flexible and easy-to-use framework for

A powerful and versatile framework
deep learning
for large-scale machine learning
Keras
Keras is a high-level neural network API built to run on top of other frameworks, such as TensorFlow and Theano.
Keras aims to make model development user-friendly on top of complex frameworks such as TensorFlow and Theano.
Keras has a user-friendly interface for beginners, but even experienced developers can use it to build models and
experiment with deep learning. It reduces the number of user actions, provides clear error messages, and makes it
easier to develop documentation and guides. It makes it easy to define and train deep learning models and is help-
ful for complex tasks such as image classification, NLP, and speech recognition. It helps researchers and scientists
develop iterative prototypes and drives faster innovation through experimentation.
Caffe
Caffe stands for Convolutional Architecture for Fast Feature Embedding. It is a deep learning framework devel-
oped by the Berkeley Learning and Vision Center. It is maintained by the community and is written in C++ with
the Python interface. Since it is developed in C++, it is more user-friendly for beginners. It is less user-friendly
than TensorFlow or PyTorch. It has extensive documentation and is intended to be fast and flexible, as it can be
extended to support new tasks.
Summary ❘ 311
NOTE Caffe is helpful for large-scale, complex tasks such as image recognition, object
detection, semantic segmentation, and speech recognition.
MXNet
MXNet is a robust deep-learning framework developed as open source by the University of Washington and the
University of Hong Kong. It is used by Baidu for its search algorithms and is developed in C++. It is tricky for
beginners and is not as user-friendly as TensorFlow and PyTorch.
NOTE MXNet is an open-source solution developed by the Apache Software Foundation

and is a powerful tool for developing high-performing models.
Scikit
Scikit stands for scientific kit. It’s written in Python and is a powerful and versatile machine-learning library
for beginners. It includes classification, regression, clustering algorithms, support vector machines, and random
forests, and K-means.
Scikit is intended to work with other Python libraries—NumPy and SciPy. It’s free software released under the
BSD license. It is a good choice for beginners who want to learn machine learning.
Chainer
Chainer is a powerful open-source deep learning tool used to develop many deep learning models written in
Python on top of other libraries such as NumPy and CuPy. It can be used for classification, object detection, and
NLP use cases. It was developed by the Japanese venture company called Preferred Networks in collaboration
with IBM, Intel, NVIDIA, and Microsoft.
Many more AI/ML frameworks exist, such as CNTK, DLib, and so on.
WORKBOOK TEMPLATE: AI/ML ALGORITHM SELECTION GUIDE

Download the “AI/ML Algorithm Selection Guide” template from the download section of the book (www.wiley
.com/go/EnterpriseAIintheCloud). Use this template to choose the right AI/ML algorithm for your AI/
ML project. You can use this to improve the quality of the predictions or classifications, the quality of data, and
the amount of computation needed.
SUMMARY
You took a comprehensive trip into the world of AI and ML algorithms. Tailored for experienced data scientists
and beginners alike, this chapter demystified the process of selecting the right algorithms based on use cases.
The chapter dived into the basics of artificial intelligence, reviewed different types of machine learning, discussed
the nuances between models and algorithms, and provided hands-on exercises to help you choose the right
algorithms.
Understanding these algorithms is akin to processing the keys that unlock the vast potential hidden in the data.
While choosing the algorithms, note that the choice of an algorithm is rooted not only in the data but also in the
business problem you are trying to solve. In subsequent chapters, you can look forward to diving deeper into the
training, tuning, evaluating, and deploying models. Your journey continues!
REVIEW QUESTIONS
1. What are the types of data that machine learning algorithms can use?
A. Text data
B. Images
C. Only structured data
D. All kinds of data
2. What is the primary purpose of AI?
A. To automate business processes
B. To improve customer service
C. To mimic human intelligence
D. To monitor events and take action
3. The purpose of labeled data in supervised learning is
A. To identify input variables
B. To predict continuous values
C. To create a model
D. To define the output variable
4. Which one of the following is a regression problem?
A. Identifying spam emails
B. Forecasting product demand
C. Recommendation systems
D. Classifying images
5. The purpose of reinforcement learning is to
A. Classify data into labels
B. Predict continuous values
C. Learn from interactions and rewards
D. Automate process interactions
6. Random forests are useful for which types of problems?
A. Classification
B. Regression
C. Both classification and regression
7. How does KNN work?
A. It uses deep learning to work.
B. It calculates the mean of all data points to make predictions.
C. It identifies the K closest data points and uses their outcomes to make predictions.
D. It randomly assigns clusters to data points.
8. Principal component analysis is useful for

A. Classification problems
B. Regression problems
C. Dimensionality reduction
D. Real-time CNN use cases
9. What is a singular value decomposition?
A. Clustering algorithm
B. Dimensionality reduction technique
C. Image processing technique
D. Classification technique
10. A key advantage of deep learning compared to traditional machine learning is which of the following?
A. Can work with smaller amounts of data
B. Can work with less human intervention
C. Produces less accurate results
D. Performs better on small datasets
11. CNNs are particularly good for
A. Image data
B. Text data
C. Sequential data
D. Tabular data
12. RNNs are particularly good for
A. Image data
B. Text data
C. Sequential data
D. Tabular data
13. Transformer models are useful for
A. Image classification
B. Text generation
C. Anomaly detection
D. Generative adversarial networks
14. This AI/ML framework is written in C++ and is less user-friendly than TensorFlow or PyTorch.
A. Keras
B. Caffe
C. MXNet
D. scikit-learn
15. ______ is known for its versatility and includes algorithms for classification, regression, and clustering
use cases.
A. Keras
B. Caffe
C. scikit-learn
D. Chainer
ANSWER KEY
1. D 6. C 11. A
2. C 7. C 12. C
3. D 8. C 13. B
4. B 9. B 14. B
5. C 10. B 15. C
17
Training, Tuning, and Evaluating
Models
Success is not final; failure is not fatal: It is the courage to continue that counts.
—Winston Churchill
Traveling the AI journey is like solving a jigsaw puzzle. In the previous chapter, we got one step closer to
solving this puzzle by learning how to choose the most apt AI/ML algorithm, a critical step in transitioning
data into actionable intelligence. This chapter builds that up to get into the actual act of modeling, which
brings these algorithms to life.
This chapter delves into the three pillars of modeling: training, tuning, and evaluation (see Figure 17.1).
The focus is on ensuring these models are secure, efficient, and high performing. We begin by looking at
the intricacies of model building and the likely challenges during the training process. From distributed
processing to container code, we dive into the tools and techniques of model development. You learn about
optimizing models for high performance using hyperparameters and model tuning. Evaluation and valida-
tion will help you align with your enterprise objectives when dealing with real-world data and ensure the
models are robust and ready to achieve tangible, transformative outcomes.
MODEL BUILDING
Model building is an iterative process, and it is the first step in the process that sets up the model for train-
ing to begin. Figure 17.2 shows the model development lifecycle.
The purpose of model building is to build a working model that can be used to make predictions on new
data. It is the process of defining the model’s structure, parameters, and hyperparameters. Taking the cook-
ing analogy, model building is like building a recipe to build your meal.
BEST PRACTICE TIPS FOR PICKING AN ALGORITHM
Start by selecting an algorithm suited to your data type and problem. However, always be open
to trying different algorithms based on model performance and the insights gained during the
training phase.
316 ❘ CHAPTER 17 Training, Tuning, and Evaluating Models
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Govern Your
Your AI
Team Process Data and Modeling
journey
09 • Process Data and Engineer Features in
ENTERPRISE AI the Cloud
Setup 04 • Choose your AI/ML Algorithm
Evolve and
Mature AI
• Train, Tune, and Evaluate Models
and Manage
Operations
08 Objective
Master the iterative process of refining AI/ML
AI
Deploy and 05 models through rigorous training,
Monitor
Models hyperparameter tuning, and evaluation,
ensuring they are ready for deployment with
07
high precision, reliability, and efficiency
06
FIGURE 17.1: Training, tuning, and evaluating models
Model Building Model Training

g Mode
Develop algorithm to identify ildin lT
rai Train the model using
l Bu ni
spam emails based on de n training dataset
content and metadata
Mo
Model
aluation
Model Tun
Development
Lifecycle
Model Evaluation Model Tuning
l Ev
Evaluate the model’s Tune the model to improve

de
ing
performance against the M

o its performance
testing dataset Mo
del Validatio
n
Model Validation
Validate the model using a
validation dataset
FIGURE 17.2: Phases of a model development lifecycle
You will be using various tools and frameworks such as TensorFlow, PyTorch, scikit-learn, and so on.
Structure, Parameters, and Hyperparameters

The structure refers to the algorithm or architecture used, such as linear regression, decision tree, or neural net-
work. The parameters refer to the internal variables related to the model, such as weights and biases. These values
are estimated from the data during training. You can think of the model as the hypothesis and the parameters as
the customization of the hypothesis to predict from a specific dataset. Hyperparameters refer to the external set-
tings configured by the user before training the model, such as learning rate, batch size, and regularization.
Figure 17.3 lists the training and tuning components discussed in this chapter.
Model Building ❘ 317
01 Building 03 Model
05 Validation
07 Training
09
Code Training Metrics Container
Algorithm Debugging/ Hyperparameter Model

Features
Selection Profiling Tuning Artifacts
02 04 06 08
FIGURE 17.3: Model training and tuning components
Steps Involved During Model Building

Model building involves the following steps:
1. Choosing an appropriate algorithm, such as a decision tree, linear regression, or neural network based
on the problem to be solved
2. Defining the features and the label that needs to be predicted
3. Choosing the loss function such as mean squared error
4. Building the model architecture if required
Model building is all about experimenting with different algorithms, features, and hyperparameters iteratively
until a satisfactory performance is achieved. As part of this process, you may be required to carry out some data
preprocessing, such as scaling, encoding, or feature engineering.
TIP Model architecture refers to the algorithm or framework that is used to predict. It can
refer to the layers, nodes, activation functions, loss functions, and optimization algorithms
that are arranged and configured in a certain way. Some common machine learning model
architectures are linear regression, decision trees, neural networks, convolution neural net-
works, and recurrent neural networks.
Developing the Algorithm Code

As a machine learning engineer, you can write your algorithm code in different languages, such as Python, R, or
Java, depending on your preferences or environment. Your code will need to do the following:
➤➤ Import the data
➤➤ Clean the data
➤➤ Split the data
➤➤ Train the model
➤➤ Make predictions
Selecting the Right Algorithm

During this model building process, you will be experimenting with different algorithms and parameters to short-
list the right algorithm. You will be evaluating the models based on different parameters such as training/predic-
tion time, accuracy, explainability, memory requirements, and so on.
For our case study, you may choose classification algorithms such as logistic regression, naïve Bayes, and support
vector machines as possible candidates to build the model.
NOTE Refer to Chapter 16 on algorithms to learn more about various algorithms and when
to use which one.
MODEL TRAINING
Once you have shortlisted a model during model building, the next step is to train your model with the training
dataset. Model training is a crucial step in machine learning that results in a working model, which you can vali-
date, test, and deploy. Using the cooking analogy, model training is like actually starting to cook the meal using
the recipe you came up with during the model building process.
NOTE A model is a set of rules or equations that the machine can use to learn from data to
make predictions. Depending on the type of problem, these predictions could be about iden-
tifying if an email is spam or classifying images as different types of animals.
During this step, you will feed the training data into the model and begin the training process. The training
process teaches the machine how to make its predictions. During this process, the weights (parameters) of the
algorithm get updated to reduce the loss, which is the difference between the prediction and the actual labels. In
other words, the training process helps to increase the accuracy with which the machine makes its predictions.
TIP Keep a close eye on metrics such as accuracy, precision, and recall to check if the model
is performing well or needs adjustment.
Distributed Training
When dealing with a lot of data or a complex model, your training process can take a long time, sometimes even
days or hours. It is where distributed training comes to our rescue. Distributed training involves using multiple
machines to train your model.
You have two types of distributed training, namely, data parallelism and model parallelism.
Model Parallelism
Model parallelism involves splitting the model into smaller parts to predict different features and assigning each
part to different machines. The models are then combined to form the final model.
Model parallelism is more complex than data parallelism but is more suited for models that cannot fit into one
device. You can also combine model parallelism with other techniques such as data parallelism to improve train-
ing efficiency.
Data Parallelism
Data parallelism involves splitting the training dataset into mini batches evenly distributed across multiple nodes.
For example, if you have 1,000 emails to identify spam, you can divide them into 10 groups of 100 emails each
and assign each group to a different computer. Therefore, each node trains on only a fraction of the dataset. You
will update the model with the results from the other machines to make the final model.
Model Training ❘ 319
NOTE An optimization algorithm helps update the parameters of the model to reduce a
loss function, such as mean squared error. A good example is the gradient descent algorithm,
which adjusts the parameters iteratively in the direction of the steepest descent of the loss
function.
Problems Faced When Training Models

When training your models, you are going to face a number of problems, some of which are listed in Figure 17.4.
We will review these challenges in this section.
• Model fitting noise in data • CPU, memory, network icons

• Poor performance on new • Slowing training
data
01 02
Overfitting Bottlenecks
CHALLENGES IN
MODEL TRAINING
Saturated
Vanishing Gradients
Activations
03 04
• Gradient slope graphic fading • Output flatlining
• Model stops learning • Model not improving
FIGURE 17.4: Problems faced when training models
Overfitting and Underfitting

Overfitting happens when you are training the model with a lot of training data. Your model starts to memorize
the labels and is unable to generalize to new data. It happens when the model captures every noise or irrelevant
pattern in the data. The model becomes too complex and is unable to predict against new data. It can lead to
errors and inaccuracies in your model predictions on validation or test data.
Underfitting is the reverse of overfitting. In this case, the model is too simple and does not capture important pat-
terns or relationships that are relevant for the prediction.
System Bottlenecks
When your system faces a bottleneck, it can bring down the entire process performance regarding memory, CPU
disk space, or network bandwidth.
Saturated Activation Functions

Saturated activation functions mean the output does not change even if the input changes, which means the
neural network cannot learn anymore. It happens when the activation functions, namely, sigmoid and tanh, have
reached their minimum or maximum levels. When this happens, we say the activation functions are saturated.
ReLU is an alternative in such situations.
Vanishing Gradients
Vanishing gradients happen when the gradients or the incremental changes to the neural network weights are
minimal or close to 0 during backpropagation. When this happens, the neural network will not learn much from
the data. It can happen when using sigmoid or tanh, as well as when the gradients are not propagated back
through the network.
TIP Think of saturated activation functions as rusty gears in a machine. If they are not
moving, neither is your model’s learning. Similarly, vanishing gradients are like whispers in a
noisy room and your model is not hearing them. Opt for RelU for activation functions and
proper weight initialization to address these issues.
To resolve this situation, you can either use ReLU or use weight initialization schemes to prevent the gradients
from becoming too small.
You can use a debugger tool to keep track of the network’s inputs, outputs, gradients, weights, metrics, logs,
errors, and so on. The tool can also capture the state of a machine learning job at periodic intervals, such as vari-
ous epochs, batches, steps, or layers.
Some examples of a debugger are the PyTorch debugger, TensorFlow debugger, SageMaker Debugger, and Azure
Machine Learning debugger. ML debuggers help you to debug and profile the performance at various intervals
using breakpoints, alerts, watches, visualizations, and filters.
Training Code Container

A training code container comes in handy when you want to include the training code and all the environmental
dependencies into one package to take care of code dependencies.
TIP Building training code containers is an essential step in the machine learning process to
build and deploy models quickly and reliably at scale.
Suppose you’re trying to build a model to predict customer churn. You have collected all your data and used the
logistic regression model to predict customer churn. And now, you have written some Python code in your local
machine to train the model. You have integrated your code with other frameworks and libraries such as scikit-
learn and pandas.
Now, you are all set to deploy your working code into production using the cloud. Unfortunately, this is where
you are going to face some issues, such as the following:
➤➤ Your code may not run due to differences in the hardware, software, and operating system configuration.
➤➤ Your code may not be able to access the data that is stored in the cloud platform.
➤➤ You may need to scale up or down your compute and storage resources depending upon the workload.
Here comes your training container to your rescue. You will start by creating a container image with the training
code and the entire dependent stack, such as the libraries, framework, tools, and environment variables for your
model to run. You can then train the model quickly and deploy it on another platform at scale because now your
code is running as part of a self-contained unit with all the dependencies taken care of.
Model Training ❘ 321
Creating a Training Code Container

To create a training code container, you need the following:
➤➤ A Docker file: This file contains the instructions to build the image.
➤➤ train.py: This file contains the training code for the machine learning model.
➤➤ requirements.txt: This file contains the dependencies such as scikit-learn and pandas for the machine
learning model to run.
You can then use the Docker file to build the container image and push it to a container registry such as the
Docker Hub or the Google Container Registry.
Using the Training Code Container

The container image can be pulled to train the model onto any machine that has docker installed. You need to
specify the data sources, compute resource requirements, and set other parameters for the training job. You can
save the file once the training is completed and you have chosen the right hyperparameters.
Using that file for real-time predictions, you can then deploy the model on any other machine.
Model Artifacts
You need to realize that the final model is one of many outputs you will get when training a model. There will
also be other model artifacts, such as the trained parameters, model definition, and metadata. These model arti-
facts are essential pieces of the puzzle to understand, use, reproduce, and improve upon your model. Figure 17.5
shows the artifacts.
Weights and biases, the

relationship between
Trained features and output
Parameters
01
02
Model architecture,
Model the layers, nodes,
Definition
data flows, output
Metadata
Preprocessing steps,
03 library versions,
hyperparameters,
validation metrics
FIGURE 17.5: Model artifacts outputted during model training
Trained Parameters
In the case of predicting the price of a house, you had trained parameters such as the weights and biases that your
model learned during the model training process. These parameters define the relationship between the features
and the price of the house. They basically determine how the model works.
Model Definition
Model definition defines the architecture or structure of the model. It defines how the data flows into the model,
the layers, the nodes, and how they interact. In the case of the house pricing prediction model, this could define
the structure of the neural network in all the regression models. It would detail how the inputs (house features)
are processed to produce the output (predicted price).
Other Metadata
This could be any information that you may have used during the model training process, such as the preprocess-
ing steps that you used to prepare the data, the versions of the libraries, the hyperparameters to tune the model,
and the validation metrics that you used to evaluate the model, and so on.
TIP It is vital to keep track of the model artifacts as they will help in reproducibility,
governance, explainability, troubleshooting, lineage tracking, and centralized artifacts
management.
MODEL TUNING
Model tuning involves fine-tuning the parameters not optimized during the training process. Model tuning helps
to modify the structure of the model itself. During model training, the internal trainable parameters, such as
weights and biases of the neural networks, are trained using the training data.
In the case of model tuning, the focus is on tuning hyperparameters. Those hyperparameters are structural set-
tings that are set before fitting the model and are not learnable from the training data. These are parameters such
as the learning rate, number of epochs, regularization strength, and so on.
TIP Think of model tuning as sharpening your pencil. While training sets the foundation,
tuning is where you achieve perfection.
Hyperparameters
Suppose you are a chef trying to create a recipe to bake bread. While creating your bread, you can change a few
parameters, such as the temperature and the baking time. You can think of these parameters as hyperparameters.
Hyperparameters control how the model works and help optimize the model performance on training and valida-
tion datasets. It helps to avoid overfitting or underfitting. These values cannot be estimated from data and are
often configured by a practitioner. Hyperparameters help to control the behavior of the algorithm or the model
and control how fast the training occurs and how complex the model is. Hyperparameters are set before the
model gets trained, and they guide how the model learns its trainable parameters.
I next discuss the learning rate, regularization, batch size, number of epochs, hidden layers, and hidden unit
hyperparameters, as shown in Figure 17.6.
Learning Rate
Learning rate is the rate at which the model parameters are updated at each iteration of the gradient descent.
You can think of the learning rate as the baking temperature. A high temperature can speed up the baking, but
it can get spoiled by overheating. Similarly, a high learning rate speeds up the convergence to minimize the loss
Model Tuning ❘ 323
function. However, it can also cause instability or overshooting. A low learning rate can ensure stability but can
take a long time to converge and potentially get stuck in local minima.
Affects convergence speed
Learning Rate
Determines
Counters overfitting
capacity Hidden Regularization
Units
Hyperparameters
Adds model
Impacts memory usage
complexity Hidden Layers Batch Size
Epochs
Controls training cycles
FIGURE 17.6: Tuning hyperparameters for optimal model performance
Regularization
Regularization is adding a penalty to the loss to reduce the complexity or magnitude of the parameters. While this
approach may reduce the variance and help avoid overfitting, it can introduce bias. Using the cooking analogy,
compare the situation of overfitting to having added more salt. To reduce this saltiness, you would add more
water. You can think of regularization in a similar manner.
Batch Size
Batch size is the number of data points in each dataset used to train the model for each iteration of the gradi-
ent descent.
A large data size can reduce the variance in the parameter updates and noise, but it takes a long time to compute
and uses a lot of memory. Smaller batch size is the reverse effect: it reduces the time taken to compute and uses
less memory. However, it increases the noise and variance in the parameter updates.
Number of Epochs
The number of epochs stands for the number of times the entire training dataset is fed to the model for training.
A higher number means the model runs the risk of overfitting, while a lower number may lead to underfitting of
the model.
To find the sweet spot for the number of epochs, you can resort to techniques such as early stopping in the case of
overfitting and adopt techniques such as cross-validation to get better results.
Hidden Layers
Hidden layers are the layers that are not visible to the user in a neural network.
NOTE The hidden layer is an important parameter that can decide the complexity of the
model.
Simple layers can catch broad features, and deeper layers can catch more intricate, specific details. A higher num-
ber of hidden layers can result in overfitting, while a lower number can lead to underfitting. To identify the right
number of hidden layers, you can try grid search or random search techniques to evaluate the performance using
different numbers of hidden layers.
Hidden Units
Hidden units are the number of neurons in a layer in a neural network. They determine the representational
capacity of the neural network. A higher value can lead to overfitting, and a lower value can lead to underfitting.
You can identify the correct optimal number of hidden units using techniques such as a grid or random search to
evaluate the network’s performance using different numbers of hidden units.
Choosing the Right Hyperparameter Optimization Technique

You can employ manual or automated model-tuning techniques such as grid search, random search, or Bayesian
optimization (see Figure 17.7).
• Tries all hyperparameter • Samples random combinations

combinations
Grid Search Random Search • Faster but may miss optimal
• Finds global optimum but slower setting
FIGURE 17.7: Grid versus random searches
You can use your intuition or experience when choosing the hyperparameters to manually tune a model. Your
goal is to choose the model with the best validation performance. While using an automated approach, the grid
search provides the best possible model. In contrast, the random search approach is more efficient.
TIP If you’re running on, you will choose the random search approach, but if you’re looking
for accuracy, you will choose the grid search approach.
Grid Search vs. Random Search

Say you’re trying to predict whether a patient has cancer, and you’re using a random forest model that contains
two hyperparameters, namely, the depth of the tree and the number of trees in the model. See Figure 17.7.
Model Validation ❘ 325
To use grid search to tune the hyperparameters, you can set the maximum depth of the decision tree to 10, 20, 30,
40, and 50. You can set the number of trees in the random forest to 100, 200, 300, 400, and 500.
When you use grid search, the system will try all 25 combinations of the hyperparameter values, which will take a
long time but will provide the best model.
If you used random search, the system would randomly sample the maximum depth of the tree from a uniform
distribution between 1 and 50 and randomly sample the number of trees from a uniform distribution between
100 and 500.
NOTE While random search will not give the best possible model, it will be faster than the
grid search approach.
Bayesian Search
Bayesian search is a hyperparameter optimization technique well suited to noise or use cases that are more com-
plex for a grid or random search. It is more efficient than grid or random search but can be complex to imple-
ment and may need some knowledge of Bayesian statistics.
TIP Bayesian search can be used for any machine learning model but may not be as effective
as a grid or random search for more straightforward objective functions.
The Bayesian search uses Bayesian statistics to find the best hyperparameters and works by building a probabil-
istic model of the objective function being optimized. It has been successfully used in many applications such as
spam filtering and credit card fraud detection.
MODEL VALIDATION
This section discusses how to validate the models and the metrics that you have to use.
Choosing the Right Validation Techniques

You will need to validate your model using the validation dataset during the training phase when it is being devel-
oped and optimized. The goal of the model validation process is to compare the predictions of the model with the
actual outputs of the new dataset. You will use techniques such as holdout validation and cross-validation. This
evaluation aims to test the model’s generalization ability, identify any issues with the model, and choose the best
model among the candidates. Its focus is to avoid overfitting the training data. Model validation happens during
the training phase, while model evaluation happens after the model is deployed in production.
There are several cross-validation techniques—the two most popular ones are K cross validation and leave-one-
out cross validation (see Figure 17.8).
Cross-validation techniques are more expensive and time-consuming but are very accurate. If you’re short on
time, you should go for holdout validation, but when accuracy is critical, you will use cross-validation techniques.
Holdout Validation
Holdout validation involves breaking the dataset into a training dataset and a test dataset. You will use the train-
ing dataset to train the model and then use the model to test the dataset using the test dataset. By comparing the
results with the actual dataset, you can evaluate how accurate the model is.
K-fold Cross Validation

• Dataset split into K groups
• Train on K-1 groups, test on held-out
group
Holdout Validation
Leave-One-Out
• Dataset split into two groups
• Dataset split into N groups
• Train on one, test on the
VALIDATING • Train on N-1 groups, test on held-out
other
MACHINE case
LEARNING
MODELS
FIGURE 17.8: Validating machine learning models
K Cross-Validation
This involves breaking down the validation set into K folds and training and testing the model on each fold. Then
you will compare the performance average across all folds with other models.
Leave-One-Out Cross Validation

Leave-one-out cross validation involves splitting the data into n data folds where n is the number of data points.
You would then train the data in n – 1 folds and evaluate the model against the fold not used in training.
TIP Using the leave-one-out cross-validation technique helps you test the data against all the
data points, which is very accurate.
Validation Metrics
As shown in Figure 17.9, validation metrics are used during the different phases of a model development lifecycle.
During Model Building

For example, during the model build phase, validation metrics can be used to choose the suitable model for
classification-type problems based on a validation metric such as area under the receiver operating characteristic
curve (AUC-ROC). In the case of regression type problems, use metrics such as mean absolute error or R-squared.
TIP Always select metrics aligned with your model’s objectives.

Model Building Model Training Model Tuning Model Evaluation
• Trying different • Training and • Hyperparameter • Test data used

model types validation sets combinations after finalizing
• Validating with • Validating to • Validating to model
AUC, F1, etc. prevent choose best • Validating on
overfitting tuning unseen data
parameters
FIGURE 17.9: Validation across ML model development phases
During Model Training

During the model training phase, you can break down the dataset into training and validation datasets. Once
the training is completed using the training dataset, you can validate the model on the validation dataset and opti-
mize the model parameters accordingly to prevent overfitting.
During Model Tuning

During the model tuning process, you will choose the combination of the hyperparameters using techniques such
as a grid or random search that optimizes the validation metrics.
During Model Evaluation

Once the model is trained and tuned, you can then evaluate the model using the validation metrics on the test
dataset. It helps you to understand how well the model will perform on new, unseen data.
TIP Validation metrics are used to measure the model performance in a quantifiable manner
during different phases of the model development lifecycle. During model building, use them
to choose the right model; during training, use them to ensure learning is on the right track;
during tuning, use them to choose the right hyperparameters; and during evaluation, use
them to finalize the model.
Validation Metrics for Classification Problems

This section discusses some validation metrics for classification problems, such as accuracy, precision, recall, and
F1 score, as shown in Figure 17.10.
TIP No single metric will tell you the whole story. While accuracy may be very impressive, it
can be misleading if your classes are imbalanced. That’s why you need to look at other met-
rics such as precision, recall, and F1 score.
Accuracy
Accuracy is the percentage of the correctly predicted labels out of all predictions calculated as follows:
Accuracy Number of correct predictions / Total number of predictions

F1 Score
• Formula combining
precision and
recall
• Balance between
precision and
recall
Recall
Precision
• Formula with
• Formula with
true positives
true and false
and false
positives
negatives
• Positive
• Completeness
prediction
of positive
accuracy
predictions
Accuracy AUC-ROC
• Formula with • Area under the receiver
correct
predictions
Key operating characteristic
curve
• Overall Classification
• How well the model can
performance Metrics to differentiate between the
metric Evaluate positive and negative
Models classes
FIGURE 17.10: Validation metrics for classification problems
Precision
Precision is the percentage of the correctly predicted positive labels out of the total positive predictions.
Precision Number of true positives / (Number of true positiives / (Total number of true positives
+ Total number off false positives))
Recall
Recall is the percentage of the correctly predicted positive labels out of the total actual positive predictions.
Recall Number of true positives / (Number of true positivess / (Total number of true positives
+ Total number of faalse negatives))
F1 Score
F1 score is the harmonic mean of the precision and recall.
F1 Score 2 Precision Recall Precision Recall
AUC-ROC
AUC-ROC stands for area under the receiver operating characteristic curve, and it helps to determine how well
the model can differentiate between the positive and negative classes. Its value ranges from 0 to 1. A value of 0.5
implies random guessing, and 1 implies its perfect classification. It plots the true positive rate (recall) against the
false positive rate (1 – specificity) for various threshold values.
Since these predictions are probabilities, we must convert them into binary values using the threshold concept. So,
for a threshold of .5, if the predicted probability is greater than .5, it means it’s positive, and any value less than
.5 is treated as negative.
NOTE AUC-ROC is a measure of how well a model can distinguish between two classes,
ranging from 0 to 1.
Confusion Matrix
Confusion matrix is a table that summarizes the number of true positives (TPs), true negatives (TNs), false posi-
tives (FPs), and false negatives (FNs) for a given class.
Consider the scenario for predicting customer churn. Table 17.1 shows The confusion matrix.
TABLE 17.1: Sample Confusion Matrix
ACTUAL NO CHURN ACTUAL CHURN
Predicted No Churn TN FN
Predicted Churn TP
FP
Assume your model returns the results shown in Table 17.2.
TABLE 17.2: Sample Results Predicted by the Model
CUSTOMER ID ACTUAL LABEL PREDICTED PROBABILITY PREDICTED LABEL
1 0 0.2 0
2 1 0.8 1
3 0 0.6 1
4 0
1 0.4
You can build the confusion matrix as shown in Table 17.3.
TABLE 17.3: Confusion Matrix with Values
ACTUAL NO CHURN ACTUAL CHURN
Predicted No Churn 90 10
Predicted Churn 80
20
Using the formula, your validation metrics will be as follows:

➤➤ Accuracy = (90 + 80) / (90 + 10 + 20 + 80) = 0.85
➤➤ Precision = 80 / (80 + 20) = 0.8
➤➤ Recall = 80 / (80 + 10) = 0.89
➤➤ F1-score = 2 * 0.8 * 0.89 / (0.8 + 0.89) = 0.84
➤➤ AUC-ROC = 0.93
These metrics reveal that this model is fairly accurate, but still, there is some scope to improve as it still has some
false predictions. You can use these metrics to choose between different models or algorithms. You can also use
these metrics to choose between different hyperparameters, such as regularization or threshold value, as part of
the model-tuning process.
MODEL EVALUATION
After training a model, the next important step is to evaluate the model for its performance and accuracy. Think
of model evaluation as your final examiner that validates if the model is ready for the real world.
As part of the evaluation, you have to take care of three things.
➤➤ Ensure the model can generalize well for new unseen data.
➤➤ Ensure the model satisfies the business rules, objectives, and expectations. Otherwise, it’s like a ship
without a compass.
➤➤ Ensure it can accommodate different scenarios and handle different trade-offs. The fact is no model can
handle every situation perfectly.
Model evaluation will help you to address these challenges. As part of it, you will select the suitable model after a
lot of experimentation, followed by validating the model and fine-tuning it with the correct hyperparameters.
NOTE Note that the model evaluation can take place using offline or online data.
Offline data evaluation will be done using holdout data, while online evaluation will take place in real time in
production, which can be risky but more effective.
Figure 17.11 shows the various components of the model performance evaluation pipeline. Model evaluation
involves testing the model against a test dataset not used for training or validation. It helps to test the final accu-
racy and quality of the model. You will use evaluation metrics such as accuracy, precision, recall, and F1 score.
During this phase, you will look for potential bias and discrimination in the model and also try to explain how
the model works using a feature importance score or confusion matrix.
TIP Model evaluation is a critical step to ensure your model is robust, fair, aligned with your
business objectives, and ready for the real world.
BEST PRACTICES
This section reviews some practices you can implement during the model development lifecycle to streamline
workflows, secure your platform, enhance model robustness, improve performance, and reduce costs.
Best Practices ❘ 331
Develop
Process Data Model Deploy Model Monitor
Feature Pipeline
Fetch Real-time/Batch
Features Inference
Train/Tune/ Apps
Data Prepare Deploy
Pipeline Evaluate
Re-Train Pipeline Monitor
CI/CD/CT Pipeline
Re-Train Pipeline (Performance)
Alarm
Model Registry
Store Artifacts (Data, Code, Model)
FIGURE 17.11: Performance evaluation pipeline
Streamlining Your ML Workflows Using MLOps

This section reviews the operational best practices you can employ during model training, tuning, and perfor-
mance evaluation to ensure streamlined and efficient ML workflows. Your goal must be to constantly monitor
these systems and continually improvise upon the processes and procedures.
Harness the Power of MLOps Through IaC and CaC

Infrastructure as code (IaC) and configuration as code (CaC) are two vital techniques to automate cloud infra-
structure management. IaC helps you define your infrastructure in code, while CaC helps you capture your code
configurations. Implementing IaC and CaC will provide you with two benefits:
➤➤ Implementing IaC will ensure consistency in the infrastructure across staging and production environ-
ments and help you avoid the tedious, error-prone manual process.
TIP IaC and CaC will lead to more predictable deployments and fewer manual errors.
➤➤ Implementing CaC will ensure the traceability of different data and model artifacts across multiple envi-
ronments and configurations. It will help you troubleshoot issues and avoid the manual, tedious process
of managing configurations across multiple environments.
TIP IaC and CaC help manage versions, making rolling back changes easier. They are life-
savers to scale ML operations.
You can use AWS CloudFormation, Terraform (open source), and AWS Cloud Development Kit to implement IaC.
Establish Scalable Ways to Access Public ML Libraries and Resources

When working with public libraries in an enterprise setup, you’ll likely encounter some challenges.
First, you must ensure that the ML library or framework fully complies with your company’s policies and is
approved for use. Second, you must ensure these libraries and frameworks are compatible with your machine-
learning environment.
You need to be able to share the latest and approved libraries across multiple environments, teams, and projects
to prevent disparities across teams.
Embrace Containerization
Containerization provides the following benefits:
➤➤ Governance and Compliance: By packaging the ML code, libraries, and frameworks into the container
and ensuring that these are approved by the organization, you can ensure compliance and compatibility
with environments as well as share these containers across multiple teams within the organization.
➤➤ Standardization: It will ensure standardization of the machine learning infrastructure and code, in addi-
tion to enabling governance across the enterprise.
Use an Artifact Repository

Here are some benefits of using an artifact repository.
➤➤ Centralized storage: Storing the model, data, and libraries in an artifact repository helps to ensure that
all teams are accessing the same, approved resources.
➤➤ Version control: Repositories help you to track the versions of these artifacts easily and roll back
changes in case of issues.
➤➤ Enhanced collaboration: Central repositories make it easier to share artifacts across multiple projects,
teams, and data scientists so that they can build upon each other’s work.
TIP Implementing MLOps best practices will not only streamline your ML development
process but also ensure consistency, governance, and collaboration, which are foundational
pillars for an agile, reliable, and scalable AI implementation.
Securing Your ML Platform

This section discusses some of the best practices to ensure the security of your models, data, infrastructure, and
other assets, as listed in Figure 17.12.
Protect from Transfer Learning Risks

Transfer learning is a technique in machine learning that leverages the learning gained by a pretrained model
when solving a specific business problem to solve another similar problem by fine-tuning it. When doing so, we
run the risk that the weights inherited from the pretrained model may not produce the correct predictions.
TIP To avoid transfer learning risk, you can use the AWS SageMaker Debugger to detect any
bias in the new predictions.
Protect from
Transfer
Learning Risks
Machine
Protect Against Securing Your
Learning
Data Poisoning ML
Security Best
Threats Environment
Practices
Encryption of
Inter-Node Cluster
Communications
FIGURE 17.12: Machine learning security best practices
Secure Your ML Environment

To secure your ML environment, follow these best practices:
➤➤ Encryption: Use encryption to secure your data at transit and at rest.
➤➤ Active monitoring: Monitor your ML environment for suspicious activity.
➤➤ Incident response: Have a plan to respond to security incidents quickly.
➤➤ Access controls: Break down your ML workloads by organizational units and environments so that you
can provide appropriate access controls.
➤➤ Environment access: Limit infrastructure access to administrators based on the dev, staging, and produc-
tion environment.
➤➤ Container security: Secure your model training and containers.
➤➤ Jupyter Notebook access: Implement a restricted Jupyter Notebook access policy through a centralized
management system and enable self-provisioning.
Enable Encryption of Inter-Node Cluster Communications

When using distributed training, you need to enable data transfer encryption between the nodes for sensitive ML
operations. You can do this by enabling inter-node encryption in Amazon SageMaker or enabling encryption in
transit when you’re using EMR.
Protect Against Data Poisoning Threats

Data poisoning is a serious threat where attackers try to gain access to training data and inject harmful data to
influence the model outputs. Here are some steps you can take:
➤➤ Anomaly detection: Use anomaly detection to check for tampering with training data.
➤➤ Change tracking: Keep track of all the changes made to the training data and the models and maintain
a baseline.
➤➤ Diversify data sources: Use more than one data source to train your models so the impact of compro-
mise to a single data source is minimized.
➤➤ Audit and oversight: Have an audit process in place to check for unusual changes and ensure alignment
with corporate security.
Building Robust and Trustworthy Models

You can enhance model robustness and trustworthiness by following these recommendations shown in
Figure 17.13.
Feature
Consistency
• Use the same feature engineering code in training
and inference environments.
• Use only trusted data sources for training.
• Before training, check the quality of the data for
outliers or incorrect labels.
Relevant Data
CI/CD/CT Building Robust • Use a validation dataset
Automation and representative of the data that
• Use the CI/CD/CT pipeline to Trustworthy the model will use in production.
automate the process of model Models • Check for underlying shifts in the
building, testing, and deploying the patterns and distributions in the
models. training data.
• Use version control system to track • Be on the lookout for drift in data
the changes in the model and data. and the model.
FIGURE 17.13: Reliability best practices
Ensure Feature Consistency Across Training and Inference

Here are some best practices to ensure feature consistency:
➤➤ Unified feature engineering: Use the same feature engineering code in training and inference
environments.
➤➤ Trusted data sources: Use only trusted data sources for training.
➤➤ Audit trails: Implement a robust audit process to replay the activity based on who did it and at
what time.
➤➤ Data quality assurance: Before training, check the quality of the data for outliers or incorrect labels.
Ensure Model Validation with Relevant Data

Here are some best practices for validating data:
➤➤ Representative validation data: Use a validation dataset representative of the data the model that will be
used in production.
➤➤ Detect data shifts: Check for underlying shifts in the patterns and distributions in the training data,
which might affect model reliability.
➤➤ Drift monitoring: Be on the lookout for drift in data and the model and address them appropriately to
ensure accuracy.
➤➤ Automated validation checks: Implement automated validation checks using relevant and updated data
to ensure the model works as expected.
Enable ML Automation and Traceability

Here are some best practices to implement automation and traceability in the ML process:
➤➤ Implement CI/CD/CT pipeline: Use the CI/CD/CT pipeline to automate the process of model building,
testing, and deploying the models.
➤➤ Version control and traceability: Combining it with a version control system to track the changes in the
model and data will ensure the traceability and reproducibility of the model.
Ensuring Optimal Performance and Efficiency

The following sections discuss some of the performance-related best practices.
Optimize Training and Inference Instances

Ensure you’re using appropriate instance types for training and inference purposes for optimal performance and
cost-efficiency.
TIP Using SageMaker Debugger, you can capture real-time training metrics that you can use
to optimize performance.
Explore Multiple Algorithms and Structures to Improve Performance

This is to reiterate what was already discussed: to explore different algorithms to best fit the problem, hyperpa-
rameters for better performance, and preprocessing data techniques to improve performance continuously.
Establish a Model Performance Evaluation Pipeline

Having a model performance evaluation pipeline will help you to assess performance continuously, detect any
deterioration in performance proactively, and send automated alerts so that you can address them immediately.
Establish Feature Statistics

Having feature statistics will help you identify the distribution of the features and improve performance by choos-
ing the right algorithms accordingly.
Conduct Performance Tradeoff Analysis

Conduct a performance trade-off analysis to ensure that you’re using your resources efficiently and yet not com-
promising on the performance.
Utilizing Cost Optimization Best Practices

The following sections discuss some of the steps you can take to optimize the cost of your ML implementation
(see Figure 17.14).
er Instance Size
Use Hyperparameter Techniques et
m
Fine-tune settings for best ra g Start small, then scale up
pa in
results
p er un Instance
y T
H Size
Checkpointing
Use Warm Starting and
Checkpointing Select Local Training
Lo inin
Reuse previous knowledge Experiment locally before
Tr
ca g
a
and save states regularly
l
moving to the cloud
Cost
Optimization
Best Practices
Choose an Optimal ML
Re Us
so ag
ML Framework
Stop Resources Not in Use Framework
ur e
ce
Pause unused resources to save

s
Pick the right tool for the job
costs
L
M
Distributed t ed g
a n
Training m ni
to rai Use Automated ML Training
Use Distributed Training Au T
Consistent, automated, and
Reduce training time with parallel repeatable training
processing
FIGURE 17.14: Cost optimization best practices
Select Optimal Instance Size

Choosing the correct type of instances for training and inference will help you save costs. Use debugging capa-
bilities to assess the resource requirements during training. Simple models may not train fast enough on larger
instances; hence, you can start with small instances and then scale them up gradually. SageMaker Experiments
can help here.
Select Local Training

For training small-scale ML workloads, you can resort to local training and experiment with different algorithms
first before moving on to the cloud.
HANDS-ON EXERCISE: SELECTING OPTIMAL INSTANCE SIZE AND

LOCAL TRAINING
Goal: Understand how to choose the correct instance size for a given problem and then
experiment with local training before scaling.
Start Small: Begin by training a simple Data scientist Trained model on a

model (e.g., linear regression) on a small instance, initial
small dataset using a small EC2 instance performance metrics
or even your local machine.
Scale Up: Gradually increase the Data scientist Performance report including
instance size, observing the training training time and metrics at
time and performance metrics. Note various instance sizes
down the changes.
Analyze: Compare the results to Data scientist Optimal instance size report,
identify the optimal balance between cost-benefit analysis
instance size and performance.
Local to Cloud Transition: Experiment Data scientist Comparison report of

with training the model locally first, then local versus cloud training,
transitioning to the cloud for further efficiency analysis, local-to-
tuning. Compare costs and efficiency. cloud transition plan
Choose an Optimal ML Framework

Choosing the proper ML framework is essential to save costs depending on your use case. SageMaker Experi-
ments can help you choose the right one. You can also choose deep learning AMIs and containers optimized for
specific ML frameworks. Some come with pre-installed TensorFlow, PyTorch, Apache MXNet, Chainer, Gluon,
Horovod, and Keras.
Use Automated ML Training

Resorting to automated ML training can help you optimize costs because they are built using a consistent, auto-
mated, and repeatable manner. They can choose the best algorithm, test different solutions quickly, and identify
the best parameters.
HANDS-ON EXERCISE: AUTOMATE MODEL TRAINING AND HYPERPARAMETER

TUNING WITH SAGEMAKER
Goal: Use AWS SageMaker to automate the training process and optimize hyperparameters.
Set Up SageMaker Configure a Data scientist Configured SageMaker

SageMaker experiment using a experiment,
predefined algorithm suitable for algorithm selection
your problem (e.g., XGBoost for
regression).
Automate Training Implement Data scientist Automated training setup,

automated training by utilizing initial model output
SageMaker’s automatic model
tuning feature.
continues
continued
Tune the Hyperparameters Data scientist Range of hyperparameters,

Define a range for hyperparameters optimal hyperparameter
and let SageMaker find the optimal values, comparison report of
values. Observe how the system tests different combinations
various combinations and settles on
the best one.
Analyze Results Data Scientist Report on best models,

Review the best models, their analysis of how they were
parameters, and how they were chosen, comparison with
chosen. Compare this with manual manual tuning (if applicable),
tuning if desired. final model
Use Distributed Training

Distributed training can also help reduce the training time, provided the algorithm allows it. The distributed data-
parallel library in Amazon SageMaker can break down the model across multiple GPUs and speed up the process.
Stop Resources Not In Use

It may be an obvious thing, but it is essential to stop resources that are not in use. It includes the Jupyter net-
works that you are using to run small models. Remember to stop the computing and storage services as well. You
can use an AWS CloudWatch Billing Alarms when the usage exceeds the threshold. Use the SageMaker lifecycle
configuration to start and stop Jupyter Notebooks when idle.
Use Warm Starting and Hyperparameter Checkpointing

Warm starting can help you to avoid the need to train a model from scratch. Instead, you can load from a previ-
ously tuned model along with its hyperparameters and reuse the knowledge gained from it.
Hyperparameter checkpoints are snapshots of a model’s state during training, including the model’s parameters,
optimizer state, and hyperparameters.
You can save your model’s state regularly, such as after an epoch or every few epochs. Restarting from check-
points can help you save time and cost.
Use Hyperparameter Optimization Techniques

Use hyperparameter optimization techniques to speed up the process of tuning your hyperparameters. For exam-
ple, SageMaker automatic model tuning runs multiple training jobs on the dataset, just as the best model uses the
hyperparameter ranges you specify.
WORKBOOK TEMPLATE: MODEL TRAINING AND EVALUATION SHEET

Download the “Model Training and Evaluation” template from the download section of the book (www.wiley
.com/go/EnterpriseAIintheCloud). Use this template to track the metrics and methods used during the
training and evaluation phases. It will help you to ensure that the model meets the required quality standards
before it is deployed.
SUMMARY
In this chapter, you began a comprehensive journey through training, tuning, and evaluating machine learn-
ing models. You examined the various stages, including model building, model training, model tuning, model
validation, and model evaluation. You also learned about some best practices in areas such as security, reliability,
performance, and cost management.
This chapter serves as a foundational guide for data scientists, machine learning engineers, and leaders to con-
tinue to explore machine learning and the model-building process in particular. As we close this chapter, it is
essential to remember that this is a continuous process, and the challenges and opportunities will continue.
In the next chapter, let’s discuss how to deploy models in production to ensure the models function optimally and
responsibly.
REVIEW QUESTIONS
1. What is one of the benefits of building models in the cloud?
A. Scalability
B. Security
C. Privacy
2. What is the purpose of model building?
A. To build a working model that can be used to make predictions on new data
B. To train the model
C. To tune the model
D. To validate the model
3. ______ is the primary step in machine learning that results in a working model that can be validated,
tested, and deployed.
A. Model training
B. Model building
C. Model tuning
D. Model evaluation
4. What is the difference between the model build process and the model training process?
A. The model build process focuses on discovering patterns in the training dataset and selecting the
best-performing model, while the model training process focuses on adjusting the model’s param-
eters to improve its performance.
B. The model build process focuses on adjusting the model’s parameters to improve its performance,
while the model training process focuses on discovering patterns in the training dataset and select-
ing the best-performing model.
C. The model build process and the model training process are the same.
D. None of the above.
5. What is data parallelism?

A. Using more than one computer or device to train the model
B. Splitting the model into smaller parts to predict different features and assigning each part to differ-
ent devices or computers
C. Combining models to form the final model
6. What is model parallelism?
A. Using more than one computer or device to train the model
B. Splitting the model into smaller parts to predict different features and assigning each part to differ-
ent devices or computers
C. Combining models to form the final model
7. What is the role of hyperparameters in machine learning models?
A. They guide the model learning process.
B. They are automatically learned from data.
C. They control the loss function.
D. They are the trainable parameters of the model.
8. How does a high learning rate affect model convergence in gradient descent?
A. It slows down convergence.
B. It stabilizes the convergence.
C. It speeds up convergence but might cause instability.
D. It does not affect convergence.
9. What is the purpose of regularization in a machine learning model?
A. It helps to speed up model training.
B. It adds a penalty to the loss function to reduce model complexity.
C. It increases the learning rate of the model.
D. It increases the number of epochs in model training.
10. What is the focus of the data prepare pipeline?
A. To prepare a clean and consistent dataset ready for feature engineering
B. To store, retrieve, and copy features into and from an online or offline feature repository for use
in a model
C. To validate and test the model before releasing it into production in various environments
D. To deploy and monitor the model’s real-time performance in production.
11. Which phase involves tweaking the data, hyperparameters, and model parameters based on evalua-
tion results?
A. Model training
B. Model selection
C. Model fine-tuning
D. Model validation
12. Which of the following steps is not involved in the model evaluation process?
A. Model selection
B. Model validation
Answer Key ❘ 341
C. Model fine-tuning
D. Model training
13. What is the benefit of implementing infrastructure as code (IaC)?
A. Consistency across different infrastructures
B. Traceability of different data and model artifacts
C. Both A and B
D. Neither A nor B
14. What is one of the risks associated with transfer learning?
A. The inherited weights from the pretrained model may not produce the correct predictions.
B. The fine-tuned model may take longer to train than a new model.
C. The pre-trained model might not have been trained on relevant data.
D. All of the above.
15. What is the purpose of conducting a performance trade-off analysis in machine learning?
A. To find a balance between model complexity and accuracy
B. To ensure optimal utilization of resources without compromising performance
C. To compare the performance of different models
D. Both A and B
ANSWER KEY
1. A 6. B 11. C
2. A 7. A 12. D
3. A 8. C 13. A
4. A 9. B 14. D
5. A 10. A 15. D
PART VII
Deploying and Monitoring
Models
Hooray! It is launch time. Part VII guides you through the process of deploying the model into produc-
tion for consumption. I also discuss the nuances of monitoring, securing, and governing models so they are
working smoothly, safely, and securely.
18
Deploying Your Models Into
Production
Do not wait to strike till the iron is hot but make it hot by striking.
—William Butler Yeats
Having navigated through data processing, algorithm selection, model training and fine-tuning, you are
now on the verge of an important step: model deployment. This is an essential chapter for data scientists,
ML engineers, IT professionals, and organizational leaders involved in deploying models into production
using the cloud.
The true value of an AI model is not just in its design or accuracy but in its real-world use. This chapter
dives deep into the nuances of deploying your model, from understanding the challenges in model deploy-
ment, monitoring, and governance to deciding between real-time and batch inferences.
Model deployment isn’t just about pushing a model live. It involves strategic decisions, systematic pro-
cesses, and synchronized architecture. Keep in mind that you will not just be launching models into pro-
duction; you will also be ensuring that the models are functioning optimally and responsibly within your
larger AI ecosystem. See Figure 18.1.
The focus of this chapter, though, is to have a successfully deployed model that is making the impact that it
was meant to have during the design process to achieve continuous innovation and growth.
STANDARDIZING MODEL DEPLOYMENT, MONITORING, AND

GOVERNANCE
Here are some considerations for your model deployment, monitoring, and governance process:
➤➤ Automate deployment: You should automate the deployment process and make it repeat-
able and standardized. This will ensure that the model is deployed consistently across multiple
environments.
➤➤ Automated performance monitoring: You need to ensure that the model is performing well over a
period of time. The monitoring process needs to be automated so that you can proactively identify
and address any issues promptly. You need a robust monitoring system to measure the model’s
performance metrics, such as accuracy, precision, recall, latency, and throughput.
346 ❘ CHAPTER 18 Deploying Your Models Into Production
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Govern Your
Your AI
Team Deploy and Monitor Models
journey
09
• Deploy Your Models into
ENTERPRISE AI Setup 04 Production
Evolve and
Mature AI
JOURNEY MAP Infrastructure • Monitor and Secure Models
and Manage
Operations
• Govern Models for Bias and Ethics

AI
Deploy and 05 Seamlessly transition AI/ML models from
Monitor development to live environments, ensuring
Models optimal deployment strategies, real-time
07 responsiveness, and consistent configuration
06 for efficient production operations
FIGURE 18.1: DEPLOY AND MONITOR MODELS: Deploy your models into production.
➤➤ Ethical use and responsibility: You need to ensure that the model is used responsibly and ethically in
a production environment without discriminating against its users. To enforce governance, you must
define policies and procedures integrated into your organization’s overall governance framework and
establish systems to detect bias proactively and address them in real time.
TIP Adopt deployment best practices such as automatically deploying through CI/CD pipe-
lines, deployment scripts, version control, continuous performance monitoring, ethical use,
proactive bias detection, and establishing a model governance framework.
Challenges in Model Deployment Monitoring and Governance

There are several challenges when deploying models and ensuring they are implemented ethically and responsibly
in production, discussed next.
Model Deployment Challenges

Consider these deployment challenges:
➤➤ Scalability: Models may not scale well enough to handle large amounts of data in production.
➤➤ Operational efficiency: Poor operational efficiency and data silos may result from a lack of integration
of the models with backend processes, systems, and data storage systems.
➤➤ Latency: High latency of the models can render them useless for practical use.
➤➤ Reproducibility: Over time, reproducing these models for retraining and redeployment becomes impossi-
ble because of numerous changes in the underlying hardware, software libraries, and other dependencies.
Deploying Your Models ❘ 347
Model Monitoring Challenges

Here are some monitoring-related challenges:
➤➤ Dynamic data: The world is constantly changing, and therefore, the data feeding the models also
changes, which can cause issues when monitoring the model performance.
➤➤ Model complexity: Most of these models are complex, and it becomes difficult to understand how they
work because of the black-box nature of the algorithms. It poses a challenge to enforcing governance of
the models because of the need for model explainability and interpretability.
➤➤ Talent shortage: There is not enough talent and resources to focus on model deployment and
monitoring.
➤➤ Unified framework: You may find that the methods adopted to develop and deploy models are not
standardized across teams.
➤➤ Data lineage tools: There is also a lack of transparency in the data quality and lineage.
Model Governance Challenges

Finally, these are governance challenges:
➤➤ Bias control: Since these models reinforce existing biases in the data, controlling biases is a challenge.
➤➤ Data privacy and regulation: Data privacy and security should be carefully protected during training and
prediction to comply with legal regulations such as GDPR, CCPA, and HIPAA.
➤➤ Industry regulations: Navigating regulations in industries such as healthcare or finance can be an issue.
➤➤ Model inventory: Models tend to increase rapidly across business units in large enterprises, and keeping
track of inventories becomes tough.
➤➤ Change logs: Keeping track of changes in AI models is not easy because you have to track the changes in
the code and the changes in data, hyperparameters, and models.
➤➤ Accountability: The question of who is accountable or responsible when the model fails is also tricky.
TIP Solving these challenges requires a mix of technology initiatives such as MLOps, pipe-
lines, explainability techniques, governance policies, data use policies, regulatory guidelines,
and culture.
DEPLOYING YOUR MODELS

This section dives into the model deployment process. It is often underestimated in the model development life-
cycle. However, it is the most crucial step from a user perspective because it makes the model available for user
consumption. The following sections review some best practices.
Pre-deployment Checklist
First, you must ensure the model is fully trained and tested. Second, ensure that you have fully optimized the
model using the best practices discussed in the previous chapter. Third, perform dry runs in a staging or sandbox
environment to uncover unforeseen issues.
Deployment Process Checklist

As shown in Figure 18.2, the deployment process involves the following:
➤➤ Features: Retrieve the features from the feature store.
➤➤ Artifacts and dependencies: Retrieve the model from the model registry and other artifacts required for
the model to run.
➤➤ Inference code container: Retrieve the inference code container, which contains the code required to
make the predictions against the model. The inference code container is typically retrieved from the con-
tainer registry using a docker pull command.
➤➤ Deployment: Once the features, the artifacts, and the inference code container are retrieved, the model is
deployed into the production environment.
➤➤ API access: Making it available as a Restful API using a web service framework like Flask or deploying
it in a cloud platform such as AWS SageMaker, GCP, or Azure Machine Learning.
➤➤ Application integration: Integrating it with other applications if necessary.
➤➤ Monitoring: Periodically review the model’s output with the actual values to ensure the model performs
properly using a manual or an automated governance process.
Feature Store
Fetch
Features
Container
Production Consuming
Inference Code Endpoint Application
Fetch Artifacts
(Models, Code)
Model Registry
FIGURE 18.2: Model deployment process
Choosing the Right Deployment Option

As part of your architecture design discussion, you need to decide which one of the model deployment options is
the best fit for you. As shown in Figure 18.3, you have four options: cloud deployment, on-premises deployment,
edge deployment, and hybrid deployment.
Cloud Deployment
Here are some considerations for cloud deployment:
➤➤ Scalability: Cloud is the best option if you have many users to serve and if the load fluctuates. It
allows you to process large amounts of data using a large computing capacity, and to store large
amounts of data.
Best for fluctuating loads, Preferred for security,

many users, comprehensive control, and data location
services large data, & preferences
computing capacity
01 02
On-premises
Cloud Deployment
deployment
CLOUD
DEPLOYMENT
OPTIONS
Edge Deployment Hybrid Deployment
03 04
Ideal for real-time processing, Combines flexibility and
high privacy, low bandwidth scalability of cloud with control
requirements, and low latency and security of on-prem
FIGURE 18.3: Model deployment options
➤➤ Service integration: The cloud providers also provide several services that make deploying and managing
your model easy. For example, you can leverage services for computing, storage, and networking.
➤➤ Data security: The cloud allows you to connect to sensitive data using robust network security protocols
offered by the cloud providers.
For these reasons, the cloud option is often the most popular option.
On-Premises Deployment
Some companies may resort to on-premises deployment of their models using their own servers, storage, and
networks in their data centers.
➤➤ Security requirement: It is preferred when security is a primary requirement, and you want to keep your
data in your own locations.
➤➤ Infrastructure control: It also gives more control over their infrastructure. However, you need to regu-
larly update your hardware and software, have robust backup and disaster recovery, and regularly moni-
tor for bottlenecks.
➤➤ Expensive: However, it can be more expensive and resource intensive to manage.
Edge Deployment
Here are some considerations for edge deployment:
➤➤ Model performance and latency: If your model performance and latency are important, then you would
resort to deploying the model at the edge where the data resides and action happens. Your edge devices
could be smartphones, IoT sensors, tablets, or industrial machines.
➤➤ Real-time processing: It allows you to process data in real time, provides greater privacy, and has fewer
bandwidth requirements.
➤➤ More complex: However, this can be the most complex deployment type because managing and updat-
ing models across multiple devices can take time and effort.
➤➤ Other considerations: Remember that you need to implement device management to manage software
updates, may encounter issues because of limited computing power, and may need to synchronize data
between central servers/cloud and devices.
NOTE AWS IoT Greengrass can run models in edge devices using models trained in the
cloud.
Hybrid Deployment
People use a hybrid deployment to get the best of both worlds, namely, cloud and on-prem deployments.
➤➤ Flexible and scalable: This approach can provide flexibility and scalability of cloud deployment.
➤➤ Greater control and security: This approach can provide greater control and security through an on-
prem deployment.
➤➤ More complex: It can be challenging to implement and manage.
With Amazon SageMaker, you can run the models on the cloud and at the edge. It has a compiler that can read
models from various frameworks and compile them into binary code to run the model in the target platform. It
also enables the “train models once and run anywhere” concept.
TIP Think long term and ensure that your deployment method suits your future growth
potential. Avoid vendor lock-in and factor in regulatory requirements.
Choosing an Appropriate Deployment Strategy

To reduce the risk involved in model deployment, there are several deployment strategies that you can adopt, as
shown in Figure 18.4. The following sections discuss them.
Blue/Green Deployment
Consider these characteristics of blue/green deployment.
➤➤ Environment consistency: In the blue/green deployment approach, you have two identical production
environments running in parallel. You will continue the live traffic on the blue production environment,
and the green environment will be used for the deployment. When a new version of the model is ready
for deployment, you will deploy it first in the green environment to test it. Once it has been tested and
found to be working fine, you will start directing the live traffic to the green environment. The blue envi-
ronment will now become the backup.
➤➤ Rapid rollback: The benefit of this approach is that you can roll back your changes if the model is not
working correctly. It is scalable and requires minimal downtime.
➤➤ Resource management: But the disadvantage is that it has additional infrastructure costs due to the need
to maintain two production environments in parallel. It is more complex to implement and needs addi-
tional resources.
Shadow Blue/Green
Deployment Deployment
Run old and new Test on green, switch
versions side-by-side for live traffic upon success.
testing. Real-life testing 3 1 Scalable with minimal
environment, scalable downtime
Canary
A/B Testing 3 2 Deployment
Split users randomly Release to a small
and serve old and new group initially. Monitor
versions. Allows data- and scale up traffic if
driven decisions successful
FIGURE 18.4: Choosing a deployment strategy
Canary Deployment
Consider these characteristics of canary deployment:
➤➤ Segmentation: You release the model to a small group of users.
➤➤ Continuous monitoring: It allows you to monitor the model’s performance closely, and once you find it
satisfactory, you can slowly ramp up the traffic to 100 percent.
➤➤ Live testing: The benefit of this approach is that you can test the model in a live, fully controlled
environment.
➤➤ Rollback: It allows you to roll out a new model without impacting all the users and can roll back if
there is an issue with the new model.
➤➤ Small sample size: The disadvantage is that you are testing it against a small sample size, and not all
errors may be caught. It won’t be easy to target a suitable set of users and assess the full impact of the
new model.
A/B Testing
Here are some characteristics of A/B testing:
➤➤ Controlled variables: You split the users into two groups in a random fashion and then test them with
the old and new versions of the model to compare their performance. The first group is called the con-
trol group and is served the old model, and the other group is called the treatment group and is served
the new model.
➤➤ Statistical relevance: The benefit of this approach is that it allows you to make data-driven decisions. It
allows you to test different features and configurations.
➤➤ Planning: It takes a lot of planning to avoid biases. It will need to run for a substantial amount of time
or need a substantial amount of traffic to reach statistical significance.
➤➤ Targeting: It is also difficult to target the right set of users and therefore understand the full impact of
the model.
NOTE AB testing is similar to canary deployment. However, it requires a more significant

number of users and will take longer, typically days or weeks, to reach a conclusion.
Shadow Deployment
With shadow deployment, you have the old and new versions running side by side.
➤➤ Data synchronization: While the old version serves the users, the new version is used purely for testing
and analysis. Both versions take in the same input data.
➤➤ Real-life testing: The advantage of this method is that you can do real-life testing in a real-life environ-
ment, which involves minimal downtime, is scalable, and can be rolled back in the case of an issue.
➤➤ Additional complexity: The cost of maintaining two models may require additional resources and are
more complex to implement.
TIP Regardless of the deployment strategy used, always ensure that the infrastructure is
robust, scalable, and have well-defined rollback plans.
Choosing Between Real-Time and Batch Inference

Another critical decision you must make is to choose between real-time and batch inferences. Your choice will
depend on the business need and the use case.
Your decision to choose between real-time and batch inference will depend on balancing immediacy, cost, and
technical feasibility.
CONSIDERATION REAL-T IME INFERENCE BATCH INFERENCE
Business requirement Requires instant feedback (e.g., Result isn’t immediately necessary (e.g.,
fraud detection). monthly reports).
Infrastructure and cost Needs robust infrastructure for Tuned for large-scale processing.
24/7 uptime and scalability. Can Generally, more cost-effective.
be more expensive.
Latency Requires quick decisions (e.g., Can tolerate delayed results (e.g., daily
algorithmic trading). analytics).
Data consistency Handles sporadic or inconsistent Deals with consistent, structured datasets.
streaming data.
Error handling Needs instant alerts and quick Requires checkpoints and logging for
correction mechanisms. post-facto error detection.
Resource allocation High availability and redundancy Needs significant storage and compute
are critical. during processing.
CONSIDERATION REAL-T IME INFERENCE BATCH INFERENCE
Reliability Should have failover mechanisms. Ensure reliability based on processing

frequency.
Hybrid approaches Possible to combine for Can be combined with real-time for
immediate predictions with comprehensive insights.
deeper analytics.
TIP Batch inference is not valuable for use cases where real-time predictions are essential.
Implementing an Inference Pipeline

When you want to make predictions based on incoming data in real time, you should explore the option of using
inference pipelines. For example, an inference pipeline helps predict customer churn in real time or make real-time
product recommendations on your website.
As shown in Figure 18.5, the inference pipeline can consist of several steps, such as data preprocessing, feature
extraction, model inference, and post-processing. It takes in data in real time and converts them into predictions
in real time. It can consist of any ML framework, built-in algorithm, or custom containers and works on both
real-time and batch inference requests.
Data Feature Post-

Collection Extraction Processing
Collect data from Extract relevant Convert predictions to
various sources features from the suitable format for
like IoT sensors processed data downstream
applications
Step 01 Step 03 Step 05
Step 02 Step 04
Data Model
Preprocessing Inference
Cleanse and Make predictions
format data for using the model
model input
FIGURE 18.5: Different steps in an inference pipeline
Data Collection
During this step, the data from various sources, such as IoT sensors, is collected and stored in data storage. How-
ever, this data is not ready yet for use in the model.
TIP Leverage reliable and scalable storage systems.

Data Preprocessing
During the data preprocessing step, the incoming data is cleansed and converted into a suitable format to be fed
into the model. The data may need to be resized, normalized, standardized, or augmented. It is an essential step
because it will determine the accuracy with which the model predicts, as it depends on the accuracy and consist-
ency of the input data.
TIP Automate the cleansing process.
Feature Extraction
The pipeline extracts the relevant features from the input dataset during this step. Again, this is an important step
that ensures accuracy in the model performance. For example, in the case of image processing, this may involve
edge detection, object segmentation, and other techniques to extract meaningful information from the raw image.
TIP Leverage domain-specific expertise and refine features periodically.
Model Inference
This is the most critical part of the inference pipeline and is the core because this is where the actual prediction
happens. This step may provide meaningful insights to drive subsequent actions.
TIP Use optimized models for faster inference and retrain models as necessary.
Post-processing
The predictions are converted into the format for other downstream applications during this step.
TIP Standardize output formats and ensure downstream application compatibility.
You can see that implementing an inference pipeline can improve efficiency, latency, and accuracy in real time.
However, it can be a complex process that requires careful planning, design, and implementation.
NOTE Amazon SageMaker Inference Pipeline has 2 to 15 containers to process inference

requests.
EXERCISE: IMPLEMENTING AN INFERENCE PIPELINE
Goal: Implement an end-to-end inference pipeline to predict customer churn.

Synchronizing Architecture and Configuration Across Environments ❘ 355
Collect and preprocess data: Data collection team Preprocessed data ready for
Gather data from various feature extraction
sources and preprocess it to
make it suitable for the model
(e.g., resizing, normalizing).
Extract features: Utilize Feature Extracted features ready for

statistical or machine learning engineering team model inference
techniques to extract
relevant features from the
preprocessed data.
Implement model Inference: Model Predictions or

Employ a pretrained model to implementation team inferences made by the
make predictions or inferences pretrained model
based on the extracted features.
Add post-processing steps: Post-processing team Formatted predictions

Format the predictions or ready for consumption
inferences so they can be used by other applications
by downstream applications or services
(e.g., convert to required
data formats).
Test the pipeline: Evaluate Testing and QA team A fully functioning inference
the entire inference pipeline pipeline validated with
using real-time data to ensure it real-time data, including a
functions as expected. testing report
SYNCHRONIZING ARCHITECTURE AND CONFIGURATION ACROSS

ENVIRONMENTS
You must ensure that the model performs consistently across multiple environments, such as development and
production.
NOTE Sometimes even a minor change in the architecture configuration can result in differ-
ent inferences in different environments.
You must also ensure the model shows the same accuracy range across development, staging, and production
environments. Any model that fails to show this level of accuracy should not be deployed in production. It
becomes imperative as you scale your operations across multiple geographies and product lines.
Ensuring Consistency in the Architecture

You can leverage infrastructure as code to ensure consistency in the architecture for both AWS and third-party
components. Using AWS CloudFormation, you can manage the infrastructure as one unit and deploy them across
multiple environments. It allows you to create, update, and delete an entire stack as one unit.
Ensuring Identical Performance Across Training and Production

Synchronizing the architecture across environments alone isn’t enough. You must also ensure the model performs
identically across the training and production environments. You can achieve this validation through the use of
AWS SageMaker Model Monitor. If you find a difference in the model’s accuracy between production and train-
ing environments, you can start troubleshooting the cause and try to address the issue.
Looking for Bias in Training and Production

SageMaker Clarify can also detect bias either during training or in production. SageMaker Clarify can help you
understand how these models predict using a feature attribution approach. It also keeps monitoring the model in
production and looks for any bias or feature attribution drift.
Generating Governance Reports

You can use the governance reports generated by SageMaker Clarify to share with your stakeholders, such as
audit teams, regulators, risk and compliance teams, and other business stakeholders.
EXERCISE: DETECT BIAS USING SAGEMAKER CLARIFY
Goal: Monitor the model for bias during training.
Identify bias metrics: Define Data science team List of selected bias metrics
and select the specific metrics tailored to the specific use
that will be used to measure case and data.
and identify bias in the
training data.
Configure SageMaker Clarify: Machine Configured SageMaker Clarify

Set up SageMaker Clarify by learning engineers environment for detecting bias.
specifying the configuration
settings such as preprocessing
and post-processing scripts.
Run bias analysis: Execute the Machine Raw output of bias analysis,
bias detection process using learning engineers including identified biases and
SageMaker Clarify, analyzing related statistical measures.
the training data against the
defined metrics.
MLOps Automation: Implementing CI/CD for Models ❘ 357
Analyze bias results: Interpret Data science and In-depth analysis report
the results of the bias analysis, analytics team highlighting the findings, causes,
understanding the underlying and potential remedies for
causes and potential impact on detected biases.
model performance.
Implement remediation Data science team Revised training data and model
strategies: Based on the training approach, minimizing
analysis, apply mitigation the identified biases.
strategies to reduce or
eliminate the identified biases
in the training data.
Monitor bias in production: Monitoring and Continuous monitoring setup,

Implement continuous analytics team with regular reports on identified
monitoring of the deployed biases in the production model.
model to detect any bias that
might emerge in production.
Generate and share reports: Compliance and Comprehensive bias reports,

Prepare comprehensive reporting team ready for distribution to internal
reports on the detected and external stakeholders,
biases and share them with ensuring alignment with
stakeholders such as audit regulations and policies.
teams, regulators, etc.
MLOps AUTOMATION: IMPLEMENTING CI/CD FOR MODELS

The previous section talked about using an inference pipeline to automate the process of converting input data
into a prediction for subsequent follow-up action.
Implementing CI/CD for models is a similar automation concept but is focused on automating the process of
deploying models into production. It involves automating the process of training, testing, and deployment of
the model into production, which is a complex process given the nature of data, the model parameters, and the
configuration versions (see Figure 18.6).
Ensure you follow some best practices such as maintaining model versions, automated testing for data preprocess-
ing, feature extraction, validation testing; also ensure environmental consistency across dev, staging, and produc-
tion, and many other practices outlined in this part of the book.
TIP Implementing CI/CD can introduce a lot of efficiency and scale, especially when
managing large enterprise AI implementations with many models to manage.
Model Development Continuous Integration Continuous Deployment

• Code repository (e.g., Git) • Automatic code pull • Deployment to
• Model versioning (upon push to the development
repository) environment
• Data preprocessing scripts
• Data preprocessing and • Deployment to staging
• Feature extraction scripts
feature extraction environment
• Model training • Deployment to
• Validation testing production environment
1 2 3
FIGURE 18.6: Implementing CI/CD for models
The following list covers some of the CI/CD tools:

➤➤ Amazon SageMaker is a fully managed service from AWS that provides end-to-end lifecycle implementa-
tion support, including continuous integration and continuous deployment.
➤➤ Jenkins is an open-source automation server that can be used in model deployment to run the scripts
necessary to train, test, and deploy models.
➤➤ GitLab CI/CD is another software automation tool that can automate training, testing, and deployment
and is particularly useful for version control of the model artifacts.
➤➤ Azure DevOps is a set of development tools that includes Azure Pipelines, which can be used to create
CI/CD workflows for models.
➤➤ Kubeflow and MLflow are open-source platforms for orchestrating machine learning workflows.
➤➤ TensorFlow Extended (TFX), Azure Machine Learning, and Google Cloud AI Platform are end-to-end
platforms to build, train, and deploy machine learning models.
EXERCISE: IMPLEMENTING CI/CD FOR MODELS USING JENKINS
Goal: Automate the process of training, testing, and deploying a model using Jenkins.
TASK NAME TASK DESCRIPTION TASK OWNERS
Set up a Install and configure Jenkins to DevOps team

Jenkins Pipeline create a continuous integration and
deployment pipeline.
Write Scripts Develop scripts for training, testing, and Development and
deploying the model as part of the CI/ data science team
CD process.
Configure Integrate the written scripts into the Jenkins DevOps and
the Pipeline pipeline and set triggers, parameters, etc. configuration team
TASK NAME TASK DESCRIPTION TASK OWNERS
Monitor and Continuously monitor the pipeline, detect Monitoring

Ensure Success any issues, and ensure the successful and QA team
deployment of the model.
WORKBOOK TEMPLATE: MODEL DEPLOYMENT PLAN

Download the “Model Deployment Plan” template from the download section of the book (www.wiley.com/
go/EnterpriseAIintheCloud). Use this template to plan the model deployment process and ensure that the
model performs well when it is integrated into the business processes.
SUMMARY
This chapter discussed the intricate process of deploying models. It started with an introduction to model
deployment followed by considerations and challenges in model deployment. It then discussed choosing the right
deployment option from cloud to on-premises, edge, and hybrid setups, and choosing an appropriate deployment
strategy, including blue/green, canary, A/B testing, and shadow deployments. It addressed the choice between real-
time and batch inference and discussed the pros and cons of each approach. It also discussed other deployment
best practices such as synchronizing architecture and configuration across environments, striving for identical
performance between training and production, and generating insightful governance reports.
In the next chapter, let’s discuss how to monitor models for data quality, data drift, concept drift, model perfor-
mance, bias drift, and feature attribution drift.
REVIEW QUESTIONS
1. Which of the following is NOT a common challenge in model deployment?
A. Models may not scale well to handle large data volumes.
B. Lack of integration with backend processes can create data silos.
C. Models become more explainable and interpretable over time.
D. Reproducing models for retraining and redeployment can become difficult due to changes in
dependencies.
2. Which deployment option can be the most complex because managing and updating models across mul-
tiple devices can take time and effort?
A. Cloud deployment
B. On-premises deployment
C. Edge deployment
D. Hybrid deployment
3. Which of the following is a disadvantage of real-time inference?
A. It is not scalable for large datasets.
B. It requires more resources and is expensive to implement.
C. It is not valuable for use cases where real-time predictions are essential.
D. Predictions can’t be stored in a database for later access.
4. Which step in an inference pipeline converts predictions into a format suitable for other downstream
applications?
A. Data collection
B. Data preprocessing
C. Feature extraction
D. Post-processing
5. What do TensorFlow Extended (TFX), Azure Machine Learning, and Google Cloud AI Platform have
in common?
A. They are open-source platforms.
B. They are end-to-end platforms to build, train, and deploy machine learning models.
C. They are specifically designed for version control.
D. They are data collection tools used in the data preprocessing stage of model development.
ANSWER KEY
1. C 3. B 5. B
2. C 4. D
19
Monitoring Models
What gets measured gets managed.
—Peter Drucker
This chapter deals with the critical components of monitoring, including making informed decisions
around real-time and batch monitoring. It delves into checking the health of model endpoints, selecting
appropriate performance metrics, and ensuring model freshness. To ensure the AI systems remain agile and
responsive, you learn how to review and update features, automate endpoint changes, and scale on demand
(see Figure 19.1).
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Govern Your
Your AI
journey
09
• Deploy Your Models into Production
ENTERPRISE AI Setup 04 • Monitor and Secure Models
Evolve and
Mature AI
JOURNEY MAP Infrastructure • Govern Models for Bias and Ethics
and Manage
Operations

AI
Deploy and 05 Implement robust real-time monitoring
Monitor mechanisms for AI/ML models, ensuring
Models optimal performance, health, and adaptability
07 in live environments, while facilitating timely
06 updates and scalability
FIGURE 19.1: Monitoring models

362 ❘ CHAPTER 19 Monitoring Models
MONITORING MODELS
Monitoring models is an essential aspect of MLOps to ensure that the model performs well and that the data
used is high quality.
Importance of Monitoring Models in Production

Models are trained based on historical data, and we expect them to work in the same fashion on new data. How-
ever, in the real world, things are constantly changing, and hence there is a need to monitor the model and keep it
updated. Here are some of the reasons to monitor the models:
➤➤ Data drift can happen due to changes in input data with respect to training data.
➤➤ Concept drift can happen due to changes in model performance.
➤➤ Detect unfair bias introduced by the model.
➤➤ Changing business needs may require a model to be updated. For example, a retailer may introduce new
product lines that require a product recommendation model to be retrained.
➤➤ New regulations require the model to be retrained.
➤➤ Model performance can decay over a period of time due to changes in the underlying data distribution.
➤➤ Ensure the model is robust against all use cases in the real world.
➤➤ Feedback from the model performance evaluation process requires retraining the model or even choos-
ing a new model.
➤➤ There are regulatory and compliance requirements to monitor models.
➤➤ New technological advancements such as the advent of a new or improved algorithm may require you
to update models.
TIP It is important to monitor the models for any changes and update them with new data,
algorithms, or any other changes required.
Challenges Faced When Monitoring Models

Monitoring models in production can be tricky because of a number of factors, some of which are listed here:
➤➤ Data drift can be hard to determine, and it needs evaluating data distributions.
➤➤ Degradation can happen slowly and go unnoticed.
➤➤ When you have multiple models in production, and especially when they interact with each other, it can
become complicated.
➤➤ Edge case scenarios that occur infrequently can be missed.
➤➤ Black-box models can be difficult to explain.
➤➤ Monitoring involves overheads, and troubleshooting can be time-consuming.
Addressing these challenges requires a mix of monitoring strategies, tools, infrastructure, and analysis techniques,
covered in this chapter.
Key Strategies for Monitoring ML Models ❘ 363
KEY STRATEGIES FOR MONITORING ML MODELS

You need to employ a combination of strategies to effectively implement model observability to track a model all
the way from data to decision-making to performance. I review them here in brief and then cover them in greater
detail in the subsequent sections of this chapter.
➤➤ Detecting and addressing data drift: Monitoring statistics such as feature distribution, missing values,
data quality issues, and so on
➤➤ Detecting and addressing concept drift: Detecting if relationships between data elements are changing
➤➤ Monitoring bias drift: Detecting unequal performance for various user segments
➤➤ Feature attribution drift: Detecting drift in the feature attribution values resulting from a drift in the
distribution of the live data in production
➤➤ Tracking key model performance metrics: Tracking key model performance metrics like accuracy, F1,
recall, and so on
➤➤ Model explainability: Explaining model behavior using interpretability techniques
➤➤ Ground truth testing: Evaluating predictions against true outcomes to validate accuracy
➤➤ Deployment testing: A/B testing, shadow deployments, canary deployments, and building validation into
CI/CD pipelines
TIP Do a phased rollout by starting with shadow deployments, and then move to canary
before a full-scale deployment.
The following sections discuss these monitoring strategies in more detail.
Detecting and Addressing Data Drift

For a model to perform well, the data must be complete and accurate. You need to differentiate between data
quality and data drift because both impact model performance. When the data is of poor quality, it could mean
two things.
➤➤ The data itself is bad, out of range, or erratic.
➤➤ The feature engineering code is not working correctly.
While monitoring data quality involves ensuring that the data used to train and deploy machine learning models
is of high quality, monitoring for data drift involves detecting changes in the distribution of data used to train and
deploy machine learning models.
Data drift can be caused by changes in how data is collected, processed, or stored. Poor data quality can result
from errors during data entry, data corruption, or data inconsistencies. While poor data quality can lead to inac-
curate results, data drift can lead to decreased model accuracy.
An example of data drift is a model that predicts if a customer will default on a loan. It may have been trained
five years ago based on customers with good credit. But over time, the model performance may decrease because
customers with poor credit ratings have started to apply in more significant numbers, and your data distribution
is no longer the same.
TIP You need to clearly distinguish between data drift and data quality issues because the
resolution is different. For data drift issues, you would need to retrain the model using new
data, while for data quality, you would need to clean the data or remove outliers. You can
use unsupervised learning algorithms to train the model so that it can adapt to the changing
data by changing the hidden structures dynamically.
There are several methods to detect data drift, such as the following (see Figure 19.2):
➤➤ Population stability index (PSI), which compares the expected probability distribution in the present
dataset with that of the training dataset
➤➤ Kolmogorov-Smirnov test
➤➤ Kullback-Leibler divergence
Model Update Pipeline Monitoring Data Quality

Implement an update Ensure data is complete,
pipeline accurate and well-
engineered
Model Explainability Monitoring Data Drift

Ensure the transparency Detect changes in data
and interpretability of Key distribution
model predictions Components
of
Monitoring
Feature Attribution Drift Monitoring Concept Drift
Monitor changes in the Identify and address
importance of features in changes in relationships
predictions within the dataset
Monitoring Bias Drift Monitoring Model Performance

Measure parity and Track model
equality metrics to performance metrics
identify any bias drift
FIGURE 19.2: Key strategies of monitoring
TIP Implement a comprehensive monitoring system that automates the monitoring process
to ensure models work accurately.
Detecting and Addressing Concept Drift

Using the same loan example, say the bank started offering a new type of loan with different characteristics, such
as different interest rates or repayment terms. In this case, the relationship between the various features in the
dataset to the default behavior would have changed, which is considered concept drift.
You can address concept drift using these steps:
1. Use neural networks to train the model to identify new relationships.
2. Try classification algorithms to capture the new relationships.
3. Use an ensemble of model instead of a single model to reduce the impact of concept drift.
Key Strategies for Monitoring ML Models ❘ 365
Here are some steps to detect concept drift:

1. Monitor the accuracy and performance of the model to check if the performance is deteriorating
over time.
2. Calculate the average confidence score of the model and check if it’s deteriorating.
3. Check the data distribution of the training dataset and the new dataset, and if there is a significant dif-
ference, then a concept drift may have occurred.
TIP Plan for periodic retraining of models when significant concept drift is detected.
Monitoring Bias Drift

Bias drift can happen when the distribution of the data changes. For example, in the case of a product recommen-
dation model, the model has been trained using data for high-income groups; it will perform poorly for lower-
income groups. The model will also show bias when the underlying data is also biased.
TIP To detect bias drift, you must measure metrics such as demographic parity, equal oppor-
tunity, and equal odds.
SageMaker Clarify can help detect bias in the models and is also integrated with SageMaker Model Monitor,
which can help detect data and concept drifts.
Watching for Feature Attribution Drift

Feature attribution drift refers to the change in the importance attributed to the feature toward a prediction.
If the importance of the feature changes, the model will no longer perform well. It’s likely to happen when the
model is trained with new features or when the data distribution changes.
TIP To monitor feature attribution drift, you must track metrics such as featuring impor-
tance score and feature correlation.
An example of feature attribution drift is for a customer churn model; it might have been that the number of
years the customer has been with the company may have been the most critical feature, but when a competitor
introduces a new product, the customers may churn. Their stay with the company will no longer be a signifi-
cant factor.
While it is true that implementing a comprehensive monitoring system can be very intensive, it can automate the
monitoring effort and save you a lot of manual effort and time to ensure the models are working accurately.
TIP SageMaker Model Monitor can help detect data and model drifts in real time by evalu-
ating the inputs and outputs and sending you alerts.
Model Explainability
Model explainability is about explaining how the model makes its predictions so that it can be trusted. It is essen-
tial as it helps you understand the root causes of a specific prediction and avoid biases. You need to know how
each feature contributes to the prediction and whether the prediction is in line with the existing domain knowl-
edge. It requires measuring metrics related to model quality, fairness, and drift.
TIP Where possible, opt for models that are inherently more interpretable. Leverage tools
like LIME or SHAP to understand complex model behaviors.
TRACKING KEY MODEL PERFORMANCE METRICS

The choice of model performance metrics depends on the type of problem that you are trying to solve.
Figure 19.3 is a brief look at some of those categorized by the type of problem.
Classification Metrics
The previous chapter discussed classification metrics, including accuracy, precision, recall, F1 score, AUC-ROC,
and confusion matrix.
Start
Problem
Type?
Classification Regression Clustering Ranking

Accuracy, Precision, Mean Absolute Error Adjusted Rand Index Precision@k, Mean
Recall, F1 Score, (MAE), Mean Squared (ARI), Silhouette Average Precision
AUC-ROC, Confusion Error (MSE), Root Coefficient (MAP)
Matrix Mean Squared Error
(RMSE), R-Squared
FIGURE 19.3: Choosing model performance metrics
Regression Metrics
Consider these regression metrics:
➤➤ Mean absolute error (MAE): This gives an idea of how wrong the predictions were and is calculated by
taking the average of the absolute differences between the predicted and actual values.
➤➤ Mean squared error: This is similar to MAE except that the differences are squared before averaging, so
they accentuate the more significant errors more than the MAE.
Tools for Monitoring Models ❘ 367
➤➤ Root mean squared error: As the name suggests, it is the square root of the MSE.
➤➤ R-squared: Also known as the coefficient of determination and is the proportion of the variation in the
dependent variable that is predictable from the independent variable(s).
Clustering Metrics
Adjusted Rand index and silhouette coefficient are two clustering metrics.
➤➤ Adjusted Rand index (ARI) measures the similarity between two data clusters.
➤➤ Silhouette coefficient identifies a model that has better-defined clusters. It provides a visual way to
assess clusters.
Ranking Metrics
Precision@k and mean average precision are two metrics used to measure rankings.
➤➤ Precision@k measures how many of the first k recommendations are in the set of true positive items.
True positive is when the actual and predicted values are the same.
➤➤ Mean average P(MAP) is used to measure the ranking of the recommended items.
TIP Set reasonable ranges for metrics, and when the performance falls out of range, trigger a
review or retraining process.
REAL-TIME VS. BATCH MONITORING

You can monitor your models either in real time or as a batch process. Your choice will depend upon several fac-
tors, some of which are mentioned next.
When to Use Real-Time Monitoring

In the case of real-time monitoring, data is collected in real time and analyzed in real time. Real-time monitoring
is required if your model is used in healthcare or finance use cases, where you need to respond quickly to data
changes or take quick action in response to predictions. The real-time option is the way to go if you have strin-
gent latency requirements.
When to Use Batch Monitoring

Batch monitoring is appropriate when the requirement is to get a comprehensive view of the model’s perfor-
mance. Data is collected over a period of time and is analyzed at a later time.
NOTE If the data under consideration is extensive, real-time monitoring may not be an
option. Real-time monitoring is also more expensive.
TOOLS FOR MONITORING MODELS

This section describes some commonly used tools to monitor models. We can broadly classify them into three
types: cloud provider tools, open-source libraries, and third-party tools. Note that this is not an exhaustive list.
Cloud Provider Tools

The following tools are available from major cloud providers:
➤➤ Amazon SageMaker Model Monitor is specifically designed to monitor models to detect and remedy
model quality issues and does not require any manual intervention.
➤➤ Azure Machine Learning and Azure Application Insights tracks web apps for performance and failures.
➤➤ Google Cloud’s AI platform and Google Cloud Operations suite monitors Google cloud services.
➤➤ IBM Watson OpenScale provides a similar functionality.
Open-Source Libraries
In this category, you have the following options:
➤➤ Evidently, Prometheus, with Grafana monitor model performance, data, and fairness.
➤➤ Amazon Deequ is a validation library for Apache Spark.
➤➤ Tensorboard tracks TensorFlow models.
Third-Party Tools
A number of third-party tools such as Whylabs, Arize, Datadog, Monte Carlo, Labelbox, and Seldon are worth
checking out.
TIP Choose cloud provider tools for scale, open-source tools for customizability, and third-
party tools for an end-to-end integrated solution.
BUILDING A MODEL MONITORING SYSTEM

You need to consider some things before deploying your model in production from a monitoring standpoint (see
Figure 19.4).
1 2 3 4 5
Employ a Conduct
Decide on the Set Up the
Monitoring Set Up Alerts Periodic
Model Metrics Thresholds
Service Reviews
Monitor Establish statistical, Use monitoring Configure alerts for Perform regular
infrastructure business, and services like issue detection checks and adjust
health, endpoint adaptive thresholds Amazon metrics and
responses, track CloudWatch and thresholds
model versions and AWS Cost Profiler
data to track metrics
FIGURE 19.4: Monitoring the health of your endpoints

Building a Model Monitoring System ❘ 369
Determining the Model Metrics to be Monitored

First, you must decide on the metrics you should monitor for your model endpoint once the model is deployed.
You should do the following:
➤➤ Monitor metrics to measure the health of the infrastructure hosting the endpoint. The metrics used to
measure the health of the underlying compute resources are memory and CPU/GPU usage, ML endpoint
invocations, and latency.
NOTE You can use Amazon SageMaker Model Monitor to monitor the quality of the
models and send metrics data to CloudWatch.
➤➤ Monitor the health of the endpoint responses. The metrics used to monitor the model responses are the
accuracy and number of errors in the model response. You can use Amazon SageMaker Clarify to iden-
tify bias and feature attribution drift.
➤➤ Track and trace these metrics to the model’s version or the data used for these predictions.
Setting Up the Thresholds for Monitoring

Thresholds determine whether a metric is acceptable within a specific range, and alerts are triggered
when exceeded.
You can set statistical thresholds for triggering alerts if the average prediction accuracy over time falls below a
certain percentage. You can also set business thresholds based on business impact, where the threshold level may
vary depending on the criticality of the metric. In other cases, you might have an adaptive threshold, where the
threshold varies depending on certain factors, such as the time of day.
Ensure you review the thresholds periodically because the models and data change and the thresholds may have
to be changed accordingly.
Employing a Monitoring Service with Dashboards

You must employ a monitoring service that provides dashboards and reports to track these metrics. For example,
you can use Amazon CloudWatch to track the metrics of your SageMaker deployments. You can use Amazon
EventBridge to automate event responses and AWS Cost Profiler to keep tabs on expenses.
Setting Up Alerts
You should also set up alerts to be notified in case of an issue, such as increased memory utilization, a sudden
increase in latency, and so on. Integrate them with other notification systems such as emails, SNS, SMS, and
even Slack.
TIP Ensure the alerts also have the necessary information as to what went wrong and pos-
sibly the root cause so that you can act on them.
Conducting Periodic Reviews

Finally, conduct periodic reviews to ensure you have the right metrics and thresholds and make necessary changes
as needed.
MONITORING MODEL ENDPOINTS

Model endpoint best practices such as automating endpoint changes, implementing a recoverable endpoint, and
autoscaling endpoints ensure models perform reliably and help identify opportunities to improve predictions,
reduce costs, and avoid failures.
Automating Endpoint Changes Through a Pipeline

Every time you change the model, you must manually deploy that change to the endpoint. It is an error-prone,
time-consuming process. A better way to do that is to use a pipeline that does the following:
➤➤ Checks out the latest version of the model code
➤➤ Builds the model
➤➤ Deploys the model to the endpoint
By integrating the endpoint into the change management system, you can track every time a new model is
deployed to the endpoint. When you face a problem during a deployment, you can roll back to the older version.
This process is not only automatic, but it also reduces downtime. You could use a change control system such as
Git and CI/CD pipeline management tools like Jenkins, GitLab, or AWS CodePipeline.
TIP By automating the model deployment to the endpoint, you can eliminate the errors
and reduce the downtime, which will reflect in the improved accuracy of your model, thus
helping you achieve your business objectives.
Implementing a Recoverable Endpoint

It is vital to have a recoverable endpoint and a solid version control strategy if your system goes down because of
errors or any other reason. You must ensure all components relevant to the hosting model endpoint are version
controlled and the relationships between these components are maintained in a lineage tracker system. Maintain-
ing all the versions will help you maintain backups of all these components, and maintaining the relationships
will help you roll back when the situation demands. Figure 19.5 shows the different steps involved to implement
a recoverable endpoint.
Consider how to achieve this using AWS services.
➤➤ You can implement Amazon SageMaker Pipeline to automate the development, training, and deploy-
ment of the models and switch between versions seamlessly.
➤➤ Using SageMaker Projects, you can enable collaboration between the data scientists and the developers
by bringing all the necessary components to build an end-to-end ML solution under one umbrella. These
components could include the pipeline executions, the registered models, the ML code, the datasets, and
the endpoints.
➤➤ At the same time, we can use the AWS CloudFormation tool to manage the infrastructure as code to
manage different versions of the infrastructure and revert to different versions if necessary. You can store
the code in AWS CodeCommit repositories.
➤➤ You can use Amazon’s ECR as a container registry to manage different versions of the containers and
revert to different versions as needed.
➤➤ Finally, you can use AWS SageMaker Lineage to track the relationships between all these components
and troubleshoot issues quickly.
Monitoring Model Endpoints ❘ 371
01 Start
02 Version Control System AWS CodeCommit
03 Container Registry Amazon ECR
Implementing a
Model Development,
Recoverable 04 Training, Deployment
SageMaker Popeline
Endpoint
Collaboration &
05 Resource Management
SageMaker Projects
06 Tracking Relationships SageMaker Lineage
07 Recoverable Endpoint
FIGURE 19.5: Implementing a recoverable endpoint
TIP Plan for scale and prepare for the unexpected and leverage tools to automate pipelines,
control versions, centralize components, track lineage, and adopt autoscaling.
Implementing Autoscaling for the Model Endpoint

As you start taking more and more workload, ensuring that your model endpoint can serve customers with the
same quality of service is essential. You can achieve this by implementing autoscaling of your model endpoint.
Configure Automatic Scaling Using SageMaker Endpoint

SageMaker Endpoint can automatically increase or decrease the number of EC2 instances attached to the model
endpoint based on the workload or the number of inference requests. The advantage of this approach is that there
is no need for manual intervention.
Leverage Amazon Elastic Inference

Another thing that you can do is, in addition to configuring the right-sized EC2 instances with the required CPU
and memory power, is configure Amazon’s Elastic Inference, which helps the system choose the correct number of
GPU-powered inference accelerations to serve the load.
EXERCISE: MACHINE LEARNING OPERATIONS (MLOps) EXERCISES
The field of MLOps is complex and requires a combination of technical skills, strategic
thinking, and effective implementation. This chapter contains a set of hands-on exercises that
cater to different aspects of a machine learning pipeline, reflecting real-world scenarios. From
continues
continued
setting up alerts and monitoring dashboards to automating deployment and ensuring scalabil-
ity, these exercises are reflective of most MLOps responsibilities. It is vital that the different
task owners in these exercises collaborate with each other to deploy these configurations. Of
course, in your specific organization, these titles may vary, and therefore I recommend you
tailor them accordingly.
TASK NAME TASK DESCRIPTION TASK DELIVERABLES

OWNERS
Setting up Define and simulate Monitoring Alerting rules,

thresholds alerting thresholds team simulation script,
including statistical, and documentation
business, and adaptive of thresholds
thresholds.
Monitoring Set up CloudWatch for DevOps team CloudWatch

service with monitoring SageMaker dashboard,
dashboards deployments and monitoring
create widgets for configurations,
specific metrics. and response
documentation
Setting up alerts Integrate alerts with Alerting team SNS topic,

email or Slack using subscription setup,
Amazon SNS and create alerting script,
conditions for alerts. and integration
documentation
Conducting Create a review Model Review schedule,

periodic reviews schedule for a review team review process
hypothetical model documentation,
and define the and potential
review process. change
recommendations
Choosing model Implement various Data science Code for

performance metrics for classification, team calculating metrics,
metrics regression, clustering, comparison report
and ranking using of different metrics
sample data.
Updating Simulate the updating ML Updated

model regularly of a model with new engineering model, data
data and algorithms, team drift simulation,
demonstrating data and update
drift handling. documentation
Optimizing Model Performance ❘ 373
TASK NAME TASK DESCRIPTION TASK DELIVERABLES

OWNERS
Keeping models Implement a retraining Pipeline team, Retraining pipeline,

fresh with pipeline using tools like DevOps team, schedule, and
scheduler Apache Airflow or AWS MLOps team performance
Step Functions. tracking over time
Automating Automate the Deployment CI/CD pipeline,

endpoint changes deployment of a model team, DevOps automated
using Jenkins, GitLab, or team, MLOps deployment script,
AWS CodePipeline. team and rollback
procedures
Implementing a Use AWS services Recovery Version-controlled

recoverable to create a team, DevOps endpoint, recovery
endpoint version-controlled, team, MLOps procedure, and
recoverable endpoint. team lineage tracker
system setup
Implementing Configure autoscaling Scalability, Autoscaling

autoscaling for a SageMaker DevOps team, configuration,
endpoint and simulate MLOps team load simulation
different loads. script, and Elastic
Inference setup
OPTIMIZING MODEL PERFORMANCE

It is vital to ensure optimal levels of model performance by continuously evaluating various aspects of models
such as data, features, and hyperparameters. Some of these techniques include the following:
➤➤ Reviewing features periodically
➤➤ Implementing model update pipeline
➤➤ Keeping models fresh with schedulers
➤➤ Monitoring data distributions
➤➤ Performing A/B testing
➤➤ Checking for skewed distributions
➤➤ Applying data balancing
➤➤ Detecting outliers and anomalies
➤➤ Tuning hyperparameters
This section delves into reviewing features periodically, implementing model update pipelines, and keeping models
fresh with scheduler pipeline (see Figure 19.6).
Process Data Develop Model Deploy Monitor
Real-time Inference Pipeline

Online
Online Feature Pipeline Feature Online Pipeline
Lineage
Store
Tracker
Offline
Offline Feature Pipeline Batch
Feature
Store
Process Data Prepare Data
Data Prepare Pipeline Application

CI/CD/CT Pipeline
[Re]Train Pipeline (Model Update)
Evaluate Deploy Monitor
Collect Preprocess Engineer
Data Data Features
Fetch Artifacts
(Model/Data Quality, Model
Explainability, Model Bias)
Alarm Model Scheduler
Retrain Pipeline
Manager Registry
FIGURE 19.6: ML lifecycle
Reviewing Features Periodically

Depending on the changes in your business environment, you must schedule a periodic review of your features
used in the model. This will help you ensure that your model is performing accurately. As part of this process, you
must not only review for new features that are impacting your model performance but also explore the opportu-
nity to improvise upon the existing feature sets in your model.
You can use AWS Data Wrangler to look for features that were not used in the original model as well as to review
existing features for improvement opportunities.
How frequently you review the features will depend upon how frequently your business environment changes.
Implementing a Model Update Pipeline

You need to have a model update pipeline in place, which gets triggered in the case of any violations in the per-
formance of the model. This pipeline should activate the data prepared pipeline, the feature pipeline, and the CI/
CD/CT pipeline.
This retraining process will ensure that the new data is cleansed before feeding into the model and new features
are engineered, if required, in the CI/CD/CT pipeline, ensuring that the model is trained and tested before deploy-
ing into production. This way, the model is continuously retrained with new data and can continue to make
accurate predictions.
Keeping Models Fresh with a Scheduler Pipeline

The scheduler pipeline plays a vital role in keeping the AI models fresh. When new data comes into the model, the
data distribution and patterns can change, impacting the accuracy of the model predictions. When that happens,
it is essential to retrain the model so that its predictions reflect the input data.
For example, the patient’s health may depend on several factors. However, these factors can change with time,
a phenomenon known as concept drift. There may have been a disease outbreak that caused a change in these
factors. Under such situations, you need to retrain your model; this is where a scheduler pipeline can greatly help.
The scheduler pipeline retrains the model at predefined business intervals, which helps keep the model always
reflective of the latest data patterns.
The data preparation pipeline gets activated when the scheduler pipeline activates the retraining process. The data
preparation pipeline collects and cleans the new data for use in a model.
At the same time, the feature pipeline gets activated, which extracts the important or relevant features from the
prepared data to feed into the model.
Once the model gets retrained with the new features and when done, the CI/CD/CT pipeline gets activated. Dur-
ing this process, the trained model gets integrated with the rest of the system, is then deployed to the production
environment, and then tested to ensure its functionality and performance are intact.
TIP A scheduler pipeline ensures that the model is constantly kept up-to-date even if the
data or model drift happens.
WORKBOOK TEMPLATE: MODEL MONITORING TRACKING SHEET

Download the “Model Monitoring Tracking Sheet” template from the download section of the book (www
.wiley.com/go/EnterpriseAIintheCloud). Use this template to ensure that the model is delivering the
expected results and is in line with the business goals. Keep track of key performance indicators, issues, and
necessary updates.
SUMMARY
This chapter discussed monitoring models and the key components of monitoring, such as monitoring data qual-
ity, data drift, concept drift, model performance, bias drift, and feature attribution drift. It discussed real-time
versus batch monitoring and the tools for monitoring models.
To ensure resilience and recovery, the chapter also covered advanced topics such as automating endpoint changes
through a pipeline, implementing recoverable endpoints, and autoscaling to handle variable workloads.
At this point, you should be equipped with the knowledge and strategies to not just deploy but also to monitor
your models to ensure reliability, scalability, and performance. Needless to say, this is just the beginning of your
learning journey, but I hope this chapter has given you a good start.
In the next chapter, let’s deal with model governance nuances such as ethics, bias, fairness, explainability, and
interpretability.
REVIEW QUESTIONS
1. What is the main purpose of monitoring model performance?
A. To ensure that the model performs above a certain threshold
B. To detect bias in the model’s predictions
C. To monitor the distribution of the data used to train the model
D. To predict future changes in the model’s accuracy
2. Which of the following performance metrics for regression problem gives an idea of how wrong the
predictions were?
A. Mean absolute error (MAE)
B. Mean squared error
C. Root mean squared error
D. R-squared
3. Which clustering metric measures the similarity between two data clusters?
A. Mean average precision (MAP)
B. Adjusted rand index (ARI)
C. Silhouette coefficient
D. Precision@k
4. What is the concept drift in the context of AI models?
A. The idea that the factors impacting the output of a model may change over time
B. The phenomenon that AI models become outdated if not updated regularly
C. A situation where the model’s accuracy improves over time
D. The occurrence of errors in AI models as a result of excessive retraining
5. How can you identify sensitive data such as personally identifiable information (PII) and intellectual
property in your system?
A. By enabling data access logging
B. By classifying your data using tools such as Amazon Macie
C. By monitoring data to look for any anomalous incidents
D. By using threat detection services
ANSWER KEY
1. A 3. B 5. B
2. A 4. A
20
Governing Models for Bias and
Ethics
In matters of truth and justice, there is no difference between large and small problems, for issues
concerning the treatment of people are all the same.
—Albert Einstein
This chapter dives into governing models, explaining why they are more than just a regulatory concern
and are an integral part of responsible AI. This chapter serves as your guide to deal with nuances such as
ethics, addressing bias, ensuring fairness, and diving into making models explainable and interpretable (see
Figure 20.1).
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Govern Your
Your AI
journey
09
• Deploy Your Models into Production
ENTERPRISE AI Setup 04 • Monitor and Secure Models
Evolve and
Mature AI
JOURNEY MAP Infrastructure • Govern Models for Bias and Ethics
and Manage
Operations

AI
Deploy and 05 Establish a comprehensive governance
Monitor framework for AI/ML models, ensuring ethical
Models deployment, fairness, transparency, and
07 cost-effectiveness, while actively mitigating
06 biases and enhancing interpretability
FIGURE 20.1: DEPLOY AND GOVERN MODELS: Govern models for bias and ethics
378 ❘ CHAPTER 20 Governing Models for Bias and Ethics
Whether you are implementing augmented AI for human review, managing artifacts, controlling costs, or setting
up an entire AI governance framework, this chapter can guide you.
IMPORTANCE OF MODEL GOVERNANCE

You know that AI is a powerful technology, but ensuring that innovation happens responsibly is also essential.
Consider these reasons why governance is critical:
➤➤ Ensuring that the models are deployed and governed ethically and responsibly and employed in a non-
discriminatory manner is critical for society.
➤➤ We need to ensure that the models protect data privacy and comply with various regulations such as
HIPAA. GDPR, and CCPA.
➤➤ Models unintentionally perpetuate bias, and it’s essential to conduct periodic audits and take corrective
actions where needed.
➤➤ We need to implement measures so that we can explain and interpret how models make their
predictions.
➤➤ We can build customer trust by adopting strict governance and compliance processes.
➤➤ By complying with regulations, we can also avoid potential financial penalties.
STRATEGIES FOR FAIRNESS

This section explores different strategies to ensure fairness in your models. See Figure 20.2.
Addressing
Fairness and
Bias in
Models
Ethical STRATEGIES Addressing Model

Considerations for TO ENSURE Explainability and
Deploying Models FAIRNESS Interpretability
IN MODELS
Implement
Augmented
AI for Human
Review
FIGURE 20.2: Strategies to ensure fairness in models

Strategies for Fairness ❘ 379
Addressing Fairness and Bias in Models

Bias and fairness are used interchangeably. Bias is a statistical concept of a model, and it refers to the difference in
its predictions for specific groups or populations. Fairness is a much broader and subjective concept, and it refers
to the absence of any discrimination against a particular group of people.
Consider these steps for mitigating bias:
1. First, you need to check if any groups are overrepresented or underrepresented and also look for labels
or features with inherent biases.
2. You can correct bias using resampling techniques or generating synthetic data to balance their rep-
resentation.
3. You can also deliberately adopt algorithms with built-in fairness mitigation constraints used dur-
ing training.
4. You can update the model’s thresholds during inference to equal error rates across different groups.
5. Lastly, you’ll adopt tools such as SHAP, LIME, or tools like Amazon SageMaker Clarify to understand
how these models work so you can address bias accordingly.
TIP Schedule periodic bias audits, preferably by a third-party to ensure an unbiased review.
EXERCISE: ANALYZING FAIRNESS AND BIAS
Goal: To understand and mitigate biases in a given dataset.
Select a dataset: Choose a public Data analyst Chosen dataset

dataset that might contain biases related
to gender, race, or other factors.
Analyze for bias: Utilize techniques such Data scientist Bias analysis report
as statistical tests to identify any inherent
biases.
Mitigate bias: Apply resampling or Data scientist Balanced dataset

synthetic data generation to balance the
representation.
Report findings: Document the process Data analyst/team Detailed report of

and findings in a detailed report. lead process and finding
Addressing Model Explainability and Interpretability

Model explainability and interpretability are interrelated but with subtle differences. Interpretability is about
understanding the model’s inner workings in terms of how it came up with its predictions based on the
input features.
Model explainability is about trying to make sense of the model’s outputs, irrespective of whether the model is
interpretable. Model explainability especially comes into play when dealing with neural networks where it is
simply impossible to interpret how the model is working due to its complexity.
For model explainability, you can use feature importance scores, partial dependence plots, or more advanced
methods like Local Interpretable Model-Agnostic Explanations (LIME) or Shapley Additive Explanations (SHAP).
TIP Consider setting up a central repository to store model decision explanations that can
be referenced later during audits or inquiries.
Implementing Explainability for Your Models

When working in highly regulated industries such as finance and healthcare, it is essential to ensure that you can
explain how the model came up with its predictions.
To begin with, you need to document the explainability needs of your business. For example, you need to
decide why you want the model to be explainable. Is it because you want to build trust in your model with
your customers by explaining how you came up with a particular decision, or do you want to comply with your
regulators?
Next, you need to understand the relationship between the complexity of the model and the difficulty in explain-
ing them. This understanding will help you choose the right model type and develop the proper set of evaluation
metrics. That way, you can come up with a performant and explainable model.
You can use SageMaker Clarify to explain the model’s predictions. It uses feature attribution to explain which
feature contributed to those predictions, and it can also help you to keep track of any drift that arises due to
changes in the incoming data distribution and, consequently, the drift in the feature attribution.
SageMaker Clarify can also generate governance reports, which can be helpful to share with your regulators and
your risk and compliance teams so they can understand how these predictions were made to ensure the model’s
fairness and explainability.
Ethical Considerations for Deploying Models

As shown in Figure 20.3, when deploying models in production, you need to factor in a few considerations, such
as the following:
➤➤ Bais and fairness
➤➤ Transparency and explainability
➤➤ Data privacy and security
➤➤ Robustness and reliability
➤➤ Accountability
➤➤ Inclusivity
TIP Regularly discuss with ethicists or an ethics review board and get their sign-off before
deploying a model in production.
Operationalizing Governance ❘ 381
Bias and
Fairness
Transparency
Inclusivity and
Explainability
Ethical
Considerations
Data Privacy
Accountability
and Security
Robustness
and Reliability
FIGURE 20.3: Ethical considerations when deploying models
Implementing Augmented AI for Human Review

Another best practice is to establish a QA system to monitor the performance of your model in production. Set
up a team of domain experts who are responsible to evaluate these predictions, especially those that have a low
confidence score.
You can use Amazon’s A2I, which redirects predictions of low confidence score team for review. It can also
randomly sample a set of predictions and forward them to the QA team to review. This approach ensures high
accuracy and performance of the model.
TIP Leverage augmented AI to provide valuable feedback to the AI team that built the
model to improve the model parameters when needed.
OPERATIONALIZING GOVERNANCE
The previous section discussed different strategies to ensure fairness in your models. This section dives into some
of the best practices to operationalize governance in the model’s lifecycle, from creation to monitoring. It involves
tracking the models, including its owners, lineage, versions, artifacts, and costs, and establishing an ethics govern-
ance framework.
Tracking Your Models

The following sections cover various steps involved in tracking your models.
Tracking Models, Owners, and Usage for Governance

Tracking the models, the owners, and usage will not only help you in organizing your operations but will also
help you with governance. Here are some recommended actions:
➤➤ Track all the models that are developed and deployed using unique identifiers. It will help you to track
their usage and performance over time.
➤➤ Track the owners of the model depending upon its stage in the lifecycle. Models may have different own-
ers for development, deployment, monitoring, and performance as well as their governance and usage.
➤➤ Track the usage of the model, how many predictions it made, how often it is used, what are the underly-
ing resources it consumed, and the business outcomes it achieved.
TIP Store all tracking information in a central repository for access by others for transpar-
ency and ease of governance.
EXERCISE: MODEL TRACKING
Goal: Tracking a model through its entire lifecycle.
Create a model lifecycle: Document a full Model architect Lifecycle design

lifecycle of a model from data collection document
to deployment.
Implement lineage tracking: Use tools Data engineer/ML Implementation

like AWS SageMaker or MLflow for engineer of tracking tools
tracking.
Analyze lineage: Analyze the tracked Data scientist Lineage analysis

information to understand how a model report
evolved, how it was built, and how
predictions were made.
Write report: Write a report detailing Data analyst/project Detailed report

the benefits and challenges of lineage manager on benefits and
tracking. challenges
Implementing Version Control

You’ve read about this already, but not from a governance perspective. Make sure you maintain the versions of
different model artifacts, which will help you to roll back in case of any issues as well as during troubleshooting.
You must track not only the models but also the underlying data, code, model parameters, hyperparameters, and
underlying infrastructure.
Adopt nomenclature best practices such as when naming versions, and make sure to include the order of the ver-
sions, timestamps, and semantic versions.
Use tools like MLflow, Data Version Control (DVC), and even Git for version control.
TIP Clean up versions regularly to optimize storage and improve clarity.
Tracking Model Lineage

Model lineage tracking is the ability to track the model through its lifecycle, and it has many advantages. Here
are some of the benefits of model lineage tracking:
➤➤ It helps to reproduce a model when required.
➤➤ Knowing exactly how it was built will help explain it to the auditors and regulators.
➤➤ Knowing how the model was built and how it came up with its predictions will help build trust with the
customers and other stakeholders.
➤➤ Tracking the lineage will help manage the different versions of the model and revert if required due to
poor performance.
➤➤ For the same reasons, it can help with debugging and troubleshooting.
Here are some actions to implement lineage tracking:

➤➤ Leverage MLOps tools such as AWS SageMaker, MLFlow, KubeFlow.
➤➤ Leverage version control systems such as Git.
➤➤ Maintain a tracker of your experiments in a central repository.
➤➤ Track the metadata such as what model, which data, what hyperparameters, and so on. Implement auto-
mation where possible to log your experiments.
➤➤ Implement access control to ensure collaboration and reuse.
TIP Ensure lineage data is backed up for future audits.
Managing Model Artifacts

An essential aspect of MLOps is to manage the artifacts to allow collaboration, reproducibility, troubleshooting,
governance, efficiency, and overall better MLOps.
The Importance of Managing Artifacts

This section examines some of the reasons that you need to manage artifacts, as listed in Figure 20.4.
➤➤ Reproducibility: Managing the artifacts will help to reproduce the experiments and train any inferences.
Knowing how the model was created, what data was used, and how the inferences were made will help
with troubleshooting, explainability, and compliance reasons.
➤➤ Collaboration and reusability: It’ll also help machine learning engineers, data scientists, data engineers,
and software engineers to collaborate and reuse the models, build upon previous work, avoid duplicate
work, and save time.
Cost Savings and Efficiency
Governance and Compliance
Performance Monitoring
Version Management and Rollback
Collaboration and Reusability
Reproducibility
FIGURE 20.4: Different benefits of artifacts management
➤➤ Version management and rollback: Maintaining the versions of these artifacts and their relation-
ships will help you to roll back to a previous version, which will be handy for troubleshooting and
auditability.
➤➤ Performance monitoring: When the model performance starts to dwindle, you can try to compare differ-
ent versions to understand what changed that caused this performance degradation.
➤➤ Governance and compliance: It is important to keep different versions of the artifacts simply for govern-
ance and compliance reasons so that we can provide the required documentation to the regulators.
➤➤ Cost savings and efficiency: By implementing an efficient artifact management process, we can remove
or archive old files, avoid duplicate files, save costs, and improve efficiency.
TIP Organize periodic knowledge sharing sessions to familiarize teams with existing models
and artifacts.
Types of Artifacts to Be Stored

This lists describes the types of artifacts that need to be properly managed (see Figure 20.5).
➤➤ Preprocessed dataset that can be used for future training
➤➤ Training dataset to reproduce the training process
➤➤ Test dataset for testing purposes
➤➤ Validation dataset for validation purposes
➤➤ Model parameters such as the weights to re-create the training model
➤➤ Model configuration, such as the hyperparameters, to re-create the training process
➤➤ Evaluation metrics to evaluate model performance
➤➤ Logs to help with troubleshooting and performance optimization

➤➤ Feature engineering artifacts such as feature dictionaries, encoders used, scalars, and so on
Feature
Engineering Logs
Artifacts
Model
Evaluation
Parameters and
Metrics
Configuration
Datasets
FIGURE 20.5: Central repository of various machine learning artifacts
Best Practices for Artifact Management

Here are some best practices to implement for artifact management:
➤➤ Implement version control using unique IDs to track different artifacts.
➤➤ Store the artifacts in a reliable storage system such as Amazon S3 or Google Cloud storage.
➤➤ Maintain a central catalog or repository of all the model artifacts along with the metadata, such as who
trained it, when it was trained, and so on.
➤➤ Implement proper access control using IAM roles and policies.
➤➤ Ensure traceability between the model and the data, and the code that was used.
➤➤ Implement proper retention and archival policies to ensure all data is deleted in a timely manner.
Controlling Your Model Costs Using Tagging

Over a period of time, your costs to host your models can increase, and you would need to manage them proac-
tively. You will need to identify how models are being spent across different environments, such as retraining and
posting, as well as across different business units, owners, environments projects, purpose, and so on.
You can implement AWS tagging to keep a tab on these expenses. For example, you can tag AWS instances as
retraining or hosting depending upon the purpose for which they are used.
You can then run billing reports based on the tags and understand how these models are used across these dimen-
sions. This visibility will help you control costs by working with various stakeholders.
The visibility of costs using these tags can help not only control the costs but also plan for future expenses based
on the cost of running the models and the benefits from them.
Setting Up a Model Governance Framework

Ensuring that a model is deployed ethically and responsibly is not just a technical endeavor but also requires
some process initiatives, such as setting up a governance framework.
Setting up a model governance framework involves the steps outlined in Figure 20.6. I provided a hands-on exer-
cise to help you get started.
Establish Implement Conduct Set Up a

Define Establish Review Ensure Continuous Periodic Impart Feedback
Ownership Policies Processes Traceability Monitoring Audits Training Loop
Establish Set policies Implement Ensure Continuously Periodically Impart training Establish a
ownership across the manual and traceability for monitor model audit the on the feedback loop
throughout model automated auditing and performance governance governance for continuous
model’s lifecycle review compliance framework framework improvement
lifecycle processes
FIGURE 20.6: Setting up a model governance framework
TIP Be inclusive of stakeholders from various departments to get a holistic view when draft-
ing an ethics policy.
HANDS-ON EXERCISE: SETTING UP A MODEL GOVERNANCE FRAMEWORK
Goal: Promote transparency, governance, compliance, quality, trust, and alignment with
organizational objectives through a model governance framework.
TASK TASK DESCRIPTION TASK OWNER DELIVERABLES
Define Start by establishing Model Ownership

ownership the ownership of the governance documentation
model for various stages team
throughout its lifecycle.
Establish Establish policies across Compliance/ Policy

policies the model lifecycle regulatory team documentation
such as fairness,
bias, interpretability,
explainability, responsible
AI, data privacy, security,
etc., in line with industry
regulations and company
guidelines.
TASK TASK DESCRIPTION TASK OWNER DELIVERABLES
Establish Establish manual and Review team Review process

review automated review guidelines
processes processes to ensure the
model works without bias
and with fairness.
Ensure Ensure traceability Audit & Traceability report

traceability by tracking the compliance
model’s activities team
across its lifecycle
for troubleshooting,
auditing, and regulatory
compliance.
Implement Implement continuous Monitoring Continuous

continuous monitoring of the models & operations monitoring
monitoring using tools such as team implementation
AWS SageMaker Model and report
Monitor to ensure they
work within prescribed
ranges.
Conduct Ensure adherence to the Internal audit Audit reports

periodic governance framework team
audits of the through periodic audits.
governance
framework
Impact Impart proper training Training and Training materials

training and awareness about the development and attendance
governance framework team records
and have clarity around
everyone’s roles and
responsibilities.
Set up a Establish a feedback loop Feedback and Feedback

feedback loop to ensure the governance improvement mechanism
framework is a living team implementation
document that gets and improvement
constantly updated and plan
improved upon.
WORKBOOK TEMPLATE: MODEL GOVERNANCE FOR BIAS & ETHICS

CHECKLIST
Download the “Model Governance for Bias & Ethics” template from the download section of the book (www
.wiley.com/go/EnterpriseAIintheCloud). Use this template to ensure that the model is not biased and is
ethical so that the results are not skewed or discriminatory in any way.
SUMMARY
This chapter delved into the governance of models. Ethical considerations were emphasized, and you explored
the topics of fairness, bias, explainability, and interpretability, as well as the implementation of augmented AI for
human review. You learned about model tracking, from ownership to version control and lineage, and discussed
managing model artifacts and controlling costs through tagging. The chapter concluded with practical steps to
establish a robust model governance framework.
REVIEW QUESTIONS
1. Which of the following services is useful for detecting threats like data leaks or unauthorized access?
A. Amazon S3
B. Amazon Macie
C. Amazon GuardDuty
D. AWS CodeCommit
2. What is the primary difference between model explainability and interpretability?
A. Explainability is about understanding the model’s inner workings, while interpretability is about
making sense of the model’s outputs.
B. Explainability is about making sense of the model’s outputs, while interpretability is about under-
standing the model’s inner workings.
C. There is no difference; they mean the same thing.
D. Both concepts do not exist in model development.
3. Which of the following is NOT typically considered an artifact to be stored and managed in machine
learning operations?
A. The preprocessed dataset used for training
B. Evaluation metrics to evaluate model performance
C. Temporary files generated during data cleaning
D. Model configuration, such as the hyperparameters
4. What is the first step in setting up a model governance framework?
A. Establish review processes
B. Implement continuous monitoring
C. Define ownership
D. Conduct periodic audits
ANSWER KEY
1. C 3. C
2. B 4. C
PART VIII
Scaling and Transforming AI
You have built it, so now you can make it even bigger! In Part VIII, I present a roadmap to scale your
AI transformation. I will discuss how to take your game to the next level by introducing the AI maturity
framework and establishing an AI center of excellence (AI COE). I will also guide you through the process
of building an AI operating model and transformation plan. This is where AI transitions from a project-
level to an enterprise-level powerhouse.
21
Using the AI Maturity
Framework to Transform Your
Business
The greatest danger in times of turbulence is not the turbulence—it is to act with yesterday’s logic.
—Peter Drucker
Having laid the foundation blocks for integrating AI into your business workflows, you will now venture
into a pivotal phase: scaling and transforming your business through AI.
This chapter signifies a transformative shift from merely deploying AI systems to aspiring to become a
genuinely AI-first company. But then, how do you know you are ready to make the transition? Enter the AI
Maturity Framework (see Figure 21.1).
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Your AI Govern Your Scale and Transform
journey Team
09 • Use AI Maturity Framework to

Transform Your Business
ENTERPRISE AI • Set Up Your AICOE
Setup 04
Evolve and
Mature AI
JOURNEY MAP Infrastructure • Build Your AI Operating Model and
and Manage Transformation Plan
Operations

Objective
AI
Deploy and 05 Assess and advance through distinct stages
Monitor of AI maturity, enabling transformation
Models toward an AI-first paradigm by leveraging a
07 structured maturity model framework
06
FIGURE 21.1: SCALE AND TRANSFORM: Use the AI Maturity Framework to transform your business
392 ❘ CHAPTER 21 Using the AI Maturity Framework to Transform Your Business
As you navigate this chapter, you will find the AI Maturity Framework to be an indispensable tool. This chapter
walks you through the five stages of maturity, starting with the initial discovery, then pilot, operationalization,
scaling, and culminating at the Transforming stage, where AI is not an addition but is core to your business strat-
egy and operations.
TIP Understanding your company’s vision and goals so you can conduct a maturity
assessment helps identify gaps and enables you to create an AI transformation plan. This, in
turn helps you move from pilot to production and makes AI a core part of your business
strategy.
This chapter focuses on leveraging the AI Maturity Framework, a vital tool tailored to guide your business
through the transformational journey of AI integration. By understanding where your organization stands in its
AI maturity and recognizing the stages ahead, you can strategically plan, execute, and optimize your AI-driven
solutions.
SCALING AI TO BECOME AN AI-FIRST COMPANY

Now that you have deployed the pilot in production, it is time for the team to explore expanding AI adoption
across the enterprise by identifying more use cases. You will soon realize that you can fully scale your AI efforts
only if a few foundational elements are in place. To elevate your AI efforts to the next level, I recommend these
three things:
➤➤ Setting up an AI center of excellence (AI COE)
➤➤ Developing the AI operating model
➤➤ Developing the AI transformation plan
These are interrelated, as you will learn after reading the rest of this chapter. The AI COE is responsible for build-
ing the AI operating model, a framework for adopting AI. The AI COE is also responsible for developing the AI
transformation plan, which is used to implement the AI operating model. See Figure 21.2.
Comprised of subject
matter experts for
everything AI
AI Center of
Excellence
Plan to implement AI
A framework for operating model and
adopting AI AI Operating AI Transformation scale AI adoption
Model Plan across the enterprise
FIGURE 21.2: Strategic pillars for an AI-first strategy

Scaling AI to Become an AI-First Company ❘ 393
This section explains how to set up an AI center of excellence to implement AI at the enterprise level. Typically,
companies face many challenges during the first six months of setting up an AI COE, so following some of the
best practices outlined in this section is wise.
NOTE Setting up an AI COE is a strategic endeavor and needs careful thought and leader-
ship support.
Why Do You Need a Maturity Model Framework?

Companies such as Deloitte, Accenture, PwC, and Gartner have researched and concluded that only around 20
percent of companies have been able to go beyond the Envision or Launch phase. One of the reasons is that
these companies have found it challenging to align the business, strategy, people, data, platform, governance, and
operational capabilities to implement and realize the true business impact of AI. This maturity framework helps
you measure where you are in the various stages of AI adoption across various dimensions, such as strategy, plat-
forms, and data. It helps you to put in place the measures to implement AI in production and optimize in the long
term for maximum business impact. See Figure 21.3.
90%
100 Strategy & Planning
80
70%
Data/Models 60 60%
People & Change Management
40
20
90% 100%
Governance
Platforms & Operations
70%
Security
FIGURE 21.3: Sample AI Maturity Framework result
Note that AI is more than just a technology; it is a new way of thinking and operating. To succeed with AI,
organizations must become digital at the core.
TIP To succeed with AI, you must measure AI maturity continuously and address gaps to
become more data-driven and AI-enabled, ultimately leading to new products, services, and
business models.
Measuring AI maturity and addressing the gaps will help you assess and scale your business impact from AI
systems. As you must have gathered from some of the case studies I presented earlier in Chapter 2, your initial
focus will be improving your operations to achieve your business goals. Eventually, your focus will move toward
making AI a core part of your organizational strategy, ultimately leading to defining new products, services, and
business models.
THE AI MATURITY FRAMEWORK

AI is a complex and multifaceted technology that requires multiple dimensions to work together, and this AI
Maturity Framework will help you do that. There are six organizational dimensions and five stages, resulting in a
five-by-six matrix that you can use to assess your organization. These six dimensions are as follows:
➤➤ Strategy and planning
➤➤ Governance
➤➤ People and change management
➤➤ Data/models
➤➤ Platforms and operations
➤➤ Security
It is essential to make progress across all dimensions. Otherwise, it’s not possible to make good progress overall
with AI. For example, even if you make good progress with your platform, people, and operations, you may not
succeed if you lack a strong business case for AI.
The Five Stages of Maturity

The five stages of maturity are Discovery, Pilot, Operationalizing, Optimizing, and Transforming. The best way to
leverage this model is to assess where your company is in those dimensions and then identify the dimensions that
will make the most significant impact when addressed. See Figure 21.4.
The Discovery Stage

You can identify your maturity level based on the following characteristics in the Discovery stage:
Understanding AI Organizations are just beginning to hear about AI and trying to understand how
it can help them.
AI adoption plans Organizations have yet to make any clear plans to adopt AI.
Business use cases This is typically handled by those focused on building business use cases and
understanding costs and benefits.
During the next stage, they are becoming good at identifying good AI opportunities and start putting together a
roadmap to define AI solutions.
Case Study Example
A retail company identified hundreds of potential AI projects but, upon deeper analysis, found that
most of them were addressable through traditional programming options. They evaluated the
remaining use cases for feasibility, viability, and desirability. This process eventually helped them to
develop a strategy and a vision for AI setting them up for the next stage, the Pilot stage.
The AI Maturity Framework ❘ 395
Identify the dimensions that will make Stage 5

the most significant impact when Stage 4
addressed
Transforming
Stage 3
Stage 2 Optimizing
Operationalizing Organizations are
Stage 1 beginning to
Pilot Organizations
develop new
are looking to
Companies are products, services,
Discovery deploy multiple
putting an AI business models
Organizations solutions with
strategy, road positive ROI
have just map and
Organizations
started launching pilot
are beginning to
experimenting into production
hear about AI
with AI
and trying to
understand
Use this AI Maturity Framework to make
the multiple capabilities work together
FIGURE 21.4: Five stages of the AI Maturity Framework
TIP Use the AI Maturity Framework to conduct self-assessment, perhaps annually, bi-
annually, or as needed to identify areas of strength and improvement areas.
The Pilot Stage

You can identify this stage based on the following characteristics:
Initial experimentation The organization has just started experimenting with AI and may have done a
proof of concept but have yet to put it into production.
Learning and iteration The organization is still learning and iterating.
Identifying blocks The organization’s focus is on identifying the blocks and enablers of AI.
and enablers
Identifying The organization is trying to identify which projects can be put into production and
potential projects how they will measure success.
AI governance The organization is focusing on AI governance aspects such as reliability and

accountability.
Case Study Example
An insurance company wanted AI to enable their claims processing division to add scale. They
implemented a proof of concept using deep learning optical character recognition and predictive
techniques to improve the process. They curated test data and evaluated metrics to evaluate the
project’s success. They eventually developed a model that improved their processing rates and saved
cost, thus enabling them to work toward putting it into production, which is the next stage—
operationalizing.
TIP Post assessment, chart out a roadmap for the next phase, but be ready to pivot based on
your progress and external landscape.
The Operationalizing Stage

During this stage, companies are trying to formalize their AI efforts and have put together an AI strategy and
a roadmap. They are putting their proof of concepts into production. This stage can be identified based on
these criteria:
Focus The focus is on achieving business outcomes for business impact. There are clear
on business business cases with agreed upon metrics and adequate processes and tools to ensure
outcomes the responsible use of AI.
Capture Lessons learned to improve operations and technology are captured; for example,
lessons learned refining the data strategy and identifying improvements to AI operations and
technology are included as a follow-up action item for the next iterative plan and
execution.
Get leadership More complex use cases require leadership sponsorship for budget and mandates and
sponsorship ensuring the responsible use of AI.
Case Study Example
A healthcare company wanted to reduce the number of patients missing appointments, as it was
causing a lot of operational and financial issues. A proof of concept was conducted, and the number of
missed appointments was reduced by 25 percent. They worked with administrative staff and doctors
to come up with the requirements for the model. As they gathered metrics on the model’s performance
and continued to improve, they became satisfied with its performance and confident to deploy it into
production.
The Optimizing Stage

During this stage, the companies have put their proof of concepts into production and are scaling AI solutions,
measuring their impact, and optimizing AI by making the necessary adjustments. Figure 21.5 and the following
table capture some of the key elements that signify this stage.
AI solution deployment By now, the organization has deployed at least one AI solution in production
and is looking to deploy multiple solutions with positive ROI.
Component and As more solutions are being deployed now, there is a greater opportunity for
model reuse component and model reuse, and organizational alignment is important.
Infrastructure and Deploying multiple AI solutions causes additional challenges in managing

operational challenges infrastructure and operations, which requires additional investment and
stakeholder buy-in.
Operational maturity In the final stages, the organization has achieved maturity in development
and deployment operations.
Ethics framework The organization has developed an ethics framework for the responsible use
of AI across the company.
C-level support The organization has obtained C-level support for multiple AI deployments
and alignment and achieved organizational alignment.
Deployed at least one AI

solution in production
Production
Deployment
Obtained C-level support Opportunity for
and achieved component and model
organizational alignment C-Level Artifacts reuse
Support Reuse
Optimizing
Stage
Developed an ethics Additional investment to

framework for solve infrastructure and
Ethics Additional
responsible AI use operational challenges
Framework Investment
Matured
Operations
Achieved maturity in development

and deployment operations
FIGURE 21.5: Key elements of the Optimizing stage
Ensure that ethics is ingrained in your AI strategy.
The Transforming Stage

During this stage, companies are transforming their businesses with AI and have the following characteristics:
New products and services They are beginning to develop new products, services, and business models.
development
AI in budget decisions AI is a key factor during budget decisions, and executives make their
decisions based on AI-driven data and insights.
continues
(continued)
Organizational Organizational silos have been broken down to further integrate the teams
silo changes across data, infrastructure, operations, and business functions.
AI in business strategy AI has become fundamental to how business strategy is defined, and the
and ethics focus is more on deploying AI in an ethical and socially responsible manner.
NOTE Very few organizations have reached this stage. The companies that have reached this
stage either were born with AI from the start, such as Uber and Airbnb, or were digital-first
companies that eventually shifted to become AI-first. These are companies such as Amazon
and Google. Other traditional companies are trying to become digital at their core before
becoming completely transformational through AI.
HANDS-ON EXERCISE: DEVELOPING AN AI MATURITY PLAN
This exercise will help your implementation team develop a clear understanding of the
opportunity areas to advance your AI maturity level and create a plan to improve it.
The following table provides the steps your team can take from self-assessment to tracking
progress continuously. At the end of this exercise, every team member should have a clear
understanding of their role in advancing your company’s maturity level.
Discovery: Ask each member to Strategy team AI maturity

individually assess the current AI maturity assessment document
level and document areas of gaps or
concerns using the five stages of AI
maturity and the six organizational
dimensions.
Group discussion: Hold a team meeting Team leads and Meeting minutes and
to discuss the gaps and areas of department heads identified gaps
concerns openly to arrive at a consensus.
Encourage open dialogue.
Goals identification: Based on the Management team Defined goals

consensus, define goals for each of the document
six dimensions based on where your
team wants to be at the end of a defined
period (e.g., quarterly, six months, one
year).
Document action items: Document clear Planning and Action items list
action items or initiatives for those six strategy teams
dimensions to achieve the goals agreed
upon.
Assign task owners: Assign a team Project manager List of task owners for
member who would be responsible for each action item
the action items or initiatives.
Develop a plan with timelines: Attach Project Project timeline and

definite start and completion dates for management schedule
the tasks or initiatives. office (PMO)
Progress tracking: Track the progress Monitoring and Progress reports and
regularly toward meeting the goals control teams updates
previously agreed upon and make the
necessary changes.
The Six Dimensions of AI Maturity

To be successful with AI, organizations must improve their maturity in all the six dimensions, namely, Strat-
egy and Planning, People and Change Management, Platforms and Operations, Data/Models, Governance,
and Security.
TIP AI maturity is not just about implementing AI projects but is more about getting
maximum value from AI.
Strategy and Planning

AI strategy is about having a plan in place to achieve the desired level of maturity in the organization. Strategy is
the most important of all dimensions, and it helps the organization to understand what it is, why it is doing this,
and where and when it needs to happen. Without having a proper AI strategy in place, organizations often find it
difficult to have a sense of direction and the business justification needed to proceed with AI. It balances the long-
term and short-term goals of the organization. Having clarity in this dimension provides momentum in the other
areas of the maturity model and helps the organization move forward more confidently with AI.
Table 21.1 summarizes the maturity stages and explains each one, along with the next steps involved.
TABLE 21.1: The Stages of AI Maturity for Strategy and Planning
MATURITY STAGE DEFINITION NEXT STEPS
Discovery There is no clearly defined strategy for AI. AI is looked Align the business and
upon as a nice-to-have tool. The efforts tend to be either technical leaders with the
too narrow or too broad, having little to no value to the need and value of having
organization. Internal experts are enthusiastic to try some an AI strategy.
projects on the side with a view to learning but no value
to the organization as such.
continues
Pilot Strategy for AI is just beginning to emerge. Some initial Try to gather leadership
use cases are being identified, but still, there is no clear support based on the
understanding of how AI can add value. success of the initial POCs.
Some proof of concepts are being explored with limited
funding from one business unit or a team. The onus still
lies with the project owner to prove the benefits of AI.
Operationalizing A clear strategy for AI has been developed and Document AI strategy
communicated throughout the organization. There is a for the organization
shared understanding of what benefits AI can give to the to establish a shared
organization and use cases have been prioritized based understanding.
on their potential impact. Executive support is available Gain a budget and secure
for the effort allowing the organization to unlock the C-Suite sponsorship for AI
budget and provide a mandate to execute its strategy. projects.
Optimizing There’s a clearly defined AI strategy for the organization, Align AI strategy with
and their strategy is integrated with the business other organizational
strategy. There are established processes to manage roadmaps.
AI projects. The organization is actively monitoring and Discover opportunities
measuring the impact of AI on the organization. to execute across other
BUs or functional teams
for maximizing business
impact.
Transforming The organization is continuously refining its AI strategy Sustain momentum by

based on the results and impact. AI is significantly driving continuing to experiment
value to the business, and there is a culture of innovation and innovate with new
and risk-taking. ideas.
People
This dimension involves having the leadership support and the skills, roles, profiles, and performance measures
required to both implement and operate AI successfully (see Figure 21.6).
Chapter 7 discussed the importance of leading people and how to go about building a change acceleration strat-
egy, as well as all the different deliverables that you have to deliver during the four phases in an implementation.
It discussed leadership support and defining job roles, responsibilities, and personas. It talked about building the
AI training courses and curriculums, as well as transforming the workforce through a talent program. Table 21.2
covers how to measure the maturity of this dimension along the five stages of a maturity model.
Platforms and Operations

This dimension deals with having the right platforms, tools, and operations set up to implement AI.
• Awareness of AI potential
Discovery • Unclear roles
• Need for cross-team collaboration
Define AI-related
roles and • Experimentation
responsibilities Pilot • Collaborative POCs
• Leadership communication
Establish performance
metrics and form • Emerging roles
communities of influence Operationalizing • Defined performance metrics
• Leadership actively engaged
Define roles clearly and

expand communities of • Clearly defined roles
excellence Optimizing • Well-defined talent strategy
• Full integration of AI in operations
Invest in skills and

capabilities to sustain AI • AI transforming business
innovation Transforming operations
• Leaders in AI space
• Proactive skills development
FIGURE 21.6: Maturity levels for the people dimension
TABLE 21.2: The Stages of AI Maturity for People
Discovery Employees are aware of the potential of AI but Develop AI literacy for both
do not know how to apply it to the business. business and technical teams so
Organizations do not know how to define roles they can gain confidence.
and responsibilities to implement AI initiatives. Obtain help from internal and
Business teams need help from their AI external AI experts.
counterparts on how to apply AI to solve business
Encourage collaboration and
problems while the data science teams need help
knowledge sharing between
from the business to understand the business
teams.
problems that are suitable for AI solutions.
Pilot Organizations are trying to experiment with AI Cross-functional and matrix

using different roles and responsibilities. They teams across technical and
should resist the temptation to allow AI teams nontechnical departments
(data scientists, BI, or data analysts) to operate should own the AI POC effort.
independently but should encourage them to use Organize learning activities such
the POC to identify the gaps in other technical as hackathons, virtual training,
and non-technical roles and responsibilities. The demos.
leadership should communicate their AI vision
Identify AI career paths
to the teams and ensure that different teams
and explore implications to
and experts are enlisted to be involved in these
workforce planning.
experimental projects.
continues
Operationalizing New roles and responsibilities are emerging. Organizations need to clearly
Performance metrics are defined to track AI define who’s responsible for
performance of employees. Centers of excellence leadership, budget, team
and communities of influence are being structure, and rules.
established to fill up resources for these new roles. Organizations need to have
Leadership is actively engaged in communicating clearly defined rewards
the vision to the employees to motivate them to and recognition programs
participate in the new vision. to recognize excellence in
performance as well as have
clearly defined performance
evaluation programs
acknowledge AI contributions.
Encourage centers of excellence
and communities of influence to
engage employees outside the
formal org structure to facilitate
collaboration and knowledge
sharing.
Optimizing Organizations have clearly defined roles and Include the AI organization in
responsibilities for AI. Communities of influence the discussions for strategy
and community centers of excellence have at the executive table with
expanded mandates including reaching out to the the accountability to achieve
external AI ecosystem. enterprise level KPI for AI.
There is a well-defined talent strategy to enable Establish formal learning
learning for employees. Leaders are very actively journeys for those involved in
engaged to facilitate organizations through the implementing and using AI.
change. AI is fully integrated into all business
operations and employees are skilled in working
with AI and are actively contributing to the
development and optimization of AI solutions.
Transforming AI solutions are transforming business operations Empower HR talent teams to

and the individuals involved in AI are recognized use AI as a business transmission
as leaders in this space. The organization is tool.
proactively investing in the skills and capabilities Communicate and encourage
of the employees and the leaders to sustain self-driven AI learning paths to
innovation and optimization in the organization. employees to continue their
learning journey.
TABLE 21.3: The Stages of AI Maturity for Platform & Operations
Discovery Platforms either do not exist or are Research new AI platforms and
disconnected. Operational processes do understand the need for CI/CD for
not exist. AI projects.
Pilot Some tools are in place for AI and some Standardize tools and platforms.
CI/CD processes are there but not Establish processes for deploying and
standardized. monitoring AI.
Operationalizing Platforms and tools are standardized Continuously monitor the platforms and
along with established operational tools and ensure scalability and reliability of
processes that are followed. AI deployments.
Optimizing Advanced tools are used to monitor Focus on automation and keep exploring
models and there is seamless integration emerging AI platforms, tools, and
between platforms and operations. operations best practices.
Transforming Platforms are self-evolving, and Foster innovation allowing teams to

operations are mostly automated. AI experiment with new platforms and
projects are deployed automatically with operational strategies.
minimal manual involvement.
Data/Models
In this dimension, you focus on ensuring the data is clean, unbiased, reliable, and available. The models need to
be scalable, accurate, ethical, optimized, and relevant.
TABLE 21.4: The Stages of AI Maturity for Data/Models
Discovery Data is siloed and is of poor quality, Centralize storage of data and set up
and few models, if any, have been data cleaning processes. Start with simple
developed. models to understand data patterns better.
Pilot Data pipelines have been started in Streamline data pipelines and focus on
development. Models have been putting models into production and
developed in pilots but not in validation of models.
production yet.
Operationalizing Data is regularly cleansed, and data Focus on getting consistent data quality
pipelines are implemented. Models are and implementing model versions.
in production but need to be monitored.
Optimizing You have a robust data infrastructure. Explore A/B testing and automated feature
Models are regularly updated and engineering updates for performance
optimized for performance. optimization.
Transforming Data is seen as a valuable asset and Encourage a data-driven decision-making

models are at the core of the decision- culture and continuously retrain models
making process. with new data.
Governance
This dimension deals with the ethical use of AI and the alignment of AI projects with the business goals.
Discovery No formal governance structures Understand the importance of

are in place. governance structures and set up initial
processes.
Pilot Initial processes are set but are not Review guidelines regularly and refine
enforced or standardized. them to align with business goals.
Operationalizing AI governance is clearly communicated, Keep refining governance structures

and there are regular audits of AI projects. based on feedback.
Optimizing Governance is integrated into AI projects Ensure governance processes are being
and ethical concerns are paramount. adapted to changing situations.
Transforming Governance is proactive and is changing Continue to foster ethics in all AI

rapidly in response to changes to the AI projects and regularly review and adapt.
landscape.
Security
This dimension deals with securing the platforms, tools, infrastructure, models, and data.
Discovery Little to no focus on security. Try to understand the need for security
and initiate some security measures.
Pilot Some security measures are in place but Prioritize data encryption and access
are not enforced or standardized. controls. Review potential security issues.
Operationalizing AI security measures are clear and Conduct regular security audits to
communicated. But there may still be address gaps and ensure all personnel
some security lapses or concerns. are well trained in security.
Optimizing Robust security infrastructure is in place Implement advanced security and threat
and threats are being regularly monitored. detection techniques and be proactive
rather than reactive.
Transforming Security is integrated in the DNA of Adopt the latest security techniques and
AI projects. Security protocols are ensure a culture of security first in all
continuously monitored and evaluated. AI projects.
HANDS-ON EXERCISE: MLOps MATURITY ASSESSMENT
The goal of this hands-on exercise is to help your team assess the maturity level in MLOps
based on various areas such as data management, model development, deployment, monitor-
ing, and team collaboration.
Workbook Template: AI Maturity Assessment Tool ❘ 405
Note that this is an iterative process, and you need to conduct this exercise periodically to
improve over time with effort and focus.
TASK NAME TASK OWNER TASK DETAILS DELIVERABLE
Identify Key Areas Team leads Identify the key focus A list of key
for assessment across various areas in the ML workflow focus areas in
disciplines that need to be ML workflow
assessed, for example,
data management,
model development,
model deployment,
model monitoring, and
collaboration among
teams.
Assess Maturity ML team leads Assess your organization’s A report

Level and the teams maturity level for each indicating the
identified area. Use a current maturity
scale such as Discovery, level in each
Pilot, Operationalizing, identified area
Optimizing, Transforming.
Define ML team Based on the assessed A detailed

Improvement leads and maturity levels, define improvement
Plans management plans to improve each plan for each
team(s) area. For instance, focus area
implement a new data
cataloging tool or develop
better data validation
procedures for data
management.
Prioritize ML team You should then prioritize A prioritized list

Improvements leads and these action items based of improvement
management on their potential impact plans
team(s) on your ML process. I
suggest you have data
management take
precedence over model
improvements.
WORKBOOK TEMPLATE: AI MATURITY ASSESSMENT TOOL

Download the “AI Maturity Assessment” template from the download section of the book (www.wiley.com/
go/EnterpriseAIintheCloud). Use this template to evaluate how advanced your organization is and to
identify areas of strength and improvement.
SUMMARY
This chapter covered the various steps you take as you move from the Align phase to the Launch phase and
eventually the Scale phase. You used the AI Maturity Framework to evaluate your progress in strategy, people,
data/models, platforms, governance, and security dimensions. It is a comprehensive framework to transform your
business by adopting AI at the enterprise level. The chapter also discussed how you can take your company from
the Launching to Optimizing to Transforming stages by evaluating your progress in these dimensions.
The heart of the chapter was a hands-on exercise to develop an AI maturity plan. It can guide your team from
self-assessment to continuous progress tracking.
REVIEW QUESTIONS
1. Conducting a company’s AI maturity assessment during the Align phase helps
A. To identify gaps in the company’s vision and goals
B. To assess the company’s AI adoption across various dimensions
C. To create a proof of concept during the Launch phase
D. To optimize the company’s AI efforts by expanding the adoption of AI
2. Identify the dimension in the AI Maturity Framework that is not discussed in this chapter.
B. Governance
C. Finance and accounting
D. People and change management
3. A company is just beginning to hear about AI and trying to figure out how AI can help them. Which
stage of maturity are they in?
A. Pilot
B. Operationalizing
C. Optimizing
D. Discovery
4. What identifies the Optimizing stage?
A. Improving AI operations and technology
B. Maximizing business impact
C. Deploying multiple AI solutions
D. Developing an ethics framework
5. What identifies the Transforming stage?
D. Developing new products
ANSWER KEY
1. B 3. D 5. D
2. C 4. C
22
Setting Up Your AI COE
Great things in business are never done by one person: they’re done by a team of people.
—Steve Jobs
As you navigate your journey through enterprise AI transformation, you might realize that creating and
deploying models into production is just one step of the larger journey toward becoming AI-first. For your
company to fully transition into enterprise-wide AI, it needs a holistic approach to integrating AI into its
core. And this holistic approach is the crux of this chapter.
This chapter delves into the compelling reasons to establish an AI Center of Excellence (AI COE), one of
the three pillars to scale and transform your business with AI. You learn about its team structure, multifac-
eted roles, and the symbiotic relationship between the COE business and platform teams (see Figure 22.1).
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
journey Team

Setup 04
Evolve and
Mature AI
Operations

Objective
AI
Deploy and 05 Establish a centralized AI Center of
Monitor Excellence (AICOE) to drive AI best
Models practices, governance, and scaling strategies
07 across the enterprise, transitioning from a
06 strategic unit to an operational guide for AI initiatives
FIGURE 22.1: SCALE AND TRANSFORM: Set up your AI COE

408 ❘ CHAPTER 22 Setting Up Your AI COE
The AI COE serves as an engine, propelling your organization through its transformation phases, and also acts as
the AI evangelist to promote AI adoption. This chapter dives into how to evolve the AI COE’s role from strategy
to operations and its transition from being a central unit to an enterprise-wide advisor.
SCALING AI TO BECOME AN AI-FIRST COMPANY

Now that you have deployed the pilot in production, it is time for the team to explore expanding AI adoption
across the enterprise by identifying more use cases. You will soon realize that you can fully scale your AI efforts
only if a few foundational elements are in place. To elevate your AI efforts to the next level, I recommend these
three things:
➤➤ Setting up an AI COE
➤➤ Developing the AI operating model
➤➤ Developing the AI transformation plan
These are interrelated, as you will realize after reading the rest of this chapter. The AI COE is responsible for
building the AI operating model, a framework for adopting AI. It is also responsible for developing the AI trans-
formation plan, which is used to implement the AI operating model.
This section explains how to set up an AI center of excellence to implement AI at the enterprise level. Typically,
companies face many challenges during the first six months of setting up an AI COE, so following some of the
best practices outlined in this section is wise.
Setting Up an AI Center of Excellence

The AI COE first focuses on transforming your organization from a traditional company into an AI-first
company. Typically, you start with a small pilot team, and once you find success, you then start scaling it into a
much larger COE. Setting up a full-fledged AI COE can take anywhere from six months to one year or more and
requires leadership commitment.
Remember that setting up a COE could also be part of your AI transformation plan to establish an AI operating
model in the long term.
TIP Before starting an AI COE, ensure everyone is clear about its objectives—to foster inno-
vation, streamline operations, or optimize customer experiences.
Why an AI COE Is Important for Scaling AI Across the Enterprise

Establishing the AI center of excellence is an iterative and time-consuming process, but it is worth the effort. It
can help set up the AI/ML infrastructure in addition to defining the AI strategy and vision, planning a portfolio,
and establishing the governance aspects of an AI initiative.
Your organization will have its own approach to the AI journey, but here are some basic guidelines and best
practices you can follow. The primary function of your AI COE is to be the subject-matter expert on all things AI,
which includes the cloud, data, models, governance, infrastructure, operations, and security.
It acts like a central team of experts from various disciplines who can guide the rest of the organization by adopt-
ing AI technologies and practices. With the help of the COE, you can ensure that AI is implemented consistently
across the enterprise. The AI COE also acts as the central hub for the dissemination of knowledge and skill sets in
addition to imparting training to your employees.
Establishing an AI Center of Excellence ❘ 409
The AI COE can also help mitigate the risks associated with AI, such as ethics, social responsibility, bias, and data
privacy concerns.
Finally, it can also act as an engine for innovation by promoting a culture of experimentation and innovation by
helping to identify new use cases as well as adopting new technologies and approaches.
TIP Clearly communicate the goals, progress, and achievements of the AI COE to the rest of
the organization to build trust and foster a culture of AI adoption.
AI COE Team Makeup

The AI COE is a multidisciplinary team focused on AI strategy, governance, best practices for AI architecture,
implementation, and operations that the rest of your organization can adopt. You should start by having an AI
sponsor responsible for the COE’s success and ensuring that the team’s KPIs are aligned with your AI strategy and
business goals.
NOTE The Center of Excellence team may sometimes also be known as the AI capability
team, AI practices team, and so on.
You should include stakeholders from the business and the technology as part of this AI COE; you could call
them the AI business and platform team. You could have people with titles such as IT manager, finance manager,
digital ops manager, operations manager, systems architect, systems administrator, application developer, database
administrator, DevOps manager, AI engineer, ML engineer, data scientist, data analyst, and BI analyst. You can
even have managers from various lines of business, such as finance, human capital, marketing, sales, operations,
supply chain, and so on.
TIP Having a diverse mix of AI experts, data scientists, cloud architects, business analysts,
and representatives across various departments is critical to get a holistic view of your orga-
nization’s needs and challenges.
AI COE Team Responsibilities

Your business stakeholders have to provide strategy, program management, and governance skills. On the other
hand, your technology stakeholders will focus on platform engineering, security, and operations. This team should
be able to make quick decisions, and their focus should be to define at least six or so goals, KPIs, key metrics, and
core services to be achieved within the next six months.
COE Business and Platform Team Responsibilities

Figure 22.2 shows the various responsibilities that fall under the AI business and platform teams.
ESTABLISHING AN AI CENTER OF EXCELLENCE

This section reviews the various steps you need to execute to form and maintain an active, high-performing center
of excellence.
AI Business Team AI Platform Team

Governance • Platform
Reference Architecture • Provisioning and
Financial Reporting Configuration
Security, Risk, and • Service Catalog
Compliance • Operations
Business Outcomes, KPIs • Operational Health
Vendor Management • Lifecycle Management
Business • Capacity Management
Product Management • Availability and Continuity
Financial Reporting • Security
• Identity and Access
People
Management
Org Change Management
• Infrastructure Security
Curriculum Strategy
• Network Security
Onboarding
FIGURE 22.2: AI core team responsibilities
From Centralized Unit to Enterprise-wide Advisor

Your AI center of excellence team should be responsible for driving standards for AI across the enterprise. They
should start as a centralized unit with a limited set of three to five A-class players who can guide the organization
with generic standards. As your organization starts to adopt AI, your AI COE team will also mature. Eventually,
your AI COE team should get into an advisory role to promote the best practices across the organization. As part
of their efforts, they should develop your company’s AI operating model and socialize it across the organization
so that various teams can adopt it.
TIP Empower your AI COE to make quick decisions that will help AI projects and accel-
erate AI adoption.
Evolving from Strategy to Operations Focus

Your center of excellence team can shift their focus from strategy to operations over a period depending upon the
phase of AI within the organization. To begin, their objectives can be the following:
➤➤ Strategy
➤➤ Training
➤➤ Pilot projects
➤➤ Scaling
➤➤ Operating AI
Figure 22.3 shows how your AI COE starts as a centralized unit to becoming an enterprise-wide advisor and
shifts its focus from strategy to operations. This is a good visual that you could use in your discussions about AI
COE within your company.
Establishing an AI Center of Excellence ❘ 411
CENTRALIZED UNIT TO ENTERPRISE-WIDE ADVISOR

Driving Continuous Improvement in
Operating AI
Model Performance
Enterprise-wide
Advisor across BUs Optimizing AI Operations Through
Scaling
MLOps
Setting Up a Production-Grade End-

Pilot Projects
to-End Machine Learning Platform
Increased
Adoption of AI
Launching a Pilot and Building AI
Training
Practices
Identifying and Addressing Gaps for AI

Strategy
Centralized Team Adoption
TEAM GROWTH STEPS IN THE EVOLUTION FOCUS AREA
FIGURE 22.3: Evolution of an AI COE
TIP Embracing an agile approach can help your AI COE adapt quickly to changes and
improve iteratively.
The steps to form an AI COE team are covered next in the form of a hands-on exercise.
HANDS-ON STRATEGIC AND OPERATIONAL PLANNING EXERCISE: SETTING UP

AN AI CENTER OF EXCELLENCE
This is a strategic and operational planning exercise focused on setting up an AI Center of

Excellence in your company.
It has a series of steps requiring active participation from your team in making decisions,
planning, execution, and review. It has components such as forming a team, assigning roles,
managing a project, developing strategy, implementing technology, and reviewing performance.
PHASES ACTIVITY DESCRIPTION
Phase 1 Gathering the Gather your team and discuss the need for an AI COE.
requirements Document the gaps that can be addressed by this AI COE
especially with regard to adoption.
Discuss the team make up and identify potential members.
continues
continued
Phase 2 Planning the Plan the roles and responsibilities of the team members.
AI COE Prepare a charter for the AI COE explaining its purpose,
goals, processes, and procedures.
Phase 3 Scoping the Identify the initial scope for the AI COE such as the
AI COE strategy, training needs, pilot projects, scaling strategies,
and other operational aspects.
Identify training needs and develop the course curriculum
working with third-party vendors and internal training
teams.
Identify the KPIs that will be used to evaluate the
performance of the AI COE.
Phase 4 Implementing Identify a pilot project that can help you to develop AI
the pilot practices that can be adopted for future projects. The goal
project is to build agile practices and develop sets of AI policies for
scaling AI.
Implement the project and document the lessons learned.
Your team must gain hands-on experience through pilots
and demo labs.
Phase 5 Setting Develop the plan for the ML platform.

up the ML Implement a production grade, end to end ML platform
platform integrated with data collection systems and appropriate
access controls.
Different environments such as development, staging, QA,
and production are set up, and budget management and
cost control practices and governance are also established.
Phase 6 Implementing Set up MLOps with automated operations, code

MLOps repositories, model version control, performance
monitoring, notifications, and escalation policies.
The team must also be involved in financial tracking, ML
model deployment policies, and ensuring that any workload
deployed for consumption is production ready.
Phase 7 Reviewing Review the AI COE’s performance, its operations, and

and identify areas of improvement and make necessary
Improving adjustments.
Track the model performance and detect bias, data drift,
and concept drift and put measures in place to address
them with processes, people, and technology.
The team must focus on automating everything, tracking
everything through a portal, tracking costs, KPIs, and
continuous training.
WORKBOOK TEMPLATE: AI CENTER OF EXCELLENCE (AICOE) SETUP

CHECKLIST
Download the “AI Center of Excellence Setup” Checklist template from the download section of the book (www
.wiley.com/go/EnterpriseAIintheCloud). Use this template to help you go through the key compo-
nents to consider when setting up an AICOE (AI Center of Excellence). You can use it to conduct strategic
planning, resource planning, and develop a well-defined roadmap.
SUMMARY
This chapter looked closely at the significance, formation, and the ongoing operation of the AI COE. It discussed
how it is one of the strategic pillars for becoming AI-first, including some of the intricacies involved in setting
it up and its evolution from a strategic focus to an operational one. Through a hands-on exercise, the chapter
offered insights on how to implement an AI COE, including a step-by-step guide to overcome the challenges that
you are likely to face.
The path to becoming an AI-first company has been laid. It’s going to take leadership commitment, strategic
planning, and an unwavering focus on execution. It’s your turn to take action.
REVIEW QUESTIONS
1. Conducting a company’s AI maturity assessment during the Align phase helps
A. To identify gaps in the company’s vision and goals
B. To assess the company’s AI adoption across various dimensions
C. To create a proof of concept during the Launch phase
D. To optimize the company’s AI efforts by expanding the adoption of AI
2. Identify the dimension in the AI maturity framework that is not discussed in this chapter.
B. Governance
C. Finance and accounting

D. People and change management
3. A company is just beginning to hear about AI and trying to figure out how AI can help them. Which
stage of maturity are they in?
A. Pilot
B. Operationalizing
C. Optimizing
D. Discovery
4. What identifies the Optimizing stage?
D. Developing an ethics framework
5. What identifies the Transforming stage?
D. Developing new products
6. Which of the following is important when developing an AI strategy?
A. It should be communicated only to the C-suite.
B. It should not be flexible and should be followed strictly.
C. It should be aligned with the overall business strategy.
D. It is fine not to consider the competitive landscape.
7. The focus of the core team implementing the AI pilot is
A. To establish the platform
B. To implement AI initiatives
C. To conduct a maturity model assessment
D. To establish a COE
8. Why do you need a label in machine learning problem framing?
A. It helps to make predictions.
B. It helps to evaluate the model.
C. It identifies the type of machine learning problem.
D. It is a feature created during feature engineering.
9. What is the primary focus of the machine learning problem framing phase in the machine learning
lifecycle?
A. Data processing
B. Model definition
C. Defining the business goal
D. Identifying the type of machine learning problem
10. Feature engineering is used
A. To evaluate the model’s features
B. To collect and preprocess features
Answer Key ❘ 415
C. To create new data and transform existing data

D. To identify the business data
11. The primary responsibility of a machine learning engineer is
A. Running models and drawing insights
B. Ensuring the data pipeline is appropriately implemented
C. Designing the overall AI strategy
D. Deploying models into production
12. The primary role of a business analyst is
A. Collecting, cleaning, and analyzing data
B. Developing and validating assumptions and hypotheses
C. Bridging the gap between business and technical team
D. Developing the project plan
13. What should be the primary focus of the AI COE team?
A. To implement AI technologies consistently across the enterprise
B. To develop AI models that perform responsibly
C. To generate revenue for the company
D. To manage customer relationships
14. What is an AI operating model?
A. A formal structure or framework to implement AI in a scaled manner
B. A model to develop AI solutions for operations
C. A way to align AI operations with business goals
D. A model to operate AI solutions without compliance
15. An AI transformation plan is
A. A plan for implementing individual AI initiatives
B. A plan to increase the maturity levels of the organization
C. A plan to develop AI technologies
D. A plan to outsource AI development
ANSWER KEY
1. B 6. C 11. D
2. C 7. A 12. C
3. D 8. A 13. A
4. C 9. D 14. A
5. D 10. D 15. B
Building Your AI Operating
23
Model and Transformation Plan
If you don’t know where you’re going, any road will get you there.
—Lewis Carroll
You are now standing at an important juncture—architecting the framework that decides how your
company will embed AI into its processes and sketch the master blueprint for your AI transformation. This
chapter builds on the work done in the previous chapter, where you focused on establishing an AI COE.
Your emphasis now shifts to understanding and implementing an AI operating model that resonates with
your organizational ethos to deliver both value and continuous innovation (see Figure 23.1). It is not just
about technology or even strategy but is more about taking a customer-centric lens and navigating the
intricacies of AI-driven innovation. You will focus on building cross-functional teams, instilling a culture of
continuous improvement and innovation, and aligning with core strategic values.
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
journey Team

Setup 04
Evolve and
Mature AI
Operations
Objective
Transform and Modeling Implement a cohesive AI operating model
AI
Deploy and 05 tailored for your organization, accompanied
Monitor by a robust transformation plan that ensures
Models
a customer-centric, product-oriented
07 approach to AI solutions, emphasizing
06 iterative growth and measurable outcomes
FIGURE 23.1: SCALE AND TRANSFORM: Build your AI operating model and transformation plan
Understanding the AI Operating Model ❘ 417
It doesn’t end there. You will be crafting a transformation plan tailor-made for you, which will act as a catalyst
for transformation to come. You will chart out the long-term vision, objectives, and timelines. A hands-on exer-
cise will help you to draft such a plan.
UNDERSTANDING THE AI OPERATING MODEL

The AI operating model is the second important pillar you need to scale your AI adoption. When you are trying
to implement AI at the enterprise level, it is essential to have a formal structure or framework to implement AI
in a scaled manner. You will find that the success of your AI initiative depends upon your organizational culture
and the existing delivery structures, such as strategic management, program management, portfolio planning,
program governance, and so on. You need a critical mass of people in the organization who can productionize
and operationalize your AI platform in an AI-centric manner. Typically, organizations add layers of AI technology
and project-based management on top of the existing operating models, but they will soon realize that they need
a different operating model that is compatible with AI to leverage some of its benefits that go beyond simple busi-
ness process automation to exploit newer, innovative opportunities as well.
The Purpose of the AI Operating Model

When trying to productionize your POC at an enterprise level, it is effective to keep an eye on establishing a
steady-state operating model so you can scale your AI solution deployment to multiple use cases using a repeat-
able, iterative, and robust development and deployment process. An AI operating model provides a structured
approach to take the AI project from ideation to deployment, along with ensuring alignment with business goals,
complying with legal regulations, and facilitating collaboration between different teams and stakeholders.
As shown in Figure 23.2, an AI operating model includes the AI strategy; the processes to build, test, and deploy
AI solutions; the different roles and responsibilities involved in the initiative; the tools and resources used; and the
success metrics to determine success.
AI Operating Model
Define Roles
Process Tools and Success
AI Strategy Ideation and
Development Resources Metrics
Responsibilities
AI Operating Model Guiding Principles
Start Small
Be Flexible
Executive Buy-in
Avoid Analysis Paralysis
Celebrate Successes
FIGURE 23.2: Components of an AI operating model
By ensuring that these aspects are taken care of, an operating model helps to increase the probability of success of
AI initiatives.
TIP To sustain long-term growth and innovation for your business, position AI as not just
a tool but as a force for business transformation through an AI operating model that instills
AI-first thinking.
418 ❘ CHAPTER 23 Building Your AI Operating Model and Transformation Plan
When Do You Implement an AI Operating Model?

There is no specific time when an AI operating model needs to be developed. However, it is often helpful to define
one when an AI initiative is getting started or when the organization has been informally implementing AI and
the time has come for them to define a formal framework to increase the probability of success. For example, this
could be after the launch, and when they’re trying to scale the adoption of AI across multiple use cases, it might
be an excellent time to develop the operating model. Another reason they may need an operating model is that
they have a clear AI strategy, platform, and resources, but then there is a need to manage AI projects more effec-
tively and measure the success of the projects.
TIP Some tips for developing a successful operating model are having executive buy-in,
starting small, avoiding analysis paralysis, being flexible to incorporate changes as the tech-
nology is fast-moving, and ensuring that you celebrate successes, however small or big they
may be.
IMPLEMENTING YOUR AI OPERATING MODEL

Figure 23.3 shows the transformational steps used to build an AI operating model.
Aligning Operating Model with Strategic Customer-Centric AI Strategy to Drive

Value and Establishing a Clear AI Roadmap 1 6 Innovation
Implement
Continuous Improvement Through an AI Product-Centric Approach to
Measurement & Testing 2 Operating
5 Building AI Solutions
Model
Starting Small and Building Iteratively

with Cross-functional Teams 3 4 Organizing Teams Around the AI Products
FIGURE 23.3: Six transformational ways to build an AI operating model
Customer-centric AI Strategy to Drive Innovation

This involves looking at your customer’s pain points and using data to identify solutions to satisfy your cus-
tomer’s needs. As you can see in Figure 23.4, you must look at the current capabilities across the six dimensions
discussed in the earlier chapter and assess them against your current AI operating model. You use that informa-
tion to prioritize business outcomes and drive innovation without compromising security to come up with new
solutions. Remember that this is an iterative, cyclical process where you identify your customer pain points, assess
your current capabilities and gaps, prioritize your business outcomes, and then drive business innovation by
implementing AI solutions continuously.
Implementing Your AI Operating Model ❘ 419
Assessment
• Business
Identify Prioritize
Customer • G overnance Implement AI
Business
Pain Points • Security Solutions
Outcomes
• Platforms
• Data/model
• Operations
Continuous Process of Improvement and Innovation
• Examples: Comparing these Use the assessment OUTCOME

dimensions with the results to build new, An AI strategy with
• Slow response
current AI operating innovative solutions an operating model
times
model to achieve business
• Lack of
outcomes that meet
personalized
customer’s needs
options
• Limited self-
service options
FIGURE 23.4: Customer-centric AI strategy to drive innovation
Case Study Example
For example, suppose a customer complains about slow response times, lack of personalized options,
or limited self-service options. In that case, you start by assessing gaps in various dimensions such as
business, governance, security, platforms, data/model, and operations, assess your current operating
model, and devise a plan considering your business strategy. You may realize a talent gap, or the
processes must align with your AI strategy. As a result, you may decide to implement a chatbot and
come up with a plan along with an operating model to address the customer’s concern.
By identifying any blocks to meeting the customers’ needs and addressing them, your company can adopt AI suc-
cessfully across the organization. The outcome is an AI strategy with an operating model that enables achieving
business outcomes while meeting the customer’s needs.
TIP Creating a map of the customer journey and tying them to the AI solutions to provide a
seamless, connected customer experience is a good practice.
Using a Product-centric Approach to Building AI Solutions

Building AI solutions from a product-centric approach implies moving away from a system and technology per-
spective into looking at the world as a set of products.
From an AI perspective, say you are building a chatbot. You can divide the AI solution into products such as
natural language understanding, user interface, model for personalized recommendations, database migration,
and data lake building, as shown in Figure 23.5.
BREAKING DOWN MONOLITHIC APPLICATIONS INTO PRODUCTS
Amazon broke down its monolithic Java e-commerce application into a collection of individual
products such as the home page, the search page, the products page, the account pages, the
shopping cart, personalized recommendations, and so on. Behind each one of these products
was a team that was entirely focused on enabling that functionality from end to end, starting
from the requirements gathering all the way to deploying the solution and making sure it
performed well.
Monolithic
Application
(Chatbot)
Natural Language
Understanding User Interface
Personalized Product-Centric
Recommendations Database Approach
Model Migration Decomposing your
monolithic application
into individual products
Data Lake Building Ethics Framework
Benefits of a Product-Centric Approach

Increased focus, improved efficiency, faster deployment, and better performance
FIGURE 23.5: A product-centric approach to building AI solutions
TIP Have a dedicated product team for each product or functionality with end-to-end own-
ership from ideation to deployment to continuous improvement.
Organizing Teams Around the AI Products

Once the products have been identified, the next logical step is to define teams around this product model so that
accountability and ownership can be established around that product (see Figure 23.6).
Establishing this clear ownership makes it possible to meet the customer’s needs. While the business stakehold-
ers within the COE are responsible for ensuring that the product requirements are aligned with the customers’
business needs in line with the business strategy, platform stakeholders are responsible for deploying the products
as reusable code and re-deployable platform components for their customers as required. Automation and self-
service capabilities are essential aspects of this productization as they help increase high availability and reliability
of the systems and security as well. These teams play a vital role in working together to build the required func-
tionality and ensuring its success by working with the business and the technical stakeholders.
Natural Language Understanding

Team
Developing the capabilities of the AI product to
understand and respond to user inputs in a natural,
human-like manner
01
Data Lake Building User Interface

Team Team
Managing the data lake, a 05 02 Implementing the user-
large storage repository that AI PRODUCT facing aspects of the AI
holds a vast amount of raw product
Chatbot Example
data
Personalized
Database Migration Team Recommendations Model
Handling the process of moving 04 03 Team
data from one database to another Managing the AI models that
generate personalized
recommendations for users
FIGURE 23.6: Organizing teams around the product
Use Case Example
In the case of the chatbot example, you would need to organize teams around the natural language
understanding, user interface, recommendation model, database migration, and data lake. These teams
would be responsible for setting up their product strategy, vision, and requirements to ensure they
perform well in production.
Starting Small and Building Iteratively with Cross-Functional Teams

Another vital aspect you need to consider when establishing an AI operating model is to start small and build iter-
atively to support AI adoption. The first product team must be an AI pilot team, which should be cross-functional
and focus on addressing concerns of viability, desirability, feasibility, and operability. As shown in Figure 23.7, as
adoption grows, the AI pilot team should scale and specialize into multiple product teams, such as the AI business
office (ABO) and AI platform engineering (APE).
Key roles in the AI pilot team include the product owner, AI architects, data scientists, machine learning engineers,
and cloud engineers, with additional roles such as financial analyst, organizational change management specialist,
and scrum master potentially added. Initially the pilot team focuses on the viability, desirability, feasibility, and
operability of the PoCs, but as the AI adoption grows, you will be able to reap benefits such as increased speciali-
zation and greater impact to the organization.
TIP Celebrating small wins will not only keep your team motivated but will also reinforce
the value of an iterative approach.
Achieving Better Outcomes from Cross-Functional Teams

Expanded AI Adoption, Increased Specialization, and Greater Impact
AI Business Office (ABO)
AI Platform Engineering (APE)
Scrum Master
Financial Analyst Change Management Specialist
AI Pilot Team
Focus
Product Owner Viability
AI Architects Desirability
Data Scientists Feasibility
Machine Learning Engineers Operability
Cloud Engineers
FIGURE 23.7: Start small and build iteratively with cross-functional teams
Continuous Improvement Through Measurement and Testing

Another critical aspect to consider when defining an operating model is ensuring that the product teams incor-
porate measurement and testing as part of building their solutions (see Figure 23.8). For example, when deploy-
ing a chatbot, they need to come up with metrics to measure the success of the chatbot in production as well as
incorporate testing scenarios across different products to ensure resiliency is built into the product. As part of this
need to build resiliency, the product owner is also responsible for identifying the other products on which this
product depends so that they can consider all the dependencies and factor them into the project plan. By identify-
ing the dependencies and incorporating them into the testing plan, it becomes possible for the product teams to
test the other dependent services, proactively plan for self-recovery options, and increase the high availability of
the products.
Use Case Example
Say your company has a product team responsible for a product recommendation engine that displays
upsell/cross-sell products on your e-commerce website. Your product recommendation engine team
could establish measures such as the recommendation engine’s uptime, response time, and error rate to
ensure they are delivering what is expected to meet the customer’s needs.
The team may also need to measure other services their product depends on, such as a database or an
authentication service. This would allow them to quickly identify any issues with these services that
could impact their own product’s performance or availability.
To ensure the continuous availability of their recommendation engine, the team could set up a testing
environment that simulates the failure of dependent services and ensures their application can still
function properly and recover itself in such scenarios. This would help the team identify and fix any
issues early on before it impacts customers.
Incorporate Improve
Product Define Success Identify Create Testing
Dependencies in Resiliency and Result
Development Metrics Dependencies Scenarios
Testing Plan Availability
• Beginning of • Teams need to • Product owner • Identified • Setup of a • Goal of the • Continuous
the process, come up with identifies other dependencies testing whole process delivery of a
when the measures of products/services are factored environment highly available
product teams success for the on which into the testing that simulates product that
are building product in their product plan. It enables various meets the
their solutions production depends the team to test situations, customer's
dependent including the needs
• For example, services and failure of
recommendation proactively plan dependent
engine's uptime, for self-recovery services
response time, options
and error rate
FIGURE 23.8: AI product testing and measurement framework
Aligning Operating Model with Strategic Value and Establishing a Clear Roadmap for
AI
Finally, when establishing the operating model, you need to ensure that it is aligned with your business objectives
and built in a manner to deliver upon the AI roadmap. You should be able to deliver iteratively, with minimal
risks and a focus on quicker time to value (see Figure 23.9).
From Basic to Advanced Functionality
Start Roadmap Deliver Enhance MVP

Align with Develop AI Deliver Essential Add Complex Develop MVP and
Business Roadmap Functionality Requirements Prioritize Key
Objectives Features
Ensure the operating Lay out the plan for AI Deliver the basic Add advanced Develop a Minimum
model aligns with the initiatives, from basic chatbot functions, like features such as Viable Product (MVP)
ACTIVITIES
company's business functionality to more responding to multilanguage with high strategic
objectives complex tasks customer queries support, personalized importance, focusing
responses, sentiment on providing accurate
analysis, and answers, high
database integration availability, reliability,
security, and reduced
downtime
BUSINESS OBJECTIVES
Align the operating model with business objectives
Deliver iteratively with minimal risks and quicker time to value
FIGURE 23.9: Aligning operating model with strategic value and establishing a clear AI roadmap
For example, suppose your team is building a chatbot. In that case, you may decide to deliver the essential func-
tionality first, such as responding to customers’ basic queries, and then add more complex requirements, such
as multilanguage support, personalization of the responses, sentiment analysis, and integration with customer
databases.
Moreover, suppose the chatbot is of high strategic importance to the organization. In that case, you may focus on
building an MVP that aims to provide more accurate answers, which would mean your operating model delivery
approach prioritizes high availability, reliability, security, and reduced downtime so as not to impact customer
satisfaction negatively.
HANDS-ON EXERCISE: IMPLEMENTING AN AI OPERATING MODEL
Goal: The goal of this exercise is to give you and your team a practical sense of how to go
about implementing some of the best practices suggested in this chapter. This exercise can be
done in a training setting, in a workshop, or even as part of a real-world project plan-
ning process.
TASK NAME TASK DESCRIPTION TASK OWNER (ROLE)
Identify Identify customer pain points and gather Business analyst

customer needs requirements for AI solutions.
Capability Assess current capabilities and gaps AI strategist

assessment against the desired AI operating model.
Develop Develop a customer-centric AI strategy AI strategy team

AI strategy to drive innovation.
Break down Break down monolithic applications into Technical architect

applications AI products.
Team Organize teams around the AI products. Project manager

organization
Iterative building Start building iteratively with cross- AI pilot team

functional pilot teams.
Continuous Implement continuous improvement Product

improvement through measurement and testing. owner, QA team
Align with Align the AI operating model with Executive sponsor, AI

strategic value strategic value and establish a clear strategist
AI roadmap.
Developing an AI Transformation Plan

The AI transformation plan is a strategic plan focusing on taking the organization forward in adopting AI at the
enterprise level.
Unlike an AI portfolio, an AI transformation plan goes beyond individual AI initiatives and instead looks at the
organization as a whole and contains a clear roadmap to increase the maturity levels of the organization in vari-
ous areas such as culture, processes, and infrastructure. It takes into consideration that for AI to succeed and to
be sustained in the long term, it needs to be integrated into the other parts of the organization, and therefore its
focus is around enabling that integration.
The next section outlines some of the components of an AI transformation plan captured in Figure 23.10.
Vision
The Vision component contains an AI transformation initiative’s overarching vision, goals, and objectives and
explains how this vision helps the organization achieve its business goals.
AI Transformation Plan – From Vision to KPI Evaluation
FUTURE
OBJECTIVES KEY
STATE VISION IMPLEMENTATION
Aims such as PERFORMANCE
Desired state of PLAN AND TIMELINE
assembling a INDICATORS (KPIs)
the organization Time taken to
cross-functional Metrics used to
after the implement each
team, measure the
implementation
01 establishing an
03 of AI 05 phase of a 07 success of the AI
AI ethical project, potential initiative, such
Transformation
framework risks, mitigation as accuracy, etc.
Plan
steps
CHANGE
CURRENT STATE ROADMAP
MANAGEMENT
ASSESSMENT Major milestones
VISION PLAN
Strengths, and phases to
Overarching goal of Manage changes
weaknesses, achieve the future
the AI
Transformation
02 opportunities, and 04 state vision such 06 and communicate
effectively during
08
threats of the as “Assess and
initiative the implementation
current AI Plan,” “Prototype
and Test” process
infrastructure,
etc.
FIGURE 23.10: Different components of an AI transformation plan

The vision outlines how the transformation initiative will impact the organization’s operations and the outcomes
it aims to achieve through this transformation initiative.
For example, the vision could be “to become a data-driven organization focused on improving productivity
and efficiency, making better decisions, and identifying new business opportunities through the adoption of AI
technology.”
These are other examples of vision:
VISION VISION STATEMENT
Thrilling customer To provide thrilling customer experiences by providing personalized and proactive
experiences solutions by adopting AI technology
Enhanced services To leverage AI technologies to provide better, more personalized services to our
and operations customers while also enhancing our own operational efficiency and risk management
capabilities
Personalized user To utilize artificial intelligence to enhance and personalize the user experience by
experience analyzing user data and behavior to recommend personalized content, improve the
streaming quality, and automate our operations
AI in content creation To use AI in content creation and production to optimize our offerings and create
and production more engaging content for our audience
Objectives
Some of the objectives could be as follows:
OBJECTIVES
AI readiness assessment To assess and document the AI readiness capabilities and gaps across the
organization
Comprehensive To develop a comprehensive strategy that includes top-level use cases for AI
AI strategy that align with the overall business strategy
Assemble AI expert team To assemble a cross-functional team of AI experts, such as data scientists,
machine learning engineers, data engineers, industry experts, and establish a
center of excellence to assist with implementing and adopting AI
Ethical AI framework To establish and implement an AI ethical framework that ensures the
responsible and ethical use of AI in accordance with social, ethical, and
regulatory considerations
Data management and To ensure data management and infrastructure plan is in place to develop and
infrastructure plan deploy AI at scale
AI deployment plan To have a plan in place to develop and deploy AI at scale to meet business
goals, deliver business value, and improve productivity, efficiency, and customer
experience
Monitoring and To have a monitoring and deployment infrastructure in place that ensures that
deployment infrastructure the AI solutions operate satisfactorily post deployment
Current-State Assessment
The plan should contain the documentation of all the findings during the current-state assessment of the AI
infrastructure process, people, and business domains. It should clearly list these strengths, weaknesses, opportu-
nities, and threats for the organization that it will face during its AI adoption journey. The plan should be well
geared to address the gaps discovered between the current-state assessment and the future vision laid out for the
organization.
Future-State Vision
The future-state vision of the company should be well documented, as well as the benefits you hope to receive
when you achieve that vision. It would depend on the use cases you identified in the Envision phase.
Roadmap
This is the most crucial part of the AI transformation plan because it documents all the significant milestones to
achieve the future state vision during this AI transformation journey. It should contain all the major phases or
milestones along with the individual projects, dates, and deliverables for each of those phases or milestones that
help address the gaps identified.
The following table shows a sample roadmap with descriptions of the major milestones:
MILESTONES DESCRIPTION
Assess and plan Includes business objectives, AI opportunities, use cases, impact and feasibility
analysis, data quality, availability and security assessment, technical infrastructure
and resources, and the project plan and funding.
Prototype and test Shortlist one to two use cases and build MVP, test MVP in a controlled
environment, gather feedback, refine it, and test again iteratively. Finally, develop
a plan to scale the solution and finalize the technical infrastructure.
Pilot and refine Release pilot in the real world, collect data, and refine the solution based on
feedback. Define a change management plan for stakeholders.
Roll out and Roll out to a broader user base, monitor impact and effectiveness, continuously
continually improve refine, and define new use cases based on experience.
Implementation Plan and Timeline

Develop an implementation plan and timeline that documents the time taken to implement each initiative phase
of a project. Document the dependencies, risks, and mitigation steps for each of those initiatives and phases. Refer
to the AI portfolio section for more information on this.
Change Management and Communication Plan

Document a clear change management plan that addresses assembling the team, aligning the leadership, and
envisioning the future. This will set the foundation that will help you implement the AI transformation plan in a
sustained, long-term manner. Refer to the change management section for more information on this.
Key Performance Indicators

The key performance indicators (KPIs) that determine the success of the AI transformation journey should be
documented here. The baseline and the target KPIs should be documented here, and the measurement framework
should also be established for periodic reporting.
You can use several metrics to measure the success of an AI transmission initiative. The following table shows
various KPIs used to measure the success of an AI transformation initiative:
KPIS DESCRIPTION
Accuracy of models Measures the percentage of accurate predictions made by the AI models.
Efficiency of models Measures the speed at which the predictions are made.
ROI Measures the return on investment, calculated as the profits generated by the AI
initiative divided by the investment made.
Customer Measures customer satisfaction, typically based on surveys, feedback forms, and
satisfaction customer reviews.
Employee Measures employee satisfaction, typically based on surveys and feedback forms.
satisfaction
Time to market Measures the time it takes to deploy a model in production. The faster, the more
competitive the organization can be.
Data quality Measures the quality of the data. The better the data quality, the better the
performance of the models.
Cost reduction Measures the cost savings from the AI effort, calculated by comparing costs before
and after the initiative.
WORKBOOK TEMPLATE: AI OPERATING MODEL AND

TRANSFORMATION PLAN
Download the “AI Operating Model and Transformation Plan” template from the download section of the book.
(www.wiley.com/go/EnterpriseAIintheCloud) Use this template to help you develop an AI O perating
Model and Transformation Plan. You can use it effectively to scale AI capabilities while minimizing risks and
maximizing ROI.
SUMMARY
This chapter took a deep dive into how to craft an AI operating model that fits businesses of all sizes. It unlocked
some of the keys for achieving seamless AI integration, such as starting small, scaling with agility, building cross-
functional teams around a product, and fostering a culture of innovation.
While the AI COE team will be responsible for developing and executing the AI transformation plan, the AI
operating model will provide the framework and the best practices to foster AI adoption at your enterprise. The
transformation plan includes vision and objectives, current-state assessment, future-state vision, roadmap, imple-
mentation plan, timeline, change management, and communication plan.
Regardless of whether you are beginning your AI journey or looking to mature your AI adoption, the lessons
hereto are pivotal in integrating AI into your business and staying ahead of the innovation curve.
In the next chapter, we will discuss implementing generative AI.
Answer Key ❘ 429
REVIEW QUESTIONS
1. What is an AI operating model?
A. A formal structure or framework to implement AI in a scaled manner
B. A model to develop AI solutions for operations
C. A way to align AI operations with business goals
D. A model to operate AI solutions without compliance
2. An AI transformation plan is
A. A plan for implementing individual AI initiatives
B. A plan to increase the maturity levels of the organization
C. A plan to develop AI technologies
D. A plan to outsource AI development
3. What are the three pillars to scale and become AI-first, as described in the chapter?
A. AI operating model, AI transformation plan, customer-centric strategy
B. AI center of excellence, AI operating model, AI transformation plan
C. AI adoption at scale, AI innovation, AI roadmap
D. AI center of excellence, AI innovation curve, AI operating model
4. Which of the following is NOT part of the AI transformation plan discussed in the chapter?
A. Vision and objectives
B. Future-state vision
C. Timeline
D. Marketing strategies
5. According to the chapter, what is vital for achieving seamless AI integration?
A. Building cross-functional teams around a product and fostering a culture of innovation
B. Focusing only on large corporations and excluding small businesses
C. Ignoring the need for strategic alignment with business goals
D. Starting with a complex plan that covers it all
ANSWER KEY
1. A 3. B 5. A
2. B 4. D
PART IX
Evolving and Maturing AI
This is where you peek into the crystal ball. I delve into the exciting world of Generative AI, discuss where
the AI space is headed, and provide guidance on how to continue your AI journey.
24
Implementing Generative AI
Use Cases with ChatGPT for the
Enterprise
The best way to predict the future is to invent it.
—Alan Kay
Welcome to the exciting world of generative AI and ChatGPT, in particular, which has been creating news
recently because of its ability to create realistic data by mimicking existing training datasets. It is like an
artist who learned the style of Picasso or van Gogh and created similar pieces of art that mimic their art
but are entirely new creations.
With every step in your enterprise AI journey, your understanding of AI and the ability to integrate AI
within your company has been increasing. It’s now an excellent time to explore the cutting-edge world of
generative AI and its implementation opportunities. See Figure 24.1.
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Your AI Govern Your Evolve and Mature
journey Team
09 • Implement Generative AI Use

Cases for the Enterprise
ENTERPRISE AI • Plan for the future of AI
Setup 04
Evolve and
Mature AI
and Manage Objective
Operations
Dive deep into the potential of Generative AI,

08 Scale and Process Data exploring its evolution, business applications,
Transform and Modeling and implementation on leading cloud
AI
Deploy and 05 platforms, while strategizing for best
Monitor practices and risk mitigation in enterprise
Models contexts
07
06
FIGURE 24.1: EVOLVE AND MATURE: Generative AI and ChatGPT use cases for your enterprise
434 ❘ CHAPTER 24 Implementing Generative AI Use Cases with ChatGPT for the Enterprise
This chapter gives you a tour of the exciting world of generative AI, starting from the innovative GANs to the
intricate diffusion models. It guides you through the technological marvels that have revolutionized industries,
from drug design to chip design.
With the strategy framework, risk mitigation plans, and other tools presented in this and other chapters in this
book, you can make generative AI applications such as ChatGPT an enterprise reality. This is where innovation
truly comes to life!
THE RISE AND REACH OF GENERATIVE AI

Even though generative AI has been creating a lot of news lately, it has not always been at the forefront of
technology until now. The story of generative AI is one of resilience and continuous innovation, starting all
the way from small statistical models to deep learning to large language models. And all of this within the last
decade or so.
But what has changed recently has been the realization that AI can be used not only to generate novel content
such as music, audio, and video, but in serious applications across industries such as drug discovery, chip design,
buildings, and so on. The impact of generative AI on the economy is vast. McKinsey has already identified close
to 63 generative AI use cases across 16 business functions that generate close to $2.6 to $4.4 trillion in economic
benefits annually across industries.
Some of the applications of generative AI have been nothing short of astounding. Take, for example, DeepArt
or DeepDream, which create lifelike, realistic artwork mimicking the style of van Gogh or Monet photos. What
is really happening behind the scenes is that a generative model has been trained based on the paintings of
these masters.
Or take the example of OpenAI’s MuseNet, which can create 4-minute compositions by using 10 musical instru-
ments and by combining new styles of music composed of a blend of country and Mozart and Beatles.
In the field of healthcare, Insilico Medicine is an example of a company that can find new drug candidates in a
few weeks, which would typically take years.
Generative AI is a fascinating field and a subset of deep learning, as shown in Figure 24.2.
NOTE Generative AI is predicted to impact almost every business, function, product,

service, and business relationship. Adopt it responsibly, ethically, and with robust testing.
The Powerful Evolution of Generative AI

Generative AI is a type of AI that uses techniques to generate new, realistic artifacts based on its learning from
existing artifacts. And these artifacts can range from audio/visual to programming and data assets, natural lan-
guage, and designs, as shown in Table 24.1.
What is essential to understand is that the artifacts produced by generative AI can span beyond just content, and
therein lies the power of generative AI.
When you have a tool that can generate not just content but also software code, music, videos, 2D/3D, art, prod-
uct designs, game worlds, process flows, drugs, building designs, and so on, you realize this is not just hype but
serious business with powerful implications for business, personal lives, and society.
Artificial Intelligence
Intelligence
ificial Tools and systems that simulate human intelligence to optimize
Art business processes and decision-making
hine Learning Machine Learning

Mac
Data-driven software algorithms that allow companies to predict
outcomes and adapt autonomously to situations
Learning
Deep
Deep Learning
Advanced ML techniques that use neural networks to analyze
massive amounts of data to identify patterns
Generative AI
Generative AI
Enterprise solutions that autonomously generate new content,
designs, or products based on learned business data patterns
FIGURE 24.2: Generative AI is a subset of deep learning
Here is the Gartner definition of Generative AI:
Generative AI enables computers to generate brand-new completely original variations of content

(including images, video, music, speech, and text). It can improve or alter existing content, and
it can create new data elements and novel models of real-world objects, such as buildings, parts,
drugs, and materials.
—Gartner
At this point, you already know what artificial intelligence, machine learning, and deep learning mean. If you are
unsure, refer to Chapter 16, where I discuss this in greater detail.
In this section, you learn how generative AI evolved over a period of time to where it is now.
Deep Learning Processes Complex Patterns Through ANNs

Recall that deep learning is a type of machine learning that uses artificial neural networks (ANNs) that enable
it to process more complex patterns than machine learning can. Artificial neural networks are inspired by the
human brain and consists of many interconnected nodes, or neurons, as shown in Figure 24.3.
Semi-supervised Learning Uses Labeled and Unlabeled Data

These neurons are arranged in many layers that allow them to process more complex patterns. And these neural
networks can process both labeled and unlabeled data. This idea of training a model using labeled and unlabeled
data is known as semi-supervised learning. This concept allows the neural network to learn using a small sample
of labeled data and a large amount of unlabeled data. While the labeled data helps the network learn the basic
tasks, the unlabeled data helps the network generalize to new examples.
TIP Use deep learning techniques only when traditional machine learning techniques are not
sufficient.
The Rise and Reach of Generative AI ❘ 437
X11
X21
X12
X13
X1[…] X22
X1[…]
X1[…]
X23
X1[…]
X1[…]
X2[…]
X1[…]
X1[…]
X1[…] X2[…]
X1[…]
X1[…]
X28
X13698
X13699
X1370
0
Input layer Hidden layer Output layer
FIGURE 24.3: Structure of artificial neural networks
Generative vs. Discriminative Models in Machine Learning

This is where generative AI comes in. It’s a form of deep learning that also uses neural networks and can process
labeled and unlabeled data using supervised, unsupervised, and semi-supervised learning techniques. As shown in
Table 24.2, there are two types of models: generative and discriminative models.
The discriminative model is focused on predicting the label for new data points after being trained on labeled
data points. Its focus is on understanding the relationships between the features and the labels.
TABLE 24.2: Discriminative vs. Generative Models
DISCRIMINATIVE MODELS GENERATIVE MODELS
Definition A type of model that tries to learn the A type of mode that tries to model on
boundary between different classes how the data is generated
Purpose Classify or predict data Generates new data based on trained

data
Training Data Labeled data Labeled and unlabeled data
Example Convolutional neural networks (CNNs) Generative adversarial networks (GANs),

Models variational autoencoders (VAEs)
Training Tries to learn the decision boundary Tries to learn the distribution of the data
Approach between classes itself
Applications Text generation, image generation Image classification, text classification,

speech recognition
Model Less intensive and complex More compute-intensive and complex

Complexity
On the other hand, generative models are focused on generating new data instances based on learning the prob-
ability distribution of the existing data.
In summary, generative models generate new data, while discriminative models try to distinguish between real
and fake data.
NOTE It is not generative AI when the output is a number or a class, meaning it cannot be
a probability or a classification. For it to be generative AI, the output needs to be in natural
language, such as text, image, audio, or video.
Foundation Models That Can Generate New Content

In the case of traditional machine learning, the classical supervised and unsupervised learning processes can take
training code and labeled data to build a model that can predict a classification or cluster something.
On the other hand, the generative AI process takes in the training code, the labeled data, and the unlabeled data
of all data types. It builds what is known as a foundation model that does not predict but instead generates new
content in the form of text, images, audio, and video, as shown in Figure 24.4.
Tasks
Question
Data Answering
Text Sentiment
Analysis
Image
Information
Training
Speech Adaption Extraction
Image
3D Signals Captioning
Object
Code
Recognition
Instruction
Foundation Model Following
FIGURE 24.4: Inputs and outputs of a foundation model
Some examples of these foundation models are the Pathways Language Model (PALM) and the Language Model
for Dialog Applications (LaMDA). These models have been trained on very large amounts of data from the Inter-
net and have built foundation models that can now answer anything by asking a question, either via prompts or
by asking verbally into the prompt.
TIP Training foundation models can be expensive, but they are impressive in performance,
and hence it’s important to put guardrails in place to ensure users are not impacted nega-
tively.
Large Language Models and Generative Image Models

You can have different types of generative models. Some would be large language models, and some would
be generative image models. Table 24.3 shows the difference between generative language and generative
image models.
The Rise and Reach of Generative AI ❘ 439
TABLE 24.3: The Differences Between Generative Language and Generative Image Models
GENERATIVE LANGUAGE MODEL GENERATIVE IMAGE MODEL
Definition A type of model that learns about patterns A type of model that produces new images
in language through training data through techniques like diffusion.
Purpose Classify or predict data Generate new data based on trained data
Training Data Text data from sources like books and Image data from datasets like ImageNet
websites and CIFAR
Example GPT-3 and GPT-4 Generative adversarial networks (GANs),

Models variational autoencoders (VAEs)
Training Tries to learn the statistical properties of Tries to learn the statistical properties of
Approach language and generates text accordingly image and generates image accordingly
Applications Text generation, chatbots Image generation, image-to-image

translation
Model Less intensive and complex More compute-intensive and complex

Complexity
Large language models are those that can understand, generate, and work with human language. The beauty of
these models is that since they have been trained on such large amounts of text data, they understand various pat-
terns, grammar, and even some real-world facts.
LLMs are considered large because they have a large number of parameters that have been learned through
the training process. The more parameters the model has, the more it can learn and remember. For example,
GPT-3 has 175 billion parameters, which is like saying they have a brain that can store about 175,000 copies
of War and Peace. So, when you ask the GPT-3 model who won the soccer World Cup in 2020, having read this
information in the past, it can immediately respond with the answer. And it gets even more interesting when you
consider that GPT-4 has 100 trillion parameters, which by the same logic means storing 100 million copies of
War and Peace in a brain.
LLMs are so huge that they demand a lot of compute power, which makes them out of reach for most enterprises
to train their own language model. Instead, the trend is to fine-tune these models for specific uses cases or indus-
tries so that the LLM can operate efficiently and provide better performance for that particular use case content.
For example, an LLM trained for healthcare will provide better accuracy for healthcare use cases.
Note that in addition to generative language models, you also have generative image models that can take an
image as input and generate text, an image, or a video.
NOTE A large language model takes in text and generates text, an image, a video, or a
decision.
Enter ChatGPT: AI Conversational Chatbot

ChatGPT is a conversational chatbot application that uses an OpenAI service to create new content. The Chat-
GPT OpenAI service serves the output from a large language model (LLM) that is trained on billions of words
from multiple sources. The model has been fine-tuned by humans with their feedback by a technique known as
reinforced learning with human feedback (RLHF).
Difference Between Foundation Models and Large Language Models

It should be noted that this is a fast-moving space, and I have attempted to capture the differences between foun-
dation models and large language models for better understanding in Table 24.4. These are subtle differences, and
the distinction between them may change as this technology and use cases evolve.
TABLE 24.4: Difference Between Foundation and Language Models
LARGE LANGUAGE MODELS FOUNDATION MODELS
Definition A type of foundation model trained on A broader category of AI systems trained

large amounts of text data, consisting on datasets of any size and includes nontext
of billions of parameters. data types and wider range of use cases and
applications
Data Types Text data Text, images, audio, video, and other sensory
Used for data
Training
Size Very large amounts of text data with Any size

billions of parameters
Primary Specifically designed for language tasks Designed to be adaptable for tasks beyond
Function such as generate text, translate text, language tasks, which could include
question answering, and summarization computer vision and robotics
Examples GPT-3, Megatron-Turing NLG BERT, RoBERTa, XLNet
Applications Text generation, chatbots Image generation, image-to-image translation
Intended Use For NLP tasks As a foundation for other AI systems
THE POWER OF GENERATIVE AI/ChatGPT FOR BUSINESS

TRANSFORMATION AND INNOVATION
This section discusses the role that generative AI and ChatGPT can play in businesses, products, services, and
relationships. Because of its ability to generate new and original content, generative AI can be a game-changer for
many industries. This section explores a few possibilities. Generative AI can enhance efficiency, innovation, and
personalization through a number of applications.
Business Transformation with Automated Content Generation

The ability to generate automated content on a regular basis, such as blogs, articles, and social media posts, is
a huge advantage for businesses. It has enormous implications for content marketing applications, as shown in
Figure 24.5.
Thought Leadership Through Improved Content Accuracy

Because gen AI trains on a large amount of data, it can look at patterns that a human eye cannot see and there-
fore can help produce high-quality content, enabling organizations and individuals to establish themselves as
thought leaders in their space.
The Power of Generative AI/ChatGPT for Business Transformation and Innovation ❘ 441
Generating different types of content
Providing feedback on content
Personalizing content for different audiences

Use Cases
Optimizing content for search engines
Bias
Detecting and removing plagiarism
Lack of transparency
Challenges Translating content into different languages
Intellectual property concerns
Job displacement
AI-Powered Content Personalized search results
Increased speed and efficiency Generation Google Google News
Improved accuracy Google Translate
Reduced costs Benefits Product descriptions

Amazon
Increased reach Personalized recommendations
Personalized content Newsfeeds

Case Studies Facebook
Personalized ads
Trending topics
Twitter
Personalized recommendations
Job postings
Linkedin
Personalized recommendations
FIGURE 24.5: The AI-powered content generation use case
Innovation with New Business Ideas, Art, and Product Design

Regarding innovation, generative AI can create new business ideas, product designs, digital art and music, and
chemical structures for potential new drugs.
By creating new and varied content that includes images, audio, videos, and text, companies can engage with their
customers and keep them captivated with their offerings.
Another example of an innovative application of generative AI is Pictionary, which uses AI to guess what you are
drawing. Behind the scenes, it uses millions of doodles from people worldwide to train itself.
In the case of product design, you have to generate many product design ideas, which would involve a lot of
brainstorming, sketching, iterations, and, finally, shortlisting them. This would be a tedious, manual process and
would consume a lot of time. On the other hand, say you are feeding hundreds of product designs from the past
into a model, and the model returns new and innovative product designs that meet your product specifications
and quality requirements. That is precisely what Autodesk’s Dreamcatcher solution tries to do.
Autodesk Dreamcatcher is like having thousands of designers trying to create new innovative designs day by day,
a round the clock.
Revolutionizing Experiences Through Personalized Emails and Product

Recommendations
Concerning personalization, generative AI can help create different versions of personalized emails for different
customers and recommend products based on past purchases. For example, using AI algorithms, Spotify recom-
mends personalized playlists for millions of users based on their individual tastes every week. Another example is
Fractal Analytics, which uses generative AI to create new landing pages and generate emails.
Innovative Product Features to Boost Customer Loyalty

Another area where generative AI can help you is in personalizing your products based on exactly what the users
need. This can help companies be laser-focused on meeting their customers’ user experience needs and drive cus-
tomer engagement and loyalty.
Impact on Personal Life Content Generation
Process Improvement Improving Content Accuracy
Generative AI
Workforce Optimization Generating New Product Ideas
Applications
Employee Collaboration Personalizing User Experiences
Transforming Customer and Improving Product Features

Supplier Relationships
FIGURE 24.6: Generative AI applications
Transforming Customer Relationships Through Chatbots, Personalization

Generative AI has been instrumental in transforming customer relationships by analyzing customer behavior,
preferences, and feedback to develop new product recommendations and responses and even create new prod-
ucts. This level of personalization can help engage with customers better and promote customer loyalty in the
long term.
Another example is the application of generative AI to build a chatbot. Instead of having preprogrammed rules to
answer the customer’s question, you can have a generative AI chatbot similar to ChatGPT to answer a customer’s
questions in real time based on their style of language, preference, sentiment, and type of question.
Breaking Language Barriers, Better Supplier Relationships, and Global Trade

In the case of supplier relationships, using the natural language translation capabilities of gen AI, you can com-
municate with suppliers across multiple geographies and countries. This ability helps you create better partner
relationships and promotes global trade.
Reimagining Workplace Efficiency with AI-Powered Assistants

Generative AI can also collaborate with internal employees to automate tasks and improve process efficiency and
employee productivity. For example, GitHub’s Copilot is an AI-powered assistant that can help developers to
develop code faster. This opens up new opportunities because this concept of having a virtual assistant that can
work with humans to improve employee productivity will be a game-changer when automating many of the busi-
ness processes within the company.
Long-Term Workforce Transformation with AI as a Partner

Figure 24.7 shows how AI can be a valuable partner that enables long-term transformation in the workforce. The
ability of the employee to work with AI will not only enhance their productivity. It will also enhance the ability
to come up with new ideas with the help of the AI assistant, which can act as a creative partner. It now becomes
possible to develop groundbreaking ideas and efficient execution strategies.
From Process Transformation to Life-Changing Applications

Generative AI has the potential to uncover hidden patterns and insights from large amounts of data that were
not possible before because of the sheer size of the data. Armed with these new insights, gen AI can also propose
changes to current workflows to eliminate process inefficiencies, remove bottlenecks, and automate tasks.
The Power of Generative AI/ChatGPT for Business Transformation and Innovation ❘ 443
Enhanced productivity, idea Automated task (coding, data

generation, and performance analysis, document drafting,
etc.)
Employee Transformation
VIRTUAL Task Automation
ASSISTANT
New Opportunities GitHub CoPilot
Generation of new ideas and Paired Programmer

process automation improved productivity
opportunities
Long-Term Workforce
Transformation
FIGURE 24.7: Long-term workforce transformation with AI as a partner

Source: Blue Planet Studio/ Adobe Stock
The uses of generative AI are not restricted to businesses alone but can also impact our personal lives. Take, for
example, the Stanford University research, where they converted human brain signals into written text.
TIP Use generative AI not to replace creativity or other relationships but to augment your
capabilities and open new possibilities to transform business and personal lives.
The Fascinating World of Generative AI: From GANs to Diffusion

Models
Now that you know what generative models do, you probably wonder how they work. Even though there are a
number of generative models out there, three models stand out: generative adversarial networks (GANs), vari-
ational auto-encoders (VAEs), and transformer models. Despite their differences, they all follow the same princi-
ples: to learn the underlying patterns in the data and re-create data with similar data patterns.
GANs: Interplay of Generator and Discriminator Networks

The GANs model was invented by Ian Goodfellow and his team in 2014. It is based on the beautiful concept of
two neural networks working against each other. One is called the Generator, which creates realistic data, such
as images, based on the data it was trained upon. The other is the Discriminator neural network, which tries to
critique the new image and find the difference between the real image and the generated version. This constant
tussle between the two networks results in an incredibly realistic image.
TIP Use GANs to create realistic human faces and even photorealistic artwork out of
sketches.
VAEs: A Solution to GAN’s Black-Box Problem

One of the drawbacks of GANs is that they operate like black boxes. We don’t know how these models work.
In such situations, VAEs can help because we can understand the hidden layers that drive the data. It is used in
product recommendation engines and text generation use cases.
Transformers: The Behind-the-Scenes Hero of GPT, BERT

Transformers are the model behind models such as GPT-3 and BERT, which revolutionized text generation,
summarization, translation, and more. Transformers were introduced in 2017 and have revolutionized the world
of natural language processing because they simultaneously consider all the words in a sentence when trying
to make a prediction. We already know that ChatGPT has taken the world by storm, as more than 100 million
people registered for the tool in just a couple of weeks. When using ChatGPT, it feels almost like talking to a
human, which is impressive. Another example of using a transformer model is when people use Gmail, where the
autocomplete feature is based on a Transformer model. As shown in Figure 24.8, a transformer model contains an
encoder and decoder.
NOTE Transformers have an encoder and a decoder, where the encoder encodes the input
sequence and passes it to the decoder, which learns how to decode the representation for a
relevant task.
Transformer
Input Generative Pretrained Output
Encoding Decoding Transformer
Component Component Model
How I am doing
are fine and
you? you?
FIGURE 24.8: Components of a transformer model
Hallucinations: When Models Go Astray

Sometimes these models can generate outputs that are nonsensical or grammatically incorrect. This is known as
hallucination and can happen when the model has not been trained adequately or on enough data or is trained on
noisy data or has not been given enough context or constraints to work with.
From Text-to-Image to Text-to-Task Using Diffusion Models

While GANs and VAEs have held the limelight for the most part, diffusion models have also been making waves
recently. Diffusion models take a unique approach to creating new data. They take the original image (the picture
of a cat) and add noise to it to distort it slightly. They continue this process until it becomes unrecognizable. Then,
it tries to recover the image by removing the noise with each step. In this process, it ends up creating a range of
new cat images depending on the noise with which it started the recovery process.
NOTE Diffusion models can generate text-to-image, text-to-video, text-to-3D, and text-to-
tasks.
Implementing Generative AI and ChatGPT ❘ 445
Text-to-3D can be used in video games, and text-to-task models can be used to carry out tasks such as ques-
tion and answer, performing a search, or taking some action such as navigating a UI or updating a document.
Table 24.5 captures the differences between GANs, VAEs, and diffusion models.
TABLE 24.5: Differences Between GANs, VAEs, and Diffusion Models
GENERATIVE GANS VAES DIFFUSION MODELS

MODEL
Working Two neural networks A type of neural network Adding noise to the original
Principle competing in a zero-sum that learns to encode image and then removing
game to generate new data and decode data by the noise gradually to create
that closely resembles the mapping it to a latent a range of new images
training dataset space depending on the noise with
which the recovery process
started
Applications • Image to image • Image and video • Text-to-image

translation generation • Text-to-video
• Scenes and people • Data compression • Text-to-3D
• Realistic human faces • Anomaly detection • Text-to-tasks
• Artwork
• Photorealistic photos of
objects
Examples • StyleGAN • Beta-VAE • DALL-E

• BigGAN • VQ-VAE • CLIP
DALL-E and CLIP are two diffusion-type models from OpenAI. DALL-E can generate images from text descrip-
tions, while CLIP can understand both text and images and can relate them to each other. For example, you can
ask DALL-E to create an orange in the form of an apple, and DALL-E will precisely do that to create a novel
image that was never seen before.
Diffusion models are not without their challenges because they can be very computationally intensive and slower
than GANs. Despite their challenges, diffusion models are fascinating and extend the limits of virtual reality and
what’s possible with Gen AI.
IMPLEMENTING GENERATIVE AI AND ChatGPT

Given its many risks, implementing generative AI and ChatGPT for your company is not a trivial task. Moreo-
ver, it is a constantly evolving space; you must keep close tabs on what’s happening out there. Having said that,
consider the different options you have when implementing generative AI or ChatGPT. Your options will span
between buying off-the-shelf solutions and building them on your own. Figure 24.9 shows the different options
laid out in front of you.
From Packaged Software to Pretraining Models

Consider these build versus buy options that are available when implementing generative AI:
➤➤ Out-of-the-box options: You already use the out-of-the-box option when you use ChatGPT for various
natural language processing tasks such as summarizing, answering questions, translation, and so on.
The issue with that approach is that you need to figure out how your data is going to be leveraged by
those companies.
➤➤ Consuming APIs: When you use APIs, cost is a critical factor because you are charged based on the
tokens, you use both as a query to the API and the response you get back. It is therefore important to
exercise controls over the amount of data input to the API as well as the amount of data generated
from the API.
➤➤ Prompt engineering: The other option is prompt engineering, which has been gaining much attention
lately. We are already seeing innovations in this space in the form of new models such as hugging space,
GPT long chain, and so on.
➤➤ Fine-tuning the LLMs: The next best option that gives you even more control is fine-tuning the LLM
model, which requires technical expertise from your company. The last option is very complex: pretrain-
ing a model, which is almost like starting from scratch; it’s costly and therefore is ruled out for most
companies.
➤➤ Using a packaged software with embedded gen AI: The easiest option is to use the packaged software
where you don’t need any technical team; anybody can use those. Examples of those are ChatGPT,
Microsoft Bing, Google BARD, Adobe Firefly, and so on. In the future, we are going to see most of these
large language models embedded in enterprise applications
Make API Calls Use Prompt Design
Out of the Box Wen

03 04 Fine Tune LLMs
Apps
02 05
Leveraging
LLMs
01 Build vs Buy 06
Use Software with Building New
Gen AI embedded Models
FIGURE 24.9: Build versus buy options when implementing generative AI
TIP Fine-tuning pretrained models can be a viable middle path between fully customized
models and off-the-shelf products.
Bridging the Gap Between External and Internal Knowledge

Implementing generative AI boils down to connecting the external knowledge gathered by these large language
models with the knowledge that you have in your organization. As shown in Figure 24.10, some of these tech-
niques are as follows:
➤➤ Prompt engineering
Implementing Generative AI and ChatGPT ❘ 447
➤➤ Customized fine-tuning
➤➤ Retrieval augmented generation (RAG)
Prompt controls the

following:
Small piece of
LLM output format,
text sent as input Prompt such as the text, image,
to the LLM
or code
LLM output style such as

Process of formal or informal,
creating the right creative, or informative
Prompt
prompt to get the Design LLM
desired response Context of the output such
as providing background
information
Optimizing the Scope of the output such as

prompts so we Prompt restricting to a specific domain
can get the best Engineering or topic
out of the LLMs
FIGURE 24.10: Using prompts to train LLMs
Enhancing LLM/ChatGPT Interactions with Prompts

A prompt is a small piece of text sent as input to a large language model and can be used to control the output in
many ways.
Prompt design is a term used to describe the process of creating the prompt to elicit the desired response from
the large language model. Think of prompt as not only a basic user request but also adding some background,
instructions, and user context to the request so that the model can respond much more accurately.
Prompt engineering is the practice of optimizing prompts so that you get the best out of these large
language models.
Improving Response Quality with Retrieval Augmented Generation

Retrieval augmented generation (RAG) is a powerful technique that combines the power of information retrieval
and the NLP capabilities of a large language model, as shown in Figure 24.11.
For example, consider a chatbot in a banking scenario. You can use RAG to enhance the quality and relevance
of your responses. Say a customer wants to know about the status of their transaction. You can break down the
process into three steps.
1. Retrieval: During the retrieval process, the chatbot looks in the customer database and retrieves cus-
tomer account information along with other relevant information related to the request from a large
corpus of text.
2. Augmentation: The chatbot then combines the user request with the information it just retrieved from
various sources and sends it to the large language model to compose a response.
3. Generation: The large language model uses the information in the user request, which contains real-
world information to generate a creative, user-friendly, and relevant response to the customer.
Internal Company Response Generation

User
Repository
Files
User
Request
Prompt Engineering
Folders
Information
Retrieval
Aggregation
Database LLM Model
Databases
Instructions
FIGURE 24.11: Retrieval augmented generation architecture
RAG has the following benefits:

➤➤ It improves the relevancy and accuracy of the information response as it’s based on real-world data.
➤➤ It saves time and effort as it saves human effort to find relevant information from various sources
related to the incoming request.
➤➤ Since you are using a single LLM, the response is always consistent.
The RAG approach has been used in various applications such as customer service, customer support, news writ-
ing, and even therapy.
RAG is also used in certain help desk use cases, where the language models can make their own API calls to other
external systems, given their ability to create their own code. This allows them to create their own tools and then
use that to respond to the user with suggestions.
TIP When integrating internal data with external data, pay attention to data privacy issues,
access control security, and leverage domain experts to tailor models efficiently.
BEST PRACTICES WHEN IMPLEMENTING GENERATIVE AI AND

ChatGPT
This section discusses some best practices to follow when implementing Generative AI.
Ensuring Performance and Mitigating Risks Through Internal Pilots

First, it is essential to test your generative AI and ChatGPT use cases internally with your employees before releas-
ing them to external customers. This will help you catch any issues with the model’s performance, especially for
things like hallucinations, which can cause severe damage to your reputation.
Communicating the Role of Generative AI in User Interactions

Second, it is imperative to be transparent with the users, whether they are employees, customers, or the public,
and admit that when interacting with the application, they are working with a machine. This helps you build the
necessary trust and avoid any misgivings in the future.
Best Practices When Implementing Generative AI and ChatGPT ❘ 449
Ensuring Fairness and Bias Mitigation Through Continuous Testing and Guardrails
Third, set up the necessary guardrails to check for biases in the system and address them accordingly. Keep testing
continuously to ensure they are working accurately.
Protecting Data Privacy and Security of Organizational Data

Ensure data privacy and security by talking to the model provider and confirming that the data input into the
model will not be used outside the organization.
Building Trust Through Beta Testing

It is also vital to ensure that the applications are kept in beta mode for a sufficient amount of time and are
released to the general public only once there is complete confidence in the accuracy of the models and the issues
have been addressed.
TIP Be prepared to manage risks proactively: starting with a pilot, maintaining transparency
with users, looking out for bias, handling data privacy issues, and planning for contingencies
are vital.
Strategy Considerations for Generative AI and ChatGPT

This section discusses the best practices to consider when implementing generative AI and ChatGPT, as shown in
Figure 24.12.
Strategy Best
Practices for
Generative AI
Data &
Disruptive Agile Shorter-term Predictive to
Automation to analytics to Talent
potential of platforms focus for generative use
augmentation enterprise- upskilling
generative AI and tools roadmap cases
wide focus
From 3-year
Process Flexible outlook to 1-
Enterprise-
automation to architecture, year outlook, Tools impact,
Update AI wide
smart integrate with monitor process
strategy with governance, AI literacy
automation to open-source frequently and impact, people
Gen AI ethics
people and licensed frequent impact
framework
augmentation software strategy
updates
FIGURE 24.12: Strategy best practices for implementing Generative AI
Risk of Disruption
Generative AI and ChatGPT can disrupt the industry you are currently in. This could mean that you need to
change your current AI strategy or complement it with generative AI and ChatGPT related strategy.
Shifting from Automation to Augmenting Humans with AI Assistants

From a vision perspective, this could mean you need to move from automating processes to smart automaton that
uses AI to augmenting people with ChatGPT. So, the focus could move from displacing humans to having virtual
assistants that act like managers or co-workers.
NOTE Virtual assistants can help employees make decisions at critical junctures, cope with
stress during the day, mentor employees, and be impartial in their interactions with the
employees.
Agile Roadmapping to Embrace Continuous Strategy Refinement

From a roadmap perspective, you need to transition from a longer three-year outlook to having an annual
roadmap, given the rapid change in technology. You need to review your strategy more frequently and refine it
continuously.
Agile and Flexible Platforms for Future-Proof Infrastructure

Greater care needs to be exercised in the choice of platforms and tools given that you cannot change them easily,
and the underlying infrastructure should be agile enough to accommodate new changes and integrate them easily
with open source and licensed products.
Shifting from Prediction to Generation of New Artifacts

The use cases will change from predictive and automation type to generating new types of artifacts such as text,
videos, images, audio, and code. As a result, the corresponding tools, processes, and skills will focus on generating
new artifacts rather than predicting outcomes.
Expanding the Scope from Data and Analytics to Enterprise-wide Governance

Given the pervasive nature and impact of these generative AI solutions, you need to ensure that the AI strategy
evolves from being just a narrow data and analytics focus to being an enterprise-wide strategy that incorporates
governance aspects of generative AI/ChatGPT. Since this can have cultural implications, all responsible parties
should be represented in these discussions, and an AI ethics committee may become more urgent.
Democratizing AI to Empower the Workforce with AI Literacy

The focus will now shift from having AI skills concentrated in a data science lab team or an AI center of excel-
lence to making the employees AI literate across the enterprise.
TIP Anticipate disruption and bake it into your business strategy to create a competitive
edge.
Challenges of Generative AI and ChatGPT

This section discusses the challenges you need to overcome when adopting generative AI, as shown in
Figure 24.13.
Best Practices When Implementing Generative AI and ChatGPT ❘ 451
Risks of disruption and addressing

01 deepfakes and misinformation
Navigating against deception

02 and fraud
Balancing surveillance and

03
ethical use of data
Challenges of Affordability for small and
Implementing 04 medium businesses
Generative AI
Integration with existing
05 workflows and processes
Lack of model transparency and

06 integration with technology
Navigating job displacement and

07 social unrest
FIGURE 24.13: Challenges of implementing generative AI
Risk of Disruption
Large industry players run the risk of being disrupted by smaller players.
Addressing Deepfakes and Misinformation

The significant challenge of generative AI is the ability to create hyperreal audio, videos, and images that can
deceive people. We have already seen a few incidents where people have been deceived, which has resulted in
monetary fraud. Work with your cyber insurance providers to ensure you are covered for AI-related breaches.
Navigating Against Deception and Fraud

Another major issue of generative AI is that of bias. Because these models are being trained on biased data, they
also generate biased results. You need to have policies or controls in place to detect when they happen and deal
with them appropriately based on company policies and legal requirements.
Addressing the Potential for Incorrect Decisions due to Hallucinations

Depending entirely on the outcomes from the generative AI and ChatGPT models can be risky and lead to
incorrect decisions, as sometimes these models can generate false information due to the hallucination effect
discussed earlier.
Affordability for Small and Medium Businesses

Another primary concern is the need for high computing resources, which makes generative AI unreachable for
small and medium-sized businesses.
Embracing Energy Efficiency and Renewable Solutions

There is also this concern about the vast environmental impact because of the large amounts of energy consumed
to build and train these models. You must choose vendors that save energy consumption and use renewable
energy more.
Integration with Existing Workflows and Processes

Another challenge is incorporating generative AI and ChatGPT use cases into existing workflows and processes.
Ethical Handling of Private Data

Generative AI uses a lot of private data to generate new synthetic data for training, especially in the case of
healthcare. While the patient’s information can be masked, getting the appropriate consent and being aware of
privacy concerns is still essential.
Balancing Surveillance and Ethical Use of Data

Another privacy-related concern is that generative AI can be used to monitor people’s activities and movements
and even predict their future behavior, which can be a serious cause of concern.
Protecting Intellectual Property and Copyright Concerns

Users should be careful about inputting their personal or company information. They should assume that this
information will become publicly available. Therefore, companies should implement controls to ensure their IP
address is not compromised.
Lack of Model Transparency

How ChatGPT or generative AI models work is unpredictable, and even the companies behind them do not know
how they work.
Integration with Technology

Integrating with existing technology and systems can be a challenge.
Navigating Job Displacement and Social Unrest

Generative AI and ChatGPT can do some of the jobs that humans can do, and this can potentially displace people
from their jobs and cause social unrest. It is especially true for creators of content and code.
Safeguarding Humanity from Potential Harmful Models

It might sound like science fiction, but it is certainly possible for those seeking power over humanity to build
power-seeking AI models whose goals are not aligned with humanity and could cause significant harm.
TIP Plan on investing in advanced deepfake detection techniques to mitigate the impact.
Strategy for Managing and Mitigating Risks

Just as with any new technology, generative AI has its own challenges, and you need to implement some mitiga-
tion steps to address those risks, as shown in Figure 24.14.
HANDS-ON EXERCISE: MANAGING GENERATIVE AI RISKS
Goal: The goal of this exercise is to help those involved in implementing generative AI to
manage and mitigate risks.
Generative AI Cloud Platforms ❘ 453
Understand security and Security team, Risk assessment

ethical risks: Document risks legal team, risk document
related to security, privacy, deep compliance team
fakes, ethics, bias, etc., for safe
generative AI implementation.
Build AI risk awareness: Educate HR team, training Training materials,

employees on the benefits and team employee
risks of generative AI, including acknowledgment forms
data handling techniques.
Define ethical guidelines: Ethics committee, Ethical guidelines

Establish policies for data legal team document, usage
privacy, ethics, fairness, policies
transparency, and security.
Address bias and ensure IT team, security Bias mitigation report,

cybersecurity: Implement tools team cybersecurity protocols
and measures to detect/address
bias and enhance cybersecurity.
Make energy-conscious Infrastructure team Energy consumption

choices: Select energy-saving report, selected
infrastructure for model training. infrastructure details
Continuous monitoring: Monitor Monitoring team, Continuous monitoring

the established controls and compliance team reports, adaptation
adapt as needed. strategies
Understanding Risks
Continuous Monitoring and Building AI Risk Awareness

Adaptation
Making Energy Conscious Risk

Responsible Data Handling
Choices Mitigation
Addressing Bias and Defining Ethical Guidelines and

Cybersecurity Policies
FIGURE 24.14: Mitigating generative AI risks
GENERATIVE AI CLOUD PLATFORMS

When it comes to generative AI from a cloud perspective, you need to have a lot of compute power to train these
large language models. By training these large language models with your own company data, you can use them
to answer customer queries based on your company data. The major cloud providers provide the ability to train
these models within your company’s premises without losing control over the data and exposing it to the Internet
by taking all security precautions and best practices. In addition to the cloud platform, you also need develop-
ment platforms that can help you build generative AI applications on top of them.
Google’s Generative AI Cloud Platform Tools

Google provides several generative AI cloud platform tools, namely, Gen AI Studio, Google AI platform, and
TensorFlow.
Simplifying AI Model Development and Deployment with Generative AI Studio

Generative AI Studio is a good option for developers and users who want to develop generative AI models with
less to no machine learning coding knowledge. It is a web-based platform with drag-and-drop features that users
can use to build and deploy generative AI models. It makes exploring and customizing gen AI models, building,
and deploying new models, and experimenting with different settings easier.
Generative AI Studio allows users to interact and train large foundation models from Google, such as PaLM 2,
Imagen, Codey, and Chirp, to develop generative AI applications, including text, chat, text-to-image generation,
and code completion.
NOTE When combined with the Model Garden provided in Google’s Vertex AI platform, it
can be a game-changer for software development agent generative AI applications.
Figure 24.15 shows the foundation models Language FMs Vision FMs
available in the model garden. One promising
application of generative AI using Google tools • BERT • Stable Diffusion v1-5
is Google’s Magenta project, which is an applica- • PaLM API for Chat • CLIP
tion that can generate unique music in a par- • PaLM API for Text • Embeddings Extractor
ticular genre or style based on the music shared • BLIP Image Captioning
by the user. • BLIP VQA
• OWL-ViT
No Code AI Development with • ViT GPT2
Generative AI App Builder
Google’s Generative AI App Builder allows you to FIGURE 24.15: Foundation models in the model garden
create generative AI applications without coding.
According to Google, it helps businesses and gov-
ernments to build their own AI-powered chat interfaces and digital assistants. It connects conversational AI flows
with out-of-the-box search experiences and foundation models and thus helps companies to build gen AI apps in
minutes or hours.
Large-Scale Deployments Using the Google AI Platform

Google’s AI platform is also an excellent tool to build, deploy, and scale generative AI models. Still, when com-
pared to Generative AI Studio, it is more complex and needs some machine learning skills. However, it can be
used for a larger-scale deployment and has more features.
Greater Control and Flexibility with TensorFlow

If you want greater flexibility and control over your generative AI machine learning process, TensorFlow will be
your choice. It is an open-source software library developed by Google. Many state-of-the-art machine learning
models, such as GANs, have been developed using TensorFlow.
Coding Made Easy with Google Bard and Colab Tool

You can use Bard to generate new code and convert it from one programming language to another. And if you
use Google’s free browser-based Jupyter Notebook, known as Colab, you can export the code from Bard directly
into the Colab tool. Bard can help you debug the code, explain the code line by line, generate SQL queries to a
database, and even create documentation and tutorials.
AWS Generative AI Cloud Platform Tools

This section discusses the AWS Generative AI Cloud Platform tools shown in Figure 24.16. It covers issues like
what are the specific features and capabilities of AWS SageMaker, AWS CodeWhisperer, AWS Bedrock, and
AWS Titan Models? How do these services work together to create generative AI applications? What are the
specific benefits and use cases for AWS Trainium and AWS Inferentia2 instances in the context of generative AI
applications?
All-in-One Solution for

Training and Deploying
Generative AI Models
AWS
SageMaker
AWS
A Family of Generative Code Faster and
Pretrained AWS Titan AWS Smarter with a
Models
AI Cloud
Foundation CodeWhisperer Coding
Models Platform Companion
Tools
AWS
Bedrock
Building Secure and

Scalable Generative AI
Applications
FIGURE 24.16: AWS Generative AI cloud platform tools
AWS SageMaker: All-in-One Solution for Training and Deploying Generative AI Models
SageMaker is another excellent tool to train and deploy generative AI models. It provides a comprehensive devel-
opment environment to build, train, and deploy generative AI models. And it comes with its own debugger and
other monitoring tools to manage bias, explainability, and so on.
Code Faster and Smarter with AWS CodeWhisperer

AWS CodeWhisperer is a coding companion that can increase the productivity of developers. Today’s develop-
ers are challenged with repetitive coding, and constantly changing programming technologies prevent them
from focusing on the creative side of programming. Imagine now as a developer, you can ask CodeWhisperer to
provide the code to parse a CSV string of songs and return a list of songwriters, titles, ranking, and so on. Like
magic, CodeWhisperer will return the entire function to you with all the code in the programming language of
your choice. During a recent challenge, it was found that CodeWhisperer can help developers complete their cod-
ing 57 percent faster, and 27 percent were successful.
CodeWhisperer is available for Python, Java, JavaScript, TypeScript, C#, Go, Kotlin, Rust, PHP, and SQL. It can
be accessed from IDEs such as VS Code, IntelliJ IDEA, AWS Cloud9, and many more via the AWS Toolkit IDE
extensions. Having been trained on billions of lines of code, CodeWhisperer can identify security vulnerabilities
such as Open Worldwide Application Security Project (OWASP) best practices and is an effective tool for handling
open-source coding responsibly. Figure 24.17 shows the different steps involved when using AWS CodeWhisperer.
Collaboration Set Up
and Version CodeWhisperer
Control
Integrate and Launch

CodeWhisperer
Test and Debug
AWS
CodeWhisperer
Select the
Programming
language
Identify Security
Vulnerabilities
Receive Ask CodeWhisperer

Customize and Code for Code
Adapt Code Suggestions
FIGURE 24.17: AWS CodeWhisperer workflow
NOTE AWS CodeWhisperer is a game-changer because anyone with no coding experience

or skills can now use it to create software code.
Building Secure and Scalable Generative AI Applications with AWS Bedrock

Amazon Bedrock is a new service offered by Amazon for builders to access foundation modules and include them
in their generative AI applications. For example, you can access FMs from AI21 Labs, Anthropic, Stability AI, and
even Amazon FMs such as Titan via an API. Because this is a serverless service, you do not have to worry about
the underlying infrastructure. All you need to do is choose the suitable model, customize it with your own data,
then deploy it and incorporate it into your applications using familiar tools such as AWS SageMaker. Using AWS
SageMaker implies that you can test different models using SageMaker Experiments and also use SageMaker
Pipelines to scale your FMs.
TIP One of the most significant benefits of using AWS Bedrock is that you can create a
private copy of the foundation module and then use your data within your own private
VPC; therefore, your data security concerns are well addressed.
Titan Models: Amazon’s Generative LLM and Embedded LLM

In addition to AWS Bedrock, Amazon has come up with two large language models; one is a generative LLM,
and the other is an embedded LLM. While the generative LLM can be used for tasks such as summarization,
text generation, question and answering classification, and information extraction, the embedded LLM cannot
generate new content but instead generates numerical representations known as embeddings. These embeddings
can be used for use cases such as product recommendations and searches. By comparing these numerical embed-
dings, the model can provide more accurate, relevant, and contextual responses that are impossible with just
word matching.
Another significant advantage of Titan models is that they can filter harmful content in the input data, the user
input, and the output generated by the model. It makes it very usable in different use cases, such as customer
service, finance, and education.
Using both Bedrock and Titan models, companies can customize these models for various use cases, such as
answering customer queries using a chatbot, creating interactive course content for students, generating patient
reports, and diagnosing diseases for patients, creating targeted ad copy and emails for customers, and detecting
fraudulent transactions.
Optimizing Performance and Cost Using EC2 Trn1n and Inf2 Instances
Now say you were working with the foundation modules, and you’re looking for a high-performing and cost-
effective infrastructure. That’s where Amazon EC2 Trn1n and Inf2 instances powered by AWS Trainium and AWS
Inferentia2, respectively, will come into play. You can use Trainium instances to reduce the cost of training. In
contrast, you can use Inferentia instances to reduce the cost of making predictions. As of this writing, AWS Train-
ium chips are built to distribute the training across multiple servers employing 800 Gbps of second-generation
Elastic Fabric Adapter (EFA) networking. It allows you to scale up to 30,000 Trainium chips, which equals more
than 6 exaflops of computing power, all within the same AWS Availability Zone. It is equal to an impressive
petabit-scale networking capacity.
AWS Inferentia chips are optimized for models with hundreds of billions of parameters. As of this writing, each
chip can deliver up to 190 tera floating operations per second (TFLOPS). On the other hand, AWS Trainium is
a second-generation ML accelerator for deep learning training of 100B+ parameter models. Amazon Search and
Runway use Trainium to train their models to generate text, translate languages and answer questions. They can
reduce the time from months to weeks or even days while reducing the cost. Money Forward uses it to train its
models to detect fraud. Figure 24.18 captures the features of AWS Inferentia and Trainium.
Azure Generative AI Cloud Platform Tools

This section reviews the generative AI cloud platform tools offered by Azure.
Transforming NLP Using the Microsoft Azure OpenAI Service

Microsoft Azure machine learning is another excellent tool for building and deploying generative AI models. It
comes with its own prebuilt model registry. The Azure OpenAI service allows you to access a number of their
models, such as ChatGPT, Codex, and Embedding models. You can access the NLP services these models offer
using APIs, SDKs, and Azure OpenAI Studio.
AWS Inferentia2 AWS Trainium
• Deploying models with • Training large language models

hundreds of billions of (LLMs)
parameters • Training mixture of experts
• Optimized for models with (MoE) models
hundreds of billions of parameters • Ability to distribute training across
• Used by Helixon, Amazon Search, multiple servers
and Runway • 800 Gbps EFA networking
• Reduced training time • Scaling up to 30,000 Trainium
• Used by Money Forward for fraud chips: equal to more than 6
detection exaflops of computing power
• Same AWS Availability Zone
FIGURE 24.18: Features of AWS Inferentia and Trainium
ChatGPT uses generative AI models to generate new content in response to users’ natural language prompts. The
OpenAI service allows you to leverage these services from ChatGPT and combine them with the security and scal-
ability of the Azure cloud platform. Codex is an AI system developed by OpenAI to develop code from text.
The first step to building a generative AI solution with Azure OpenAI is to provision an OpenAI resource in your
Azure subscription, as shown in Figure 24.19.
FIGURE 24.19: Provisioning an OpenAI service in Azure

NOTE At this point, the OpenAI Azure service is available with limited access to ensure its
ethical use. Certain models are available in certain regions only.
Effortless Gen AI Development Using Azure OpenAI Studio

Azure OpenAI Studio provides access to building, managing, and deploying models. It’s a cloud-based service that
makes building gen AI applications using OpenAI’s APIs easy. It is a preconfigured environment that comes with
a GPU-powered compute instance, a Jupyter notebook, a collection of pretrained models, and access to OpenAI’s
API documentation.
To access OpenAI Studio, you need to create an account, launch the Studio instance, and then start building by
importing a pretrained model or by training your own model using the Jupyter Notebook. Once done, you can
deploy your application to Azure Cloud using the OpenAI API.
NOTE Azure OpenAI Studio is a great way to build your gen AI apps without worrying
about the underlying infrastructure.
Customizing Models in Azure OpenAI Service

When building apps with Azure OpenAI service, you need to choose a base model, and then you can customize it
if required. These models are grouped by family and capability, as shown in Table 24.6.
TABLE 24.6: Base Models Grouped by Family and Capability in Azure
FAMILY DESCRIPTION BASE MODELS WITHIN THE FAMILY
GPT-4 Models that generate natural language and code. gpt-4, gpt-4-32k
These models are currently in preview.
GPT-3 Models that can understand and generate natural text-davinci-003, text-curie-001, text-
language. babbage-001, text-ada-001, gpt-35-turbo
Codex Models that can understand and generate code, code-davinci-002, code-cushman-001
including translating natural language to code.
To make API calls to the model, you must deploy it. Note that you can deploy only one instance of the model.
Once deployed, you can test it using the prompts. A prompt is the text input that is sent to the model’s endpoint.
Responses can come in the form of text, code, or other formats.
Prompts can be of various types, such as classifying content, generating new content, holding conversations,
translating, summarizing content, picking up where you left off, and getting facts.
Azure OpenAI Studio Playgrounds for Model Testing

Playgrounds are useful interfaces that you can use to test your deployed models without developing a client appli-
cation. Azure offers multiple playgrounds with different parameter-tuning options.
FIGURE 24.20: Azure OpenAI Studio playgrounds for model testing
Optimizing Model Behavior Using Playground Parameters

These playground parameters are used to change the performance of the models. Consider this list:
➤➤ Temperature: A lower value provides more predictable answers, while a higher value will result in more
creative answers.
➤➤ Max length (tokens): A token is equal to roughly four characters in the English language. This sets the
maximum number of tokens per model response.
➤➤ Stop sequences: These instruct the model to stop the response at a desired point, such as the end of a
sentence or list.
Other parameters include top probabilities (top P), frequency penalty, presence penalty, pre-response text, and
post-response text.
AI-Powered Code Assistance with GitHub Copilot

GitHub Copilot is like a pair programmer who can assist you in completing the code. It is trained based on an
OpenAI Codex model and can provide coding suggestions and even write full functions. Using the tool can help
improve the developer productivity by doing the following:
➤➤ Generating code for functions, classes, loops, and other programming tasks
➤➤ Filling in missing code
➤➤ Suggesting alternative code
➤➤ Translating code from one programming language to another
As shown in Figure 24.21, it can also troubleshoot bugs quickly by catching errors, enhance developer creativ-
ity by suggesting new ideas, reduce learning curves, and improve collaboration by increasing the readability and
maintainability of the code.
GitHub CoPilot
Your Coding Companion
Features Benefits
Generating code Helps developers stay in the flow
Filling in missing code Learn new programming languages
Suggesting alternative code Find and fix bugs
Translating code Improve the quality of the code
Troubleshooting bugs
Enhancing creativity
Reducing learning curves
Results Integration
74% focus on more Visual Studio
satisfying work Neovim
88% feel more productive VS Code
96% are faster with JetBrains IDE
repetitive tasks
FIGURE 24.21: GitHub Copilot’s features and benefits
To use GitHub Copilot, you just need to install the extension in your IDE, such as Visual Studio, Neovim, and
JetBrains. Currently, you need to have an active GitHub Copilot subscription to use Copilot. While GitHub
Copilot may differ in the details from AWS CodeWhisperer, it’s similar in terms of its goals to improve developer
productivity.
Additional Tools and Platforms

With the advent of generative AI, several tools and platforms have also come up to benefit businesses and indi-
viduals. This section discusses the ones shown in Figure 24.22.
OpenAI API Integrations

OpenAI has released many APIs for developers to access their models, such as GPT-3 and DALL-E. Take, for
example, the Kuki.ai chatbot that leverages OpenAI APIs to provide an impressive conversational chatbot. Its
ability to chat in almost 50 languages and handle so many conversational scenarios makes it a good candidate for
customer service gaming, and so on.
Transforming Conversations with GPT-3, GPT-4, and ChatGPT

GPT-3 and GPT-4 are large language models created by OpenAI, and ChatGPT is a chatbot built on those mod-
els. These LLMs can generate text, translate languages, answer questions, and respond to you in a personalized
manner depending upon your prompts.
NOTE The G in GPT stands for generative, P for pretrained, and T for transformer models.
Runway ML
Develop Generative
Hyper-Realistic AI code without
Faces with NVIDIA coding OpenAI
StyleGAN
GPT-3 and DALL-E integrated into
Creating highly realistic faces Kuki.ai chatbot
that do not exist in reality
Image Style OpenAI Integrations

Transfer with GPT-3 and DALL-E integrated
DeepArt into Kuki.ai chatbot
Combine the style of one
Gen AI
image with another, to create
new artwork Innovative
Tools
ChatGPT
GPT-3, GPT-4, and ChatGPT in
generating text, translation, answering
questions, and providing personalized
Psychedelic Art with
responses
DeepDream
Generate psychedelic Text-to-Image Using
images by adding variations DALL-E
and patterns to ordinary Create hyper-realistic images from text
images for creation of fictional characters and
product designs
FIGURE 24.22: Additional gen AI tools and platforms
Text-to-Image Generation Using DALL-E 2

As discussed earlier, DALL-E2 is a generative AI model that can create images from text. It can create hyper real-
istic images that can be used to create fictional characters or even new product designs.
Unveiling Psychedelic Art with DeepDream

DeepDream is a generative AI tool created by Google that can create psychedelic images from ordinary images. It
does this by learning the existing patterns in the image and creating strikingly new patterns and stunning images
by adding variations.
Creativity Through Image Style Transfer with DeepArt

DeepArt is a generative AI tool created by Google that combines one image’s style with another’s content to cre-
ate entirely new art.
Hyperrealistic Faces with NVIDIA StyleGAN

NVIDIA StyleGAN is a generative AI model that can create stunningly realistic faces that do not exist in
the world.
Unleash Creativity with Runway ML

Runway ML helps users to develop generative AI code without coding. Whether it is creating a new artwork or
designing a new font, Runway ML can provide you with generative AI models to build upon.
WORKBOOK TEMPLATE: GENERATIVE AI USE CASE PLANNER

Download the “Generative AI Use Case Planner” template from the download section of the book (www
.wiley.com/go/EnterpriseAIintheCloud). Use this template to plan for Generative AI use cases and
ensure you have factored in all aspects of planning for the generative AI use cases.
SUMMARY
In this chapter, you journeyed into the complex world of generative AI. You learned about the critical components
of deep learning, semi-supervised learning, generative versus discriminative models, diffusion models, foundation
models, and large language and large image models.
Ultimately, generative AI opens a new chapter of innovation and potential. It can revolutionize and disrupt busi-
nesses, augment human potential, and provide a new perspective for solving business problems. However, it is not
without risks, potential ethical concerns, and calls for an urgent need to protect data and user privacy and ensure
fairness and transparency in decision-making.
It is undoubtedly a new era for businesses and personal lives, and I look forward to seeing how this technology
evolves and transforms our lives.
REVIEW QUESTIONS
1. What is the estimated economic benefit generated by generative AI use cases annually across industries
according to McKinsey?
A. $1.5 to $2.3 trillion
B. $2.6 to $4.4 trillion
C. $5.5 to $7.2 trillion
D. $6.3 to $8.5 trillion
2. Discriminative models in machine learning are focused on
A. Predicting the label for new data points
B. Generating new data instances
C. Understanding the relationships between the features and the labels
D. Learning the probability distribution of the existing data
3. What is the main limitation of large language models (LLMs)?
A. They can only understand text.
B. They require a large amount of compute power.
C. They can only generate text.
D. They can work only with small amounts of data.
4. Variational autoencoders (VAEs) are beneficial because
A. They can understand the hidden layers that drive the data.
B. They are used in text generation use cases.
C. They can operate like black boxes.
D. Both A and B.
5. In terms of workforce transformation, AI can act as

A. A competitor
B. A partner
C. A replacement
6. Which of the following is NOT an advantage of using retrieval augmented generation (RAG)?
A. Improves the relevancy and accuracy of the information response
B. Saves time and effort
C. Inconsistent response
7. Which of the following is NOT a benefit of using Amazon EC2 Trn1n instances?
A. Reduced training time
B. Reduced training cost
C. Increased model accuracy
D. Increased model scalability
8. What is the unique feature of the DALL-E 2 generative AI model?
A. It can create images from text.
B. It can create new fonts.
C. It can generate text, translate languages, and answer questions.
D. It can create psychedelic images from ordinary images.
9. Which of the following is NOT a step in ensuring fairness and bias mitigation through continuous testing
and guardrails?
A. Using a variety of metrics to measure fairness
B. Identifying and addressing any potential biases in the model
C. Monitoring the model’s performance over time
D. Releasing the model to external customers
10. What is one potential outcome of the disruption caused by generative AI in the industry?
A. A decrease in the use of automation
B. An increase in the use of automation
C. A need to adjust or complement the current AI strategy
D. A need to abandon the current AI strategy
11. What privacy-related concern is associated with the use of generative AI?
A. It cannot access private data.
B. It can be used to monitor activities and predict future behavior.
C. It protects all data with advanced encryption.
D. It is incapable of handling private data.
ANSWER KEY
1. B 5. B 9. D
2. A 6. C 10. C
3. B 7. C 11. B
4. D 8. A
25
Planning for the Future of AI
The future belongs to those who believe in the beauty of their dreams.
—Eleanor Roosevelt
You have learned that AI is not a static field; instead, it continuously evolves rapidly, reshaping industries
and personal lives. While generative AI resembles the latest pinnacle of current AI capabilities, it is equally
vital to gaze into the future and discern the emerging trends.
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Your AI Govern Your Evolve and Mature
journey Team
09 • Implement Generative AI Use

Cases for the Enterprise
ENTERPRISE AI • Plan for the future of AI
Setup 04
Evolve and
Mature AI
and Manage Objective
Operations
Dive deep into the potential of Generative AI,

08 Scale and Process Data exploring its evolution, business applications,
Transform and Modeling and implementation on leading cloud
AI
Deploy and 05 platforms, while strategizing for best
Monitor practices and risk mitigation in enterprise
Models contexts
07
06
FIGURE 25.1: EVOLVE AND MATURE: Plan for the future of AI
In this chapter, you learn about the trends in the smart world powered by AR/VR, the captivating concept
of the metaverse, and the intricacies of quantum machine learning. You also learn about the productiv-
ity revolution in industries, the impact of powerful technologies such as digital humans, and the growing
popularity of AI at the edge.
466 ❘ CHAPTER 25 Planning for the Future of AI
This chapter also takes a quick look at the critical enablers poised to reshape AI’s future, such as foundation mod-
els, knowledge graphs, hyper-automation, and the democratization of AI/ML.
This chapter provides you the opportunity to build a futuristic roadmap for continuous innovation and sustain-
able growth in the ever-evolving field of AI.
EMERGING AI TRENDS
We are living in an exciting world, and we haven’t seen anything yet. The future of the world is smart. It is being
shaped by several emerging technologies, such as AR and VR, the metaverse, and digital humans—all of which
are covered in this chapter. The smart world is focused on improving the quality of its citizens. In addition, AI is
transforming businesses to become more productive and efficient.
Behind this revolution are several emerging trends, such as AI in the edge, intelligent apps, compressed models,
self-supervised learning, and generative AI. There are also several critical enablers, including ChatGPT, trans-
former models, foundation models, knowledge graphs, and the democratization of AI/ML, which are all accelerat-
ing this trend. This section looks at these trends.
Smart World
A smart world is about creating virtual replicas of the physical environment and the world we live in. It includes
technologies such as virtual reality, augmented reality, the metaverse, and digital humans. The smart world is
focused on improving citizens’ quality of life. It can be done in many ways, such as the following:
➤➤ Improving transportation: Self-driving cars reduce traffic congestion and improve safety.
➤➤ Managing energy: AI can be used to optimize energy consumption and reduce the cost.
➤➤ Improving healthcare: AI can be used to cure diseases, find new drugs, and improve patient treatments.
➤➤ Increasing security: AI can be used to improve security, detect crime, and prevent terrorism.
➤➤ Protecting the environment: AI can be used to protect the environment through initiatives such as
climate control, pollution control, tracking endangered species, and managing resources.
One of the key initiatives in the smart world category is Singapore’s smart nation initiative, which includes devel-
oping a data exchange program where government agencies can share data and understand the customer’s needs
and problems better.
AR and VR Technology
AR and VR technology can be used to improve business operations and customer service and create new prod-
ucts. Here are some examples:
➤➤ Retail: Augmented reality can be used in the retail industry for customers to try on clothes, makeup, and
other products. This can help customers make better product choices and increase customer satisfaction.
➤➤ Education: Virtual reality can be used for students to go on virtual field trips and experiment with diffi-
cult-to-access or hazardous products.
➤➤ Healthcare: In healthcare, VR can be used to reduce pain for patients, help them recuperate from
surgery, and help them to walk after a stroke.
➤➤ Manufacturing: Employees can be trained under challenging terrains and environments and thus
improve safety and reduce costs using virtual reality.
➤➤ Marketing: Both virtual reality and augmented reality can be used in marketing to create immersive
experiences; for example, augmented reality can be used to insert products in the user’s environment,
while virtual reality can be used to create virtual experiences that customers can explore.
Emerging AI Trends ❘ 467
Metaverse
The metaverse can be thought of as a platform with which businesses can interact with their customers, employ-
ees, and partners.
Here are some examples of how it can be used:
➤➤ Virtual meetings: Meetings can be conducted virtually where participants can interact with each other
and the environment.
➤➤ Product demonstrations: Businesses can demonstrate their products to their customers in a virtual world
using AR and VR technology.
➤➤ Virtual training: Companies can provide immersive simulation environments where employees can train
under challenging circumstances.
➤➤ Remote work: Employees can interact with each other through a platform that allows them to interact
with other employees in a virtual environment.
AI models, natural language processing, and other AI/ML services can be used to enable these services.
Digital Humans and Digital Twins

Digital twins are virtual representations of physical objects, and digital humans are computer-generated represen-
tations of humans. They can be used in education, entertainment, and customer service. They are generated using
AI machine learning and computer graphics technologies. From an enterprise perspective, digital humans can be
used to give presentations to answer customer queries.
The benefit of having digital humans is that they are available 24/7, they always respond consistently to the
customers, and they can be engaging. The disadvantage of using digital humans is that customers may think
the interactions are fake, and they are often expensive to develop and challenging to scale, and maintenance is
intensive.
HANDS-ON EXERCISE: STRATEGIC TECHNOLOGY IMPLEMENTATION
This exercise involves exploring the emerging AI technologies within a specific business or
industry context, understanding the potential of these technologies, identifying potential use
cases, and coming up with an initial high-level implementation plan.
You can ask your workshop participants to explore each of these areas and come up with
deliverables such as presentations to explore real-world examples such as Singapore’s Smart
Nation initiative, AR in retail, VR in education, and digital humans in customer service,
among others.
STEP OBJECTIVE DELIVERABLES EXAMPLES
Form groups To get into groups Groups formed. Team 1, Team 2,

of three to four Team 3
people
Understand To grasp the smart Write up or Singapore’s Smart

smart world world concept presentation on smart Nation initiative
concepts and its potential world concepts and
benefits benefits.
continues
continued
Study AR and To learn the Document or Virtual reality in

VR technologies possibilities of AR presentation detailing education, AR in
and VR in various the applications of AR retail
industries and VR.
Explore the To understand how Concept note on the Virtual meetings,

concept of the metaverse can metaverse and its product
metaverse transform the way business applications. demonstrations in
businesses interact the metaverse
Learn about To understand Write up or report Digital twins in

digital twins the concept and on the application of manufacturing,
applications of digital twins. healthcare, etc.
digital twins
Understand the To understand the A document or Digital humans for

use of digital concept of digital presentation detailing customer service,
humans humans and their the use of digital entertainment,
applications humans in various etc.
fields.
Analyze use To identify use Identify specific use For example,

cases cases where these cases for your chosen using VR in
technologies could technology in your education for
be impactful selected industry. immersive
learning
experiences or
using digital
humans in
customer service
to provide 24/7
assistance
Assess To evaluate the Consider the technical For example, skill

feasibility feasibility of and organizational sets, funding
implementing readiness, the
these technologies availability of resources
in the chosen use and potential
cases challenges or risks
associated with
implementing your
chosen technology for
the identified use cases.
The Productivity Revolution ❘ 469
Plan the To develop a Sketch out an initial

implementation preliminary plan implementation plan,
for implementing including the necessary
the technology steps, resources,
timeline, and expected
outcomes. This can
also include identifying
necessary partnerships,
training requirements,
or changes in
organizational
structures or processes.
Present the To share and Speak about the salient

implementation discuss the plan points and explain and
plan justify the plan.
THE PRODUCTIVITY REVOLUTION

The productivity revolution refers to a set of technologies geared toward increasing an organization’s produc-
tivity and potentially transforming it. It covers technologies such as AI/ML in the edge, intelligent applications,
synthetic data, edge computer vision, self-supervised learning model compression, and generative AI.
All these technologies work together to enable this productivity revolution. The following sections discuss how.
AI in the Edge
AI in the edge involves deploying ML models in edge devices such as smartphone sensors rather than on the
cloud. Deploying AI models in edge devices opens more possibilities, such as intelligent apps, compressed models,
and self-supervised learning.
The benefits of AI on the edge are faster response times, lower bandwidth requirements, reduced cost, and higher
security. An example of a challenge related to AI in the edge is that data security can be compromised when unau-
thorized individuals get access to those devices.
Intelligent Apps
Intelligent apps are software applications that use AI algorithms to provide intelligent and personalized services
to users. They provide benefits such as personalized experiences, process automation, predictive analytics, and
real-time decision-making. Intelligent apps include recommendation engines, chatbots, virtual agents, predictive
maintenance applications, and fraud detection applications. These apps have the potential to revolutionize a
companies’ operations and customer service. See Figure 25.2.
TIP Be a part of the productivity revolution by leveraging technologies like AI/ML on the
edge, intelligent apps, compressed models, self-supervised learning, and generative AI to
increase productivity and transform your organization.
FIGURE 25.2: Intelligent apps provide intelligent services to users.

Source: ImagineDesign / Adobe Stock
Compressed Models
Compressed models are about compressing the models in size and complexity without losing accuracy and per-
formance. They open new possibilities concerning the deployment of the models on the edge. Model compression
results in reduced storage, reduced bandwidth requirements, improved performance, lower cost, and improved
security. However, it is not without drawbacks because, with the reduced size, it can lose accuracy, is also chal-
lenging to develop, and is available for only some models. As technology develops, some of these drawbacks may
be eliminated, making this more effective.
Some examples of model compression techniques are defined here:
➤➤ Pruning the weights: Involves removing certain connections (weights) between the nodes of a neural net-
work to reduce the size of the network.
➤➤ Quantization: Reduces the number of bits used to identify a weight. For example, using 8-bit integers
instead of 32-bit floating-point numbers can reduce the model size and increase the speed of arithmetic
operations.
➤➤ Activations: Refer to the outputs of the neural network. Reducing the size of the activations can also
reduce the compute power required to train or use the model.
➤➤ Knowledge distillation: Involves using a smaller, simpler (student) model to replicate the behavior of a
larger, complex (parent) model. By transferring the knowledge from the larger to the smaller model, you
can deploy it on smaller devices with limited resources.
➤➤ Low-rank factorization: Involves approximating the weight matrices with the product of two lower
ranked matrices to reduce the size of the model. This approach can result in loss of accuracy.
Self-Supervised Learning
Supervised learning is a type of machine learning where the model can be developed by training a large unlabeled
dataset, and it develops its own features without supervision. Self-supervised learning provides benefits such as
Critical Enablers ❘ 471
data efficiency, reduced bias, and improved performance. The advantage of self-supervised learning is that it can
look at the existing data and try to come up with patterns and anomalies, which can then be used to make further
predictions, classify new data, and make recommendations. In healthcare, it can look at a lot of imaging data
and come up with patterns and anomalies that can then be used to treat new patients. Similarly, it can look at
customer tickets and come up with patterns that relate customers with issue types and then use that information
to automate responses to customers as well as to provide insights to the customer service reps to address those
types of issues.
CRITICAL ENABLERS
This section looks at some of the critical enablers that are important for the development of AI and ML solutions
in an enterprise that provides a company competitive edge (see Figure 25.3). They are critical enablers because
they help to democratize the use of AI/ML, speed up the deployment of complex models, automate complex tasks,
and increase efficiency, thus driving rapid productivity revolution in companies.
Foundation Knowledge
Models Graphs
Building blocks for AI Connect disparate
apps data
CRITICAL
ENABLERS
Hyper-automation Democratization
Automate processes with Make AI accessible to all
AI
Critical enablers help:
• Speed deployment
• Increase productivity
• Drive innovation
• Provide a competitive
edge
FIGURE 25.3: Critical enablers for enterprise AI
Foundation Models
You can think of foundational models as basic building blocks that form more complex AI solutions. These are
large, pretrained deep learning models pretrained to create a particular type of content and can be adapted for
other use cases. Foundational models, such as GPT 4 and Bard, can be pretrained models on large amounts of
data and can be used for developing more complex models. They’re primarily used for natural language process-
ing, machine translation, and image recognition use cases. Once you have a foundation model, you can develop
an application on top of it to leverage its content creation capabilities. For example, if you consider GPT model,
you have other applications built on top of it such as Jasper and Copy.ai. This is one of the main reasons that we
are going to see an explosion of a number of generative AI applications in the future.
Knowledge Graphs
Knowledge graphs are databases where entities and their relationships are stored so that models can quickly
leverage them and be trained to make further predictions. This ability to represent data from various sources in
graphs enables developers to use them in use cases such as product recommendations, and to resolve customer
issues. Figure 25.4 shows a knowledge graph that displays a diverse set of entities and relationships to model real-
world contexts and domains.
ACTED IN The Godfather Part II PRODUCED BY Paramount Pictures

Robert De Niro
Tom Hanks ACTED IN Forrest Gump BASED ON

Forrest Gump by Winston Groom
Harrison Ford
ACTED IN Star Wars
PRODUCED BY
Meryl Streep Lucas Film Ltd.
ACTED IN
The Devil Wears Prada
PRODUCED BY
Steven Spielberg Fox 2000 Pictures
DIRECTED Jurassic Park PRODUCED BY
Keanu Reeves ACTED IN The Matrix DIRECTED BY The Wachowskis
FIGURE 25.4: Knowledge graph reflecting complex real-world data as meaningful relationships
TIP When developing AI and ML solutions, consider critical enablers such as foundational
models, knowledge graphs, and hyper-automation to create a competitive edge for your
enterprise.
Hyper-Automation
Hyper-automation refers to the ability to automate business processes using machine learning, robot process
automation, and natural language processing.
Democratization of AI/ML
The mobilization of AI/ML means making AI and ML services accessible to a broad group of users through user-
friendly user interfaces, pretrained models and templates, and services that are hosted on the cloud.
Transformer Models
Transformer models are a class of neural networks useful for natural language processing. They are known for
capturing long-range dependencies between different parts of a sentence and have been found to have exceptional
performance in more effective conversations.
Keras and TensorFlow in the Cloud

These are popular machine learning libraries and frameworks used in the cloud. They use the AWS, Microsoft,
and Google platforms.
Critical Enablers ❘ 473
Quantum Machine Learning

Quantum machine learning can revolutionize AI as it combines quantum computing with machine learning and
enables the use of superposition and entanglement in calculations for faster computations and to solve complex
problems previously unsolvable:
➤➤ Superposition refers to a principle where a quantum system can exist in multiple states simultaneously.
Unlike classical bits, which can be only a 0 or 1, a qubit (quantum bit) can exist in both 0 and 1 states
simultaneously with certain probabilities, which allows quantum computers to process a vast number of
calculations simultaneously.
➤➤ Entanglement refers to a phenomenon where a group of particles can interact in such a way that the
state of each particle depends on the state of the others, which allows for more interconnected and syn-
chronized processing power.
HANDS-ON EXERCISE: EXPLORING CRITICAL ENABLERS FOR ENTERPRISE

AI EXERCISE
This is a hands-on self-guided learning exercise where the participants conduct independent
research to understand various emerging technologies and relate theoretical concepts to
practical applications.
STEP OBJECTIVE INSTRUCTIONS DELIVERABLES EXAMPLES
1 To understand Research A summary of GPT-4 by

the role of foundation the features, OpenAI used
foundation models like GPT benefits, and to develop
models and BARD. Look real-world conversational
into applications applications agents like
built on top of of foundation Jasper and
these foundation models Copy.ai
models.
2 To understand Learn about A brief report Google’s

the role of knowledge graphs, on knowledge Knowledge
knowledge their creation, graphs with Graph used
graphs and their uses. their uses to enhance
Analyze how they in different search engine
can contribute to industries results
the understanding
of context and
connections.
continues
continued
3 To understand Research about A document Application

the concept hyper-automation, describing of hyper-
of hyper- how it combines hyper- automation
automation machine learning, automation and in a
robot process its potential manufacturing
automation, and business setting to
natural language applications streamline
processing to production
automate business processes
processes.
4 To understand Explore how AI/ A presentation Google’s

the concept of ML technologies on the AutoML
democratization are becoming democratization products
of AI/ML more accessible of AI/ML and its that allow
through user- implications for developers
friendly interfaces, businesses and with limited
pre-trained individuals machine
models, and learning
cloud-based expertise
services. to train
high-quality
models
5 To learn about Investigate A report Transformer

transformer Transformer detailing the models used
models models and their architecture for machine
effectiveness in and uses of translation
natural language transformer tasks like
processing tasks. models Google’s
Neural
Machine
Translation
(GNMT)
system
6 To familiarize Learn about A tutorial or Building

yourself with these popular guide on using a neural
Keras and machine learning Keras and network
TensorFlow in frameworks TensorFlow on model using
the Cloud and their cloud AWS, Microsoft, TensorFlow on
implementations. and Google Google Cloud
platforms
Emerging Trends in Data Management ❘ 475
7 To understand Study the A written essay Quantum

the concept principles of or report on machine
of quantum quantum machine the principles learning
machine learning learning and how of quantum algorithms
it can revolutionize machine implemented
AI by combining learning and on quantum
quantum its potential computers
computing with impact on AI like IBM Q
machine learning.
EMERGING TRENDS IN DATA MANAGEMENT

There are a couple of exciting new trends that are happening related to AI modeling. Some are federated learning,
AI at the edge, and quantum computing.
Federated Learning
Federated learning is a new approach that involves training the model across decentralized devices or servers that
hold local data. This is helpful when you cannot move data to cloud servers due to security or latency concerns.
In the case study example about the retail business, this means you can train models using local store data while
maintaining overall model consistency across the company.
AutoML
AutoML and the democratization of AI is another trend that is creating many citizen data scientists by creating
intuitive and accessible models.
AutoML is a significant advancement that automates parts of the machine learning process. It not only makes
model development faster and more accessible but also lowers the barrier to entry to implement AI.
NOTE Federated learning, AI at the edge, quantum computing, AutoML, and explainable
AI are some exciting new trends in machine learning.
Data Flywheels
A data flywheel describes a concept where, as more data is fed into the machine learning applications, more users
start using the information, which creates more demand on the data, which in turn creates more insights that lead
to new machine learning applications. This cycle continues and has come to be known as the data flywheel.
DataOps and Data Stewardship

The need to manage the data pipeline all the way from data creation to storage to consumption and deletion has
come to the fore. This practice is known as DataOps. Data stewardship is an important role that ensures the qual-
ity, trust, governance, privacy, security, and ethics of the data assets.
Distributed Everything
Distributed everything involves connecting the data to the people and enabling no-code, low-code, self-service
options to facilitate greater collaboration between people. These might be data scientists, machine learning
engineers, business analysts, data engineers, and other stakeholders who want to facilitate data visualization
and analysis.
WORKBOOK TEMPLATE: FUTURE OF AI ROADMAP

Download the “Future of AI Roadmap” template from the download section of the book (www.wiley.com/
go/EnterpriseAIintheCloud). Use this template to capture key areas to focus on as part of planning your
organization’s journey with future AI innovations. You can use it effectively by constantly updating it frequently
as the field of AI continues to evolve.
SUMMARY
You just traveled through the exciting world of AI, touching upon many innovations such as the metaverse, AR/
VR, digital humans and twins, and similar technologies that are shaping the smart world. You also reviewed some
of the productivity revolution technologies, such as AI on the edge, intelligent apps, self-supervised learning, and
generative models like ChatGPT. Together, these technologies propel the advancements in the smart world in our
personal lives and the productivity revolution in business.
This chapter also delved into the critical enablers that make these advancements possible, such as foundation
models, knowledge graphs, ChatGPT, hyper-automation, transformer models, and leading tools like Keras and
TensorFlow.
REVIEW QUESTIONS
1. What is the metaverse?
A. A platform for businesses to interact with their customers, suppliers, and employees
B. A virtual world for remote work and virtual training
C. A computer-generated version of the metaverse
2. What are compressed models?
A. Models that are compressed in size and complexity without losing accuracy and performance
B. Models that are uncompressed and have high complexity
C. Models that have low accuracy and performance due to compression
D. Models that have not been trained using machine learning
3. What is self-supervised learning?
A. Models can be developed by training a large unlabeled dataset.
B. Models are supervised by a human expert.
C. Models that are trained on labeled data only.
D. Models that are not used in industry.
4. What is generative AI?

A. It can generate new types of data like the data it was trained on.
B. It can perform only classification tasks.
C. It can be deployed only on the cloud.
5. Choose one of the advantages of self-supervised learning.
A. It requires a large dataset of labeled examples to be effective.
B. It allows the model to develop its own features without supervision.
C. It always outperforms supervised and unsupervised learning methods.
D. It requires more human intervention compared to supervised learning.
6. What does a smart world focus on improving?
A. Transportation only
B. Healthcare only
C. Citizens’ quality of life
D. Environment only
7. In the context of the metaverse, what can AI models and services enable?
A. Product manufacturing
B. Virtual meetings and interactions
C. Paper documentation
D. Physical transportation
8. Which technology refers to computer-generated representations of humans that can be used in
customer service?
A. Metaverse
B. Digital humans
C. Virtual reality
D. Digital humans
9. What emerging trend involves creating virtual replicas of the physical environment and includes tech-
nologies like virtual reality and digital twins?
A. Intelligent apps
B. Self-supervised learning
C. Smart world
D. Compressed models
10. Which of the following benefits is NOT associated with AI on the edge?
A. Faster response times
B. Increased bandwidth requirements
C. Reduced cost
D. Higher security
11. What is a potential drawback of model compression?
A. Improved security
B. Reduced bandwidth requirements
C. Loss of accuracy
D. Reduced storage
12. What principle allows a quantum system to exist in multiple states simultaneously?
A. Entanglement
B. Superposition
C. Quantization
D. Activation
13. What tool can businesses use in the metaverse to interact with customers, employees, and partners?
A. VR 360 goggles
B. AI Fairness 360 toolkit
C. Digital twinning
D. Encryption techniques
14. What is a disadvantage of using digital humans in business interactions?
A. They are available 24/7.
B. Customers may think interactions are fake.
C. They respond consistently to customers.
D. They are engaging.
15. Which term refers to the concept where training the model occurs across decentralized devices or
servers holding local data, especially when data cannot be moved to cloud servers due to security or
latency concerns?
A. Explainable AI
B. Quantum computing
C. Federated learning
D. Edge computing
ANSWER KEY
1. A 6. C 11. C
2. A 7. B 12. B
3. A 8. D 13. A
4. A 9. C 14. B
5. B 10. B 15. C
26
Continuing Your AI Journey
The journey of a thousand miles begins with a single step.
—Lao Tzu
Congratulations on making it this far! I’m impressed with what you’ve achieved, having deployed an AI/
ML and potentially Gen AI solutions that have transformed your organization. This is a significant mile-
stone and a testament to the hard work and dedication of everyone involved.
You gained insights from case studies, addressed challenges and initiated pilots, built your team, developed
an infrastructure, processed data, and deployed and monitored models.
And you didn’t stop there. As your expertise matured, you scaled your AI initiative using the AI maturity
framework and AI center of excellence and then stepped up your game with generative AI use cases.
Figure 26.1 shows this impressive journey.
02
01 Strategize
and Prepare
Plan and
Launch Pilot
03
Continue Build and
Your AI Govern Your Continue Your AI Journey
journey Team
09
Equip your company with strategies and
ENTERPRISE AI insights to sustain AI initiatives, foster a
Setup 04 culture of relentless progress, and
Evolve and
Mature AI
and Manage
responsibly advance on the path of AI-driven
Operations transformation
The journey never ends!

AI
Deploy and 05
Monitor
Models
07
06
FIGURE 26.1: Continue your AI journey

480 ❘ CHAPTER 26 Continuing Your AI Journey
But it is only the beginning of many more exciting things you and your company want to achieve. AI is an evolv-
ing field, and you must watch out for the latest opportunities and trends to capitalize upon.
Your current implementation would have also unlocked many more opportunities and opened up a new chapter
in your digital journey. You must explore how to build upon your past success and lessons learned as you move
on to the next phase in your AI adventure.
REFLECTING ON YOUR PROGRESS

It is important to take time to review what has been done so far.
Reviewing the Lessons Learned

As part of your previous AI/ML implementation, you have captured the lessons learned, including the successes,
challenges, and opportunities to improve across different areas of your implementation. You should revisit the AI
strategy, project management, governance, platform, operations, talent, and security areas and identify opportuni-
ties to work on in the next phase. For example, you may find that more training is required for your team or that
you need to improve your data collection and preparation.
Exploring Opportunities for Improvement

Specifically, concerning your current implementation, you may want to explore the following:
➤➤ Adding new features to your solution
➤➤ Improving the accuracy of the solution
➤➤ Making your solution more user friendly
Embracing the Culture of Continuous Improvement

I know it is a cultural thing, but it is imperative. Every AI/ML implementation should be looked upon as an
opportunity for continuous improvement. It is an opportunity to learn, iterate, and optimize.
Engage Your Business Stakeholders

Your business stakeholders are essential partners in this digital journey; this section discusses how to engage
them. You should install a culture where you can encourage ideas from your stakeholders and promote innova-
tion across your company.
Conduct Discovery Sessions

You will have to take your business owners with you because they are crucial in identifying those use cases for
your implementation. They’re very close to the business processes, and they are the ones who know what works
well and what doesn’t. Knowing that is important because that will help you to define your next scope of work.
Identify your stakeholders across different business units and functions in your company and set up a series of
calls with them to explain the virtues of AI, including machine learning, deep learning, and generative AI. Explain
clearly how AI/ML can help solve their business problems. Then solicit their pain points, concerns, and opportu-
nities to improve their business processes and daily tasks.
Find Opportunities for Automation, Decision-Making, Customer Service

Your goal must be to continue to find new ways to improve your company’s business processes. Here are some
ways you can achieve that goal:
➤➤ Find tasks that you can automate with AI/ML
Planning for the Future: Building a Roadmap ❘ 481
➤➤ Make better decisions with AI/ML

➤➤ Improve your customer service with a chatbot
Engage Technical Stakeholders in the Discussion

Ensure a partnership between the business and technical stakeholders in your organization. Your data scientists,
business analysts, machine learning engineers, data engineers, and DevOps engineers should all be included in this
discussion so that you can validate the feasibility of their ideas and then groom them properly to get them ready
for execution when the time comes. Looking at the technical aspects of the implementation method, initial status
is one of the critical factors to ensure the success of your machine learning implementation.
PLANNING FOR THE FUTURE: BUILDING A ROADMAP

You must look at your existing business processes and operations to scope your work properly.
Mapping Your AI/ML Opportunities

Review all your business processes by talking to your corresponding stakeholders and documenting all their pain
points and opportunities to improve. Once you have documented them, look at the use cases that are possible
through the use of AI technologies such as prediction, classification, clustering, forecasting, personalization, and
generative AI. Try to map the AI/ML opportunities to the previously identified pain points and scope them into
the body of work for the next phase.
Prioritizing Your AI/ML Opportunities

Once you have a comprehensive list of AI/ML opportunities and the corresponding business processes you will be
working on, the next step is to work with your leadership to prioritize these opportunities for execution. To prior-
itize these proposals, you will factor in aspects such as feasibility, strategy alignment, business impact, availability
of skill sets, resources, underlying platform, and data to build the models for this project.
Mobilizing Your Team for the Journey

Now that all the hard work has been done, the fun part of the journey is to mobilize your team for the actual
implementation.
Promote Cross-Functional Collaboration

Talk to various stakeholders across business and technology and share ideas. Ensure you collect all the neces-
sary implementation details from each of them and use that to drive the project implementation plan. Your team
should consist of domain experts from the business, such as business analysts, data engineers, machine learning
engineers, software engineers, data scientists, and even legal and risk management experts.
Develop a Project Implementation Plan

Leverage your experience from your previous project implementation and develop a new project plan. Ensure
you capture all the activities, resources, and timelines accurately. You must ensure that you have clearly defined
all the roles and responsibilities across different businesses and technology at every implementation stage. Talk to
everybody and ensure that you have a commitment from them to execute the project.
Assess Your Current Skills and Build Your Team

This is probably one of the most critical parts of your activities because you will go in circles if you don’t have a
strong, capable team. The good news, though, is that you would have identified the gaps in the skill sets during
482 ❘ CHAPTER 26 Continuing Your AI Journey
the previous implementation, which should now help you address them as part of this planning phase. Capture
all the gaps in a document and create an action plan to fill those gaps. Have a mix of action items such as send-
ing people to the training, encouraging them to attend workshops and seminars, and encouraging them to take
up certifications. Where necessary, talk to the recruiters to hire new talent either as permanent employees or as
consultants for the duration of the project.
ENSURING RESPONSIBLE AI/ML IMPLEMENTATION

Ensure that you implement the best practices for your AI/ML implementation to ensure data privacy and security,
ethics, etc.
Enabling Awareness Around AI Risks and Data Handling

It is important to educate your employees about the benefits of AI and the risks around it. Doing so will help you
implement AI responsibly and ethically and manage secure data handling.
Implementing Data Security, Privacy, and Ethical Safeguards

You need to understand data privacy, security, and ethical risks well. You should implement effective security pro-
tocols such as access control, secure storage, and data encryption. Finally, ensure compliance with data privacy
and security.
Defining Ethical Framework and Data Usage Policies

Ideally, you already have an ethical framework by now, and the data usage policy is well-defined. Take a second
look at those policies and frameworks and try to refine them again based on the learnings and new regulations
that may have popped up, given the fast-evolving nature of AI.
PREPARING FOR THE CHALLENGES AHEAD

The goal of this book is to help you navigate some challenges and adopt some best practices to avoid failures
during the implementation. Unfortunately, there is only so much one can include in a book, and obviously, there
is no substitute for real-life implementation experience because every company is unique. However, if you employ
some of the best practices in this book, you will increase the probability of success by a long way. As we discussed
during the first chapter, close to 80 percent of the implementations fail due to the complexity of an AI/ML effort.
However, knowing the pitfalls and putting some measures in place would have gone a long way. Preparing for
risks and eventualities is half the battle won, and that’s the best we can do.
Encouraging Innovation, Collaboration, and High-Performing Teams

What is important is that you embrace a culture of innovation and continuous improvement, develop an atmos-
phere of openness and collaboration among your stakeholders across both the business and technology, and
ensure that you have a high-performing team with all the skill gaps addressed.
Take note that this is a highly evolving space. You must always be on your toes, watching out for new develop-
ments and opportunities in AI. You need to ensure that your AI strategy is always current and take advantage of
essential opportunities to stay ahead and avoid disruption. It will need strong leadership, and some tough deci-
sions may need to be made. Be quick to celebrate successes and recognize individuals for their contributions.
Preparing for the Challenges Ahead ❘ 483
Leveraging the Transformational Nature of AI

Leveraging AI/ML is a once-in-a-lifetime opportunity with a scale and impact many times larger than what we
have seen thus far. AI is highly transformational in nature and can revolutionize your company’s processes, drive
innovation, and even transform your business models. Remember that with a strategic mindset, a committed
team, and a culture of continuous improvement, you can achieve sustained success and leverage the full potential
of this exciting technology.
My Personal Invite to Connect

Best of luck, and stay in touch with me at www.linkedin.com/in/rabijay1. You can also visit www.
rabiml.com for my latest blogs and plans. I hope this is not the end but the beginning of our relationship on
this AI journey. I hope to hear from you regarding your AI success and challenges, and remember that I will be
around to celebrate your victories as well. Thank you!
INDEX
A AI/ML Algorithm Selection Guide workbook template,

311
A/B testing, 351–352 AI/ML algorithms
access control management, 177, 232 about, 268–269, 311
account setup, as a role for cloud engineers, 177 AI/ML framework, 309–311
accountability, 37, 347 answer key, 314
accuracy, 114, 212, 285, 327–328 artificial intelligence, 269–270
activation functions, saturated, 319–320 Cafee, 310–311
activation layer step, in customer 360-degree categories of machine learning, 273
architecture, 188 Chainer, 311
activations, 470 compared with models and AWS services, 113
Active Directory, 95 convolutional neural networks, 303–305
additive models, EDS and, 123 data-driven predictions using machine learning,
Adjusted Rand index (ARI), 367 272–309
affordability, 451 deep learning, 302–303
Agile, 22–23, 450 factors to consider when choosing machine learning
agriculture, smart farming in, 14 algorithms, 270–272
AI acceleration charter, 67–69 features and weights of predictive algorithms, 270
AI architects, 175–176 generative adversarial networks (GANs), 308–309
AI assistants, 450 Keras, 310
AI Center of Excellence (AI COE) machine learning, 269–270
about, 407–408, 413 MXNet, 311
answer key, 415 PyTorch, 309–310
establishing, 409–413 recurrent neural networks, 305–306
review questions, 413–415 reinforced learning, 300–301
scaling AI to become an AI-first company, 408–409 review questions, 312–313
workbook template, 413 Scikit, 311
AI Center of Excellence (AICOE) Setup Checklist TensorFlow, 309–310
workbook template, 413 transformer models, 306–307
AI for Data Analytics (AIDA), 93 types of supervised learning algorithms, 276–291
AI Hub, 137, 138 using supervised learning, 274–276
AI Maturity Assessment Tool workbook template, 405 using unsupervised learning to discover patterns in
AI Maturity Framework unlabeled data, 292–300
about, 391–392, 394, 406 workbook template, 311
answer key, 406 AI/ML Pilot Launch Checklist workbook template, 158
five stages of maturity, 394–398 AI/ML Platform Evaluation Sheet workbook template,
review questions, 406 148
scaling AI to become an AI-first company, 392–394 AI-powered assistants, 442
six dimensions of AI maturity, 398–404 AI-powered customer service, transforming retail
workbook template, 405 with, 13
AI Operating Model and Transformation Plan AI-powered personalization, in gaming industry, 15
workbook template, 428 AI-powered tools/workflows, 10
AI-first strategy, 5, 23–24, 392–394 Air France, 98
Airbus – artificial intelligence
Airbus, 98 anomaly detection

Aladdin, 93 about, 292
alerting, 207, 369 Amazon DevOps Guru and, 127
algorithms. See also AI/ML algorithms Amazon Lookout for Metrics and, 125
best practices for selecting, 315 data poisoning threats and, 333
deployment of, as a role for model owners, 180 as a use case for autoencoders, 298
developing code for, 317 as a use case for support vector machines (SVMs), 288
level of, as a mitigation technique, 258 Application Teams, 232
optimization, 319 applications, 139, 140, 232
scalability of, 45 appointment scheduling, Amazon HealthLake and, 131
selecting, 180, 317–318 AR technology, 466
Align phase, of AI strategy, 53–55 architecture, 176, 355–357
Alteryx, 93 area under the receiver operating characteristic curve
Amazon, 13, 15, 276 (AUC-ROC), 326, 328–329
Amazon Code Guru, 126–127 artifact repository, 332
Amazon CodeGuru Reviewer, 92 artifact versioning, as an artifact, 230, 231
Amazon Comprehend, 114 artifacts, shifting from prediction to generation of new,
Amazon Comprehend Medical, 92, 131 450
Amazon Connect, 92 artificial intelligence (AI). See also AI/ML algorithms
Amazon DevOps Guru, 92, 127 about, 269
Amazon ECR (Elastic Container Repository), 229, 370 accelerating adoption of with strategic change
Amazon ECR Public Gallery, 232 management, 67–75
Amazon Elastic Inference, 371 adoption of by digital natives, 35
Amazon Forecast, 98, 122, 128 AI enabled contact center use cases, 92
Amazon Fraud Detector, 124, 128 applications, reviewing in various industries, 84–87
Amazon Go, 98 Azure Cognitive Search and, 144
Amazon HealthLake, 92, 131 cloud computing as an accelerator for the growth of,
Amazon Kendra, 96, 125–126, 128 7–8
Amazon Lex, 116–117, 120 collaborative, 43–44
Amazon Lookout for Equipment, 130, 133 creating sustainability plan, 74–75
Amazon Lookout for Metrics, 93, 124–125, 128 culture evolution for, 166–167
Amazon Lookout for Vision, 130, 133 democratization of, 472
Amazon Monitron, 130, 132 developing skills, 73
Amazon Personalize, 98, 121, 128 in Discovery stage of maturity, 394
Amazon Polly, 117–118, 120 as a driver for process transformation, 25
Amazon Rekognition, 95, 97, 98, 118, 119 driving competitive advantage through, 59
Amazon S3, 229, 252 in the edge, 469
Amazon SageMaker, 98, 112, 134–135, 197, 230, emerging trends, 466–469
233–234, 358, 455 features and weights of predictive algorithms, 270
Amazon SageMaker Canvas, 136 for financial services, 93
Amazon SageMaker Clarify, 258 fluency as a change management focus area, 166
Amazon SageMaker Data Wrangler, 255 as a focus areas for building competitive advantage, 59
Amazon SageMaker Endpoint, 233–234, 371 frameworks, 112
Amazon SageMaker Lineage, 370 governance in Pilot stage of maturity, 395
Amazon SageMaker Lineage Tracking Service, 229 growing industry adoption of, 11–15
Amazon SageMaker Pipeline, 255, 370 healthcare and, 12–13
Amazon SageMaker Projects, 370 human-centric, 45–48
Amazon SageMaker Studio Lab, 135 implementing for human review, 381
Amazon Textract, 95, 115–116, 120 integrating into business processes, 4–5
Amazon Transcribe, 117, 120 literacy, 450
Amazon Transcribe Medical, 92, 131–132 machine learning, 269–270
Amazon Translate, 96, 114–115 mapping opportunities, 481
Amazon’s Business Metrics Analysis ML, 93 measuring success of, 69
Amazon’s Embedded LLM, 457 for Netflix, 24–25
Amazon’s Generative LLM, 457 optimizing business processes with, 8
analysis, 82, 92–93 prioritizing, 5–6, 481
486
artificial intelligence – AWS Console
redesigning organization for, 168 workbook template, 148

revolutionizing operations management with, 21 artificial intelligence (AI) portfolio, 57–65
robust, 43 artificial intelligence (AI) strategy
scalable, 44–45 about, 50, 55–56, 65
scaling to become an AI-first company, 392–394 across phases, 63–64
services, 112 Align phase, 53–55
solution deployment in Optimizing stage of maturity, answer key, 56, 65
396 defining, 52, 60–62
solving business problems with, 80–81 developing, 57–65
success and failure with, 4 Envision phase, 51–53
to transform current roles, 174 initiating and planning to kickstart enterprise AI, 59–64
transformational nature of, 483 leveraging organizational capabilities for competitive
trustworthy, 44 advantage, 57–59
as a use for generative adversarial networks (GANs), managing, 60
309 review questions, 56, 65
artificial intelligence (AI) challenges step-by-step methodology for, 50–55
about, 31, 39 workbook template, 55, 65
AI transformation compared with digital artificial intelligence (AI) transformation, 173–174,
transformation, 35–37 424–428
answer key, 40 asset sharing, AI Hub and, 137
choosing between smaller PoC point solutions and Associated Press, 93
large-scale AI initiatives, 37–39 Atomwise, 92
digital natives and adoption of AI, 35 audio-related generative AI use cases, 102
faced by companies implementing enterprise-wide AI, auditing, 11, 227–228, 334
31–34 augmented AI, implementing for human review, 381
review questions, 39–40 autoencoders, 297–298
workbook template, 39 automated content generation, business transformation
artificial intelligence (AI) journey with, 440
about, 479–480 automated ML training, 337
building roadmaps, 481–482 automated validation checks, 335
ensuring responsible AI/ML implementation, 482 automatic deletion, of resources, 235
preparing for challenges, 482–483 automatic detection, Amazon Lookout for Metrics and,
progress reflection, 480–481 125
artificial intelligence (AI) operating model automation, 218–225, 345, 480–481
about, 416–417, 428 AutoML, 138–139, 141, 475
answer key, 429 autonomous vehicles, transforming transportation
customer-centric AI strategy to drive innovation, industry with, 14
418–424 Auto-Regressive Integrated Moving Average (ARIMA),
developing AI transformation plan, 424–428 122, 124, 260
implementing, 418–428 autoscaling, implementing for model endpoints, 371
purpose of, 417 availability, Amazon DevOps Guru and, 127
review questions, 429 AWS
when to implement, 418 about, 112
workbook template, 428 AI frameworks, 112
artificial intelligence (AI) platforms and services AI services, 112
about, 108, 148–149 Amazon SageMaker, 112
answer key, 151 as a cloud provider, 226, 228
AWS, 112–114 controlling access to resources, 235
benefits and factors to consider, 107–111 generative AI cloud platform tools, 455–457
core AI services, 113–120 machine learning algorithms compared with models
Google AI/ML services stack, 136–142 and, 113
machine learning services, 134–135 AWS Bedrock, 456–457
Microsoft AI/ML services stack, 142–146 AWS CloudFormation, 370
other enterprise cloud AI platforms, 147–148 AWS CodeWhisperer, 455–456
review questions, 149–151 AWS Comprehend, 96, 120
specialized AI services, 121–133 AWS Console, 124, 128, 129
487
AWS Elemental Media Insight – case studies
AWS Elemental Media Insight, 97 Bitbucket, 229

AWS Forecast, 93 BlackRock, 93
AWS IOT twin maker, 95 blocks, in Pilot stage of maturity, 395
AWS Machine Learning Interpretability Toolkit, 258 blue/green deployment, 350–351
AWS Panorama, 119, 120 bottlenecks, system, 319
Azure, 95, 226, 228, 234, 457–461 budgets, 69, 397
Azure Applied AI services, 142 bug detection, Amazon Code Guru and, 127
Azure Blob Storage, 253 build vs. buy decision, 200–204
Azure Bot Service, 144 business analyst, 177–178
Azure Cognitive Search, 144 business case, 38, 53–54
Azure Cognitive Services Business Case and AI Strategy workbook template, 65
about, 95, 98, 99, 142–143 business goals, 81, 155, 174–182
Azure Bot Service, 144 business interpretation, as a role for data analysts, 178
Azure Cognitive Search, 144 business metrics analysis, 92–93
Azure Form Recognizer, 144 business outcomes, 60–62, 396
Azure Stream Analytics, 144 business problems, 155, 180
Azure Video Indexer, 144–145 business processing services
Cognitive Services for Language, 143 Amazon Forecast, 122
Cognitive Services for Vision, 143 Amazon Fraud Detector, 124
Azure Container Registry, 230 Amazon Lookout for Metrics, 124–125
Azure Data Catalog, 230 Amazon Personalize, 121
Azure Data Factory, 230 Auto-Regressive Integrated Moving Average (ARIMA),
Azure DevOps, 92, 230, 358 122
Azure Fairlearn, 258 DeepAR, 123
Azure Form Recognizer, 144 Exponential Smoothing State Space Model (ETS),
Azure HDInsight, 97 123–124
Azure Machine Learning, 98, 145–146, 358 Extended Kalman Filter Dynamic Linear Model (EDS)
Azure Machine Learning Datasets, 230 Prophet, 123
Azure Machine Learning Services, 230 business requirements, real-time vs. batch inference and,
Azure ML, 93 352
Azure OpenAI Service, 459 business strategy and ethics, in Transforming stage of
Azure OpenAI Studio, 459–460 maturity, 398
Azure Personalizer, 98 business transformation and innovation, power of
Azure Pipelines, 92 generative AI/ChatGPT for, 440–445
Azure Stream Analytics, 144 business value, measuring, 155
Azure Video Indexer, 144–145 business-related challenges, 32
buy decision, build vs., 200–204
B
C
balance, 257–259
base images, 232 Caffe, 112, 310–311
batch data processing, 251 calculated features, as a feature creation technique, 263
batch inference, 352–353 call centers, Amazon Transcribe for, 118
batch monitoring, 367 canary deployment, 351
batch size, 323 capabilities, 108, 135, 143
Bayesian search, 325 Capital One, 92
best practices, 315, 330–338, 385, 448–453 Capital One and How It Became a Leading Technology
beta testing, 449 Organization in a Highly Regulated Environment
bias, 37, 356, 379, 449. See also model governance case study, 21–24
bias checking, 180, 181 Cartesian products of features, as a feature
bias control, as a model governance challenge, 347 transformation technique, 263
bias drift, monitoring, 365 case for change, creating, 72
binary classification, 275 case studies
binary features, 260 about, 27
binning, as a feature creation technique, 262 answer key, 28
488
categorical features – collaborative filtering
building business case and planning during Align phase, data management architecture, 207–209
55 data science experimentation platform, 209–211
customer-centric AI strategy for driving innovation, 419 developing event-driven architecture using IoT data,
of enterprise transformation with AI in the cloud, 19–28 190–193
evaluating AI opportunities across industries, 11 as a driver for process transformation, 25
How a Company Transformed Itself with AI Using a factors to consider when building ML platforms,
People-Centric Approach, 164–165 200––205
How a Supply Chain Management Company Managed hybrid and edge computing, 211–213
their AI Strategy to Achieve their Business key components of enterprise AI/ML healthcare
Outcomes, 62 platforms, 206–214
How Netflix leveraged their capabilities for competitive key components of ML and DL platforms, 206
advantage, 58 multicloud architecture, 213–214
IBM’s New Collar Job Initiative, 71 personalized recommendation architecture, 193–195
identifying initiatives during Envision phase, 53 real-time customer engagement, 195–198
maturity, 394–396 reference architecture patterns for typical use cases,
Netflix and Portfolio Management, 63 188–200
review questions, 27–28 review questions, 215–216
workbook template, 27 as a role for cloud engineers, 176
categorical features, 260 workbook template, 214
categorical outcomes, predicting using logistic regression, cloud providers, 204–205, 368
278–281 cloud-based data processing, benefits and challenges of,
Central Team, 232 247–250
centralized model management, 227 cloud-first thinking, impact of on DevOps, Agile
centralized unit, 410 development, and machine learning, 23
Chainer, 311 cloud-native principles, 23
change acceleration, 69, 72–75, 166 clustering, 292, 296
change impact assessment, conducting, 71–72 clustering metrics, 367
change logs, as a model governance challenge, 347 CNNs, 98
change management, 36, 153, 166–167, 427 code authoring, as a feature of data science
change tracking, data poisoning threats and, 333 experimentation platform, 210
chatbots, 116–117, 307, 442 code build service, as a component of automation
ChatGPT, 99, 433–434, 440–453, 461–462 pipelines, 223
Chevron, 95 code quality, Amazon Code Guru and, 127
classification, 274–276, 280, 285–286, 327–330 code repository, as a component of automation pipelines,
classification metrics, 366 223
classification models, 260 code versioning, as an artifact, 229–231
C-level support and alignment, in Optimizing stage of CodeCommit, 229
maturity, 397 code-related generative AI use cases, 100–101
climate control, as a use for generative adversarial cognitive service for decisions, Cognitive Services for
networks (GANs), 309 Vision and, 143
clinical notes, Amazon Transcribe Medical and, 132 Cognitive Services for Language, 143
clinical support system, Amazon HealthLake and, 131 Cognitive Services for Vision, 143
clinical trials, Amazon HealthLake and, 131 COIN platform, 93
cloud computing, 7–8, 24–25, 174 Colab, 455
cloud deployment, 348–349 collaboration
cloud engineer, 176–177 AI Hub and tools for, 137
cloud fluency, as a change management focus area, 166 Amazon SageMaker Canvas and, 135
Cloud Metadata Catalog, 231 as a benefit of artifact repositories, 332
Cloud Natural Language, 139, 141 as benefit of using cloud AI/ML services, 108
cloud platform infrastructure with data scientists as a role for data analysts,
about, 187–188, 214–215 179
answer key, 216 encouraging, 482
build vs. buy decision, 200–204 managing artifacts and, 383
choosing between cloud providers, 204–205 SageMaker Studio Lab and, 135
customer 360-degree architecture, 188–190 collaborative AI, 43–44
data anomaly and fraud detection, 198–199 collaborative filtering, 121, 122, 296
489
collaborative robots – data
collaborative robots, transforming manufacturing cost-effective solutions, as benefit of using cloud AI/ML
with, 13 services, 108
collaborative systems, 299 costs
communication, 72, 140, 427 controlling as a challenge of cloud-based data
competitive advantage, 57–59 processing, 247
complex classifications, 288 controlling for models using tagging, 385
complexity, 350, 352 management and billing of, 234
compliance, 11, 114, 132, 179, 247, 384 on-premises deployment and controlling, 349
component reuse, in Optimizing stage of maturity, 396 optimization best practices, 335–338
compressed models, 470 reducing, 8, 15, 83, 131, 384
computer vision, 93, 98–99 course recommendations, as a use case for recommender
computing capacity, as a benefit of cloud-based data systems, 300
processing, 247 Coursera, 96
concept drift, detecting and addressing, 364–365 credit risk assessment, as a use case for decision tree
configuration, synchronizing across environments, algorithms, 282
355–357 criteria, reviewing and refining, 89
configuration as code (CaC), 331 critical enablers, 471–475
confusion matrix, as a validation metric for classification cross-account access, as an AWS best practice, 233
problems, 329–330 cross-functional collaboration, promoting, 481
consistency, in architecture, 356 cross-functional teams, 421–422
container image management, 232–233 culture evolution, as a change management focus area,
container security, as an AWS best practice, 233 166–168
container versioning, as an artifact, 229–231 Current-State Assessment component, of AI
containerization, 211, 332 transformation plan, 427
content accuracy, 440–441 custom models, Amazon Fraud Detector and, 124
content moderation, 93, 118 customer 360-degree architecture, 188–190
content-based filtering, 121, 122, 299 customer analytics, 277
contextual meaning, as a use case for transformer models, customer behavior, DeepAR and, 123
306, 307 customer churn prediction, as a use case for logistic
continuous improvement, 422–423, 480–481 regression, 280
continuous integration and continuous delivery (CI/CD), Customer ECR Instance, 232
156, 218, 335, 357–359 customer engagement, real-time, 195–198
continuous monitoring, 258, 351 customer experience, 8–9, 14–15
continuous testing, 449 customer loyalty, boosting, 441–442
control, hybrid deployment and, 350 customer needs, identifying new, 9–10
controlled variables, as a characteristic of A/B testing, customer satisfaction, 83, 126
351 customer segmentation, 281, 300
convolutional neural networks (CNNs), 303–304 customer service, 12, 82, 282, 301, 480–481
Copy.ai, 93 customer support, AWS Translate for, 115
copyright, protecting, 452 customer support knowledge management system,
core AI roles Amazon Kendra and, 125
about, 175 customer-centric AI strategy, for driving innovation,
AI architect, 175–176 418–424
business analyst, 177–178 customers at scale, enabling real-time service for, 6–7
cloud engineer, 176–177 cybersecurity, 94
data engineer, 176 Cylance, 94
data scientist, 176
machine learning engineer, 176
MLOps engineer, 176 D
core AI services
about, 113 DALL-E 2, 462
chatbots, 116–117 Danske Bank, 44
speech, 117–118 Darktrace, 93
text and document services, 114–116 dashboards, 207, 369
vision services, 118–120 data
correlation matrix, as a feature selection technique, 261 anomaly, 198–200
490
data analysis – democratization
artifacts lineage tracking and, 228–231 benefits and challenges of cloud-based, 247–250
augmentation of, 259 as a component of automation pipelines, 223
defining usage policies for, 482 data exploration and preprocessing stage, 253–259
as a focus areas for building competitive advantage, 59 data needs, 244–247
managing different types of, 247–248 as a feature of data science experimentation
as a maturity dimension, 403 platform, 210
needs for, 244–247 in machine learning lifecycle, 156
selection of, as a role for model owners, 180 phases of ML lifecycle, 250–253
storage options for, 252–253 review questions, 265–267
data analysis, 82, 209, 291 workbook template, 265
data analyst, 178–179 Data Processing & Feature Engineering Workflow
data capability, in Envision phase, 154 workbook template, 265
Data Catalog, 208, 231 data quality assurance, 334
data centers, moving into the cloud, 22–23 data querying, as a feature of data science
data cleaning, 206, 255–257 experimentation platform, 209
data collection, 178, 250–251, 353 data reputation, in large-scale AI initiatives, 38
data combinations, in real-time customer data scalability, 45
engagement, 196 data scaling, 257
data compliance, in large-scale AI initiatives, 38 data science experimentation platform, 209–211
data consistency, real-time vs. batch inference and, 352 data science library access, as a feature of data science
data consolidation, building experiences due to, 22 experimentation platform, 211
data denoising, as a use case for autoencoders, 298 data scientists, 136–138, 176, 177
data drift, detecting and addressing, 363–364 data security, 38, 349, 482
data engineer, 176, 244–247 data shifts, 335
data engineering, 243–244, 259–267 data source diversification, data poisoning threats
data ethics, in large-scale AI initiatives, 38 and, 334
data extraction, Amazon Textract for, 115 data sources, 188, 208
data flywheels, 475 data stewardship, 475
data generation, as a use case for autoencoders, 298 data synchronization, 213, 214, 352
data governance, importance of, 36 data transformation, 209
data infrastructure, importance of, 36 data versioning, 209
data ingestion, 196, 206, 208, 250, 251 data visualization, 196, 254
data initiatives, prioritizing, 5–6 data warehouses, 252
data integration, as a benefit of cloud-based data data wrangling, 255
processing, 247 data-driven predictions, using machine learning, 272–309
data lakes, 207–208, 252 Dataflow, 231
data latency, 213 Dataiku, 147
data level, as a mitigation technique, 258 DataOps, 475
data lineage tools, as a model monitoring challenge, 347 data-related challenges, 33
data management, 179, 207–209, 475–476, 482 DataRobot, 147
data parallelism, 318 dataset compression, as a use case for principal
data partitioning, 257 component analysis, 296
data poisoning threats, protecting against, 333–334 dataset versioning, as an artifact, 229–231
data preparation, 145, 146, 247, 253–254 datasets, as a use case for logistic regression, 280
data preprocessing deception, navigating against, 451
about, 253–255, 354 decision trees, 260, 281–283
data augmentation, 259 decision-making, 10, 480–481
data cleaning techniques, 255–257 deep learning, 94, 206, 260, 302–303
data partitioning, 257 DeepAR, 123, 124
data scaling, 257 DeepArt, 462
managing inconsistent formats, 259 DeepDream, 462
unbias and balance, 257–259 deepfakes, 451
data privacy, 38, 347, 449, 482 Dell, 94
data processing demand forecasting, 277, 282
about, 243–244, 265 demand predictions, ARIMA and, 122
answer key, 267 democratization, 135, 472
491
deployment – external knowledge
deployment, 45, 225, 345. See also model deployment enablers, in Pilot stage of maturity, 395
design involvement, as a role for domain experts, 178 encryption, enabling of inter-node cluster
detection, 257–259 communications, 333
developers, Google AI/ML services stack for, 138–141 endpoints, 230, 233–234, 370–371
development tools, AutoML and, 139 end-to-end automated pipeline, 225
device support, Dialogflow and, 139 energy consumption analysis, 277
DevOps, 23, 92 energy efficiency, 451
Dialogflow, 139, 141 energy industry, transforming, 15
diffusion models, 443–445 engagement strategy, creating, 73
digital humans, 467 ENGIE, 94
digital natives, AI adoption by, 35 entanglement, 473
digital twins, 94–95, 467 enterprise AI, 59–64, 300
Digitata, 93 Enterprise AI Cloud Platform Setup Checklist workbook
dimensionality reduction, 296–298 template, 214
dimensions, reducing using principal component analysis, enterprise AI opportunities, 6–15
294–296 enterprise AI transformation
direct downloads stage, as a data collection method, 250 about, 3, 16
disaster recovery, in multicloud architecture, 214 AI-first strategy, 5
discovery sessions, 480 answer key, 18
Discovery stage, as a stage of maturity, 394–395, 399, into business processes, 4–5
401, 403, 404 case studies, 19–28
discriminator, 308 prioritizing Ai and data initiatives, 5–6
discriminator networks, 443 review questions, 16–17
disruption, risk of, 449, 451 success and failure in, 4
distributed everything, 476 workbook template checklist, 15
distributed model management, 227 enterprise AI/ML healthcare platform, key components of,
distributed training, 318–319, 338 206–214
Docker build pipeline, 225 enterprise applications, Amazon Polly for,
Docker file, for creating training code containers, 321 117
Docker Hub, 232 enterprise-wide advisor, 410
document analysis, Amazon Textract for, 115 enterprise-wide governance, expanding scope from data
document services, 114–116 and analytics to, 450
documentation, streamlining, 11 environment consistency, as a characteristic of blue/green
domain expert, 178 deployment, 350
domain-specific features, as a feature transformation Envision phase, 51–53, 153–154
technique, 263 epochs, number of, 323
drift monitoring, 335 equipment maintenance, 130, 282
drug discovery, as a use for generative adversarial error checking/handling, 156, 181, 352
networks (GANs), 309 ethical considerations, 154, 380–381
DVC, 229 ethical framework, 397, 482
dynamic data, as a model monitoring challenge, 347 ethical handling, of private data, 452
ethical risks, addressing, 37
ethical safeguards, implementing, 482
E ethical use and responsibility, 346, 452
ethics. See model governance
EC2 Trn1n and Inf2 instances, 457 event-driven architecture, developing using IoT data,
ecommerce, 115, 126 191–193
edge deployment, 349–350 experimentation, 135, 395
education industry, 116, 466 explainability, 366, 379–380
efficiency, 8, 15, 281, 335, 384 exploratory data analysis, 178, 253–254
e-learning and training platforms, Amazon Polly for, 118 Exponential Smoothing State Space Model (ETS),
emails, 280, 441 123–124
emerging trends, 466–469, 475–476 export results, in real-time customer engagement, 196
employee collaboration, enhancing, 10 Extended Kalman Filter Dynamic Linear Model (ED)
employee knowledge base, Amazon Kendra and, 125 Prophet, 123, 124
employee productivity, as a success metric, 83 external knowledge, 446–447
492
F1 score – Google
F fraud, navigating against, 451

fraud detection
F1 score, as a validation metric for classification about, 277
problems, 328 architecture for, 198–200
facial reconstruction, as a use for generative adversarial in finance and insurance industries, 12
networks (GANs), 309 as a pain point, 82
fairness, 379, 449 as a use case for decision tree algorithms, 282
false negatives (FNs), 329–330 as a use case for logistic regression, 280
false positives (FPs), 329–330 as a use case for manifold learning, 300
fault diagnosis, as a use case for decision tree algorithms, Future of AI Roadmap workbook template, 476
282 future planning
feasibility, 87–88 about, 465–466, 476
feature attribution drift, monitoring, 365 answer key, 478
feature engineering critical enablers, 471–475
about, 156, 254, 259 emerging AI trends, 466–469
feature creation, 262–263 emerging trends in data management, 475–476
feature extraction, 261–262 productivity revolution, 469–471
feature imputation, 264 review questions, 476–478
feature selection, 260–261 workbook template, 476
feature transformation, 263 future-proof infrastructure, 450
feature types, 259–260 future-state AI vision, defining, 69
feature extraction, 261–262, 296–298, 354 Future-State Vision component, of AI transformation
feature importance scores, as a feature selection plan, 427
technique, 261
feature store, as a feature of data science experimentation
platform, 210 G
features
consistency of, across model training and influence, 334 Gamestatix, 95
creating, 262–263 gaming industry, transforming, 15
imputing, 264 GCP Explainable AI, 258
reviewing periodically, 374 GE Healthcare, 94
selecting, 260–261 generative adversarial networks (GANs), 93, 98,
transforming, 263 308–309, 443–445
federated learning, 475 generative AI
feedback loops, importance of, 218 about, 99–103, 433–434
feedback reports, as a key component of enterprise AI/ML best practices when implementing, 448–453
healthcare platforms, 207 challenges of, 450–452
FICO, 94 evolution of, 434–440
finance industry, 93, 115, 116, 301 as a focus areas for building competitive advantage, 59
financial analysis, 277 growth of, 434–440
financial data forecasts, ARIMA and, 122 power of, 440–445
financial data generation, as a use for generative strategy considerations for, 449–450
adversarial networks (GANs), 309 Generative AI App Builder, 454
financial trading, as a use case for convolutional neural generative AI cloud platforms, 453–462
networks (CNNs), 304 Generative AI Use Case Planner workbook template, 463
Firefly’s Helix, 94 generator, 308
flexibility, hybrid deployment and, 350 generator networks, 443
food, smart farming in, 14 Git, 230, 231
Forbes, 93 GitHub, 232
forecasting, as a use case for recurrent neural networks GitHub Copilot, 460–461
(RNNs), 306 GitLab, 229, 358
form extraction, Amazon Textract for, 115 global reach, Translation AI and, 140
foundation models, 471 global trade, 442
Foxconn, 93 goals, AI initiative, 69
framework images, 232 Goldman Sachs, 95
frameworks, 176, 385–386 Google, 95, 226, 228, 234, 454–455
493
Google AI platform – image-related generative AI use cases
Google AI platform, 98, 454 H

Google AI/ML services stack
AI Hub, 137 hallucinations, 444, 451
AutoML, 138–139 healthcare industry
Cloud Natural Language, 139 AI and, 12–13
for data scientists, 136–138 Amazon Comprehend Medical, 131
for developers, 138–141 Amazon HealthLake, 131
Dialogflow, 139 Amazon Kendra and, 126
image and video services, 140 Amazon Rekognition for, 119
media translation, 139–140 Amazon Textract for, 115
speech-to-text, 140 Amazon Transcribe for, 118
translation AI, 140 Amazon Transcribe Medical, 131–132
Vertex AI, 136 AR and VR technology in, 466
Vertex AI Workbench, 136 AWS Translate for, 115
Video AI, 141 chatbots for, 116–117
Vision AI, 141 use cases for, 92
Google Bard, 455 HealthTab, 95
Google BigQuery, 93 hidden layers, 324
Google Cloud AI Platform, 231, 358 hidden units, 324
Google Cloud Auto ML, 98 hierarchical clustering, 298–299
Google Cloud Build, 92 high dimensional data, 288
Google Cloud Container Registry, 231 high-performing teams, encouraging, 482
Google Cloud Healthcare API, 92 HIPAA, 92
Google Cloud Identity Platform, 95 holdout validation, 325–326
Google Cloud IOT Gore, 95 Honeywell, 99
Google Cloud Language, 97 HR Block, 92
Google Cloud Recommendations API, 98 HSBC, 44, 95
Google Cloud Source Repositories, 230, 231 human resources industry, 115, 116
Google Cloud Storage, 231, 252 human-centric AI, 45–48
Google Cloud Translate, 96 human-like interactions, Dialogflow and, 139
Google Cloud Vision, 95, 98, 99 hybrid and edge computing, 211–213
Google Contact Center AI, speech-to-text and, hybrid approaches, real-time vs. batch inference and, 353
140 hybrid deployment, 350
Google DialogFlow, 92 hyper-automation, 472
Google Firebase Authentication, 95 hyperparameters, 316–317, 322–325, 338
Google search engine, 44, 45, 92
Google Translate API, Translation AI and, 140
Google Video AI, 141 I
Google Vision AI, 141
governance IBM, 71
in Envision phase, 154 IBM Watson, 147
as a focus areas for building competitive advantage, IBM Watson Assistant, 92
59 IBM Watson Financial Services, 93
managing artifacts and, 384 IBM Watson Health, 92, 96
as a maturity dimension, 404 IBM’s Qradar, 94
mechanisms for, 67–69 identity verification, 95
of models (See model governance) image and video services, 140
standardizing, 345–347 image building, as an Azure best practice, 233
governance reports, 356 image classification, 280, 288, 304
government industry, Amazon Textract for, 115 image deployment, as an Azure best practice, 233
GPT-3/-4, 461–462 image editing, as a use for generative adversarial networks
gradients, 289, 320 (GANs), 308
grid search, 324–325 image processing, 297, 307
grid-like structure, as a use case for convolutional neural image recognition, Amazon Textract for, 115
networks (CNNs), 303 image storage, as an Azure best practice, 233
guardrails, 449 image-related generative AI use cases, 101–102
494
impact – LLMs
impact, assessing, 88 J
impact criteria, defining, 87
Implementation plan and timeline component of AI Jenkins, as a CI/CD tool, 358
transformation plan, 427 job displacement, navigating, 452
imputation, as a technique for handling missing values, Johnson Controls, 94
256 Johnson & Johnson, 95
incident response, as a role for security engineers, 179 JP Morgan, 93
inconsistent formats, managing, 259
independent component analysis (ICA), as a feature
extraction technique, 261–262 K
industrial solutions, 129, 130
industry adoption, Amazon Lookout for Equipment and, K Cross validation, 326
130 Keras, 310, 472
industry compliance, as a role for domain experts, 178 key performance indicators, 427–428
industry regulations, as a model governance challenge, key phrase extraction, Amazon Comprehend for, 114
347 K-Means clustering, segmentation using, 292–294
industry trends, 84 K-Nearest Neighbors (KNN), 290–291
inference pipelines, implementing, 353–354 KNIME, 147
inference results, in real-time customer engagement, 196 knowledge distillation, 470
information extraction, as a use case for transformer knowledge graphs, 188, 471–472
models, 307 KPMG, 95
infrastructure Kubeflow, 231, 358
complexities of integration, 35–36
on-premises deployment and control of, 349
in Optimizing stage of maturity, 396 L
real-time vs. batch inference and, 352
scalability, 45 labeled data, 250, 290
infrastructure as a code (IaC), 331 labeling, 206, 274
innovation, 60, 108, 166–168, 482 language barriers, 442
insight gathering, in real-time customer engagement, 196 language detection, Amazon Comprehend for, 114
instance size, selecting optimal, 336 language support, Translation AI and, 140
integration, 108, 125, 180 languages, speech-to-text and, 140
Integration of Services, 142 large data, 293, 303
intellectual property, protecting, 452 large-scale AI initiatives, choosing between smaller proof
intelligence layer step, in customer 360-degree of concepts (PoCs) and, 37–39
architecture, 188 latency, 212, 213, 346, 352
intelligent apps, 469–470 layers, hidden, 324
intelligent document processing (IDP), 95 leadership alignment, ensuring, 69–72
intelligent search, 95–96 leadership sponsorship, in Operationalizing stage of
interaction features, 260 maturity, 396
interface development, as a role for software engineers, learning and iteration, n Pilot stage of maturity, 395
180 learning rate, 322–323
internal knowledge, 446–447 leave-one-out cross validation, 326
internal pilots, 448 legal compliance/documentation, 180, 181
inter-node cluster communications, enabling encryption legal industry, 115, 118, 126
of, 333 lessons learned, 396, 480
interpretability, 281, 286, 379–380 life sciences, use cases for, 92
Intuit, 92, 98 lifecycle, machine learning (ML), 155–158
inventory control, transforming retail with, 13 limited data, 288
IoT data, developing event-driven architecture using, linear data, 288
191–193 linear discriminant analysis, as a feature extraction
issue detection, Amazon Lookout for Equipment and, 130 technique, 262
IT auditor, 181 linear regression, predicting value using, 276–278
iterative team-building, 421–422 live testing, as a characteristic of canary deployment, 351
IVR systems, Amazon Polly for, 117 LLMs, 447
495
loan default prediction – MLOps best practices
loan default prediction, as a use case for logistic manifold learning, nonlinear dimensionality reduction
regression, 280 using, 300
logging, auditing and, 227–228 manufacturing, 13, 119, 466
logistic regression, 260, 278–281 marketing, 115, 118, 466
long term dependencies, RNNs and, 306 max length, as a playground parameter, 460
low-rank factorization, 470 mean absolute error (MAE), 366
mean average P(MAP), 367
mean squared error, 366
M mean/median imputation, as a technique for handling
missing values, 256
machine learning (ML). See also AI/ML algorithms measurement, continuous improvement through, 422–423
about, 3 media, Amazon Transcribe for, 118
artificial intelligence and, 269–270 media intelligence, 97
choosing optimal framework, 337 media translation, 139–141
cloud computing and hardware demands, 7 medical coding, Amazon Comprehend Medical and, 131
data processing phases of lifecycle, 250–253 medical diagnosis, 280, 282
data-driven predictions using, 272–309 medical imaging, as a use case for convolutional neural
democratization of, 472 networks (CNNs), 304
different categories of, 273 metadata, 322
enabling automation, 335 metadata tracking, as an artifact, 229–231
as a focus areas for building competitive advantage, 59 metaverse, 467
impact of cloud-first thinking on, 23 metrics, 326–330, 369
integrating into business processes, 4–5 MICE, as a technique for handling missing values, 256
leveraging benefits of for business, 7 Microsoft, 92
lifecycle, 155–158 Microsoft AI/ML services stack, 142–145
mapping opportunities, 481 Microsoft Azure Learning, 98
modernization, 97–98 Microsoft Azure OpenAI Service, 457–459
planning for, 154 Microsoft Azure Translator, 96
prioritizing opportunities, 481 Microsoft Cognitive Services, 97
revolutionizing operations management with, 21 Microsoft Graph, 96
securing environment, 333 Microsoft Healthcare Bot, 92
solving business problems with, 80–81 Microsoft’s Azure Form Recognizer, 95
strategies for monitoring models, 363–366 Microsoft’s Azure IOT hub, 95
streamlining workflows using MLOps, 331–332 Microsoft’s Azure Machine Learning Platform, 98
success and failure with, 4 min-max scaling, as a technique for handling outliers, 256
tagging to identify environments, 235 misinformation, 451
machine learning engineer, 176 missing values, 255–256
machine learning (ML) platforms and services mitigation, 257–259
about, 108, 134–135, 148–149 ML Operations Automation Guide workbook template,
answer key, 151 237
AWS, 112–114 MLflow, as a CI/CD tool, 358
benefits and factors to consider, 107–111 MLOps automation, implementing CI/CD for models,
core AI services, 113–120 357–359
in Envision phase, 154 MLOps best practices
factors to consider when building, 200–205 about, 217, 237
Google AI/ML services stack, 136–142 answer key, 239
key components of, 206 automation pipelines, 222–225
machine learning services, 134–135 automation through MLOps workflow, 218–221
Microsoft AI/ML services stack, 142–146 central role of MLOps in bridging infrastructure, data,
other enterprise cloud AI platforms, 147–148 and models, 217–221
review questions, 149–151 container image management, 232–233
securing, 332–334 data and artifacts linage tracking, 228–231
specialized AI services, 121–133 deployment scenarios, 225
workbook template, 148 importance of feedback loops, 218
machine translation, 96 logging and auditing, 227–228
maintenance, as a role for software engineers, 180 model inventory management, 225–227
496
MLOps engineer – model testing
model operationalization, 221–237 answer key, 376

review questions, 237–238 automating endpoint changes through pipelines,
streamlining ML workflows using, 331–332 370
tag management, 233–237 Azure Machine Learning and, 145, 146
workbook template, 237 bias drift, 365
MLOps engineer, 176 building a system for, 368–369
ML-powered personalization, 98 challenges of, 362
model artifacts, 321–322, 383–385 classification metrics, 366
model deployment cloud provider tools, 368
about, 345, 359 clustering metrics, 367
answer key, 360 conducting periodic reviews, 369
Azure Machine Learning and, 145, 146 dashboards for, 369
challenges in, 346–347 detecting and addressing concept drift, 364–365
choosing appropriate strategies for, 350–352 detecting and addressing data drift, 363–364
choosing between real-time and batch inference, endpoints for, 370–373
352–353 feature attribution drift, 365
choosing options for, 348–350 implementing autoscaling for model endpoints, 371
deployment process checklist, 348 implementing recoverable endpoints, 370–371
ethical considerations for, 380–381 importance of in production, 362
implementing CI/CD for models, 357–359 key strategies for ML models, 363–366
implementing inference pipelines, 353–355 in machine learning lifecycle, 157
as a key component of enterprise AI/ML healthcare metrics for, 369
platforms, 206 model explainability, 366
in machine learning lifecycle, 157 open-source libraries, 368
MLOps automation, 357–359 optimizing performance, 373–375
pipeline for, 225 ranking metrics, 367
pre-deployment checklist, 347 real-time vs. batch, 367
review questions, 359–360 regression metrics, 366–367
standardizing, 345–347 review questions, 375–376
synchronizing architecture and configuration across as a role for model owners, 180
environments, 355–357 setting up alerts, 369
workbook template, 359 third-party tools, 368
Model Deployment Plan workbook template, 359 thresholds for, 369
model development, 139, 156 tools for, 367–368
model governance tracking key performance metrics, 366–367
about, 377–378, 388 workbook template, 375
answer key, 388 Model Monitoring Tracking Sheet workbook template,
controlling model costs using tagging, 385 375
ethical considerations for deploying models, 380–381 model operationalization
explainability and interpretability in models, 379–380 about, 221–222
fairness and bias in models, 379 automation pipelines, 222–225
implementing augmented AI for human review, 381 container image management, 231–232
importance of, 378 data ad artifacts lineage tracking, 228–231
managing model artifacts, 383–385 deployment scenarios, 225
operationalizing, 381–387 logging and auditing, 227–228
review questions, 388 model inventory management, 225–227
setting up frameworks for, 385–386 tag management, 232–237
strategies for fairness, 378–381 model owners, 180
tracking models, 381–383 model parallelism, 318
workbook template, 388 model registry, as a component of automation pipelines,
Model Governance for Bias and Ethics workbook 223
template, 388 model risk manager, 179
model hosting, 221, 223 model testing
model inference, 354 as a key component of enterprise AI/ML healthcare
model monitoring platforms, 206
about, 221, 361, 375 as a role for data analysts, 179
497
model training – noncore AI roles
as a role for model risk managers, 179 reuse of, in Optimizing stage of maturity, 396
as a role for model validators, 181 review questions, 339–341
model training security of, as a role for security engineers, 180
Azure Machine Learning and, 145, 146 selecting algorithms, 317–318
as a component of automation pipelines, 223 size of, 212
as a feature of data science experimentation platform, structure, 316–317
210 tracking, 381–383
as a key component of enterprise AI/ML healthcare tracking lineage of, 383
platforms, 206 training, 318–322, 327
pipeline, 225 tuning, 322–325, 327
in real-time customer engagement, 196 validating, 325–330
Model Training and Evaluation Sheet workbook template, workbook template, 338–339
338–339 monolithic applications, 420
model tuning, as a feature of data science experimentation movie recommendations, as a use case for recommender
platform, 210 systems, 299
model validation multiclass classification, 275
Azure Machine Learning and, 145, 146 multicloud architecture, 213–214
as a role for data analysts, 179 multimodal models, 308
as a role for model risk managers, 179 multivariate imputation, 264
as a role for model validators, 181 music composition, as a use for generative adversarial
model validators, 181 networks (GANs), 308
model versioning, as an artifact, 230 MXNet, 112, 311
model-related challenges, 33
models
about, 315, 339 N
accuracy of, 212
answer key, 341 Naïve Bayes, 286–287
architecture of, 317 natural language processing (NLP), 94, 144, 304
Azure Machine Learning and building, 145, 146 natural language tasks, Cloud Natural Language and,
Azure Machine Learning and management of, 145, 146 139
best practices, 330–338 Netflix, 15, 58, 63, 97
building, 315–318, 326–327 Netflix and the Path Companies Take to Become
choosing, 89–91 World-Class case study, 24–27
complexity of, as a model monitoring challenge, 347 network connectivity, 213
compressed, 470 network effect, 7
consumption of, 222 neural language processing, 93
defined, 318 neural machine translation, Translation AI and, 140
defining, 322 neural networks, data analysis using, 291
developing algorithm code, 317 neural point forecasting (NPTS), 124
edge deployment and, 349 New York Times, 97
evaluating, 327, 330 news personalization, as a use case for recommender
evaluating, as a feature of data science experimentation systems, 299
platform, 211 noise handling, 288
executing, 221 noise reduction, as a use case for principal component
harmful, 452 analysis, 296
hyperparameters, 316–317 noisy data, 303, 306
improving, as a role for model validators, 181 nominal variables, 279
inventory management for, 225–227, 347 noncore AI roles
lack of transparency in, 452 data analyst, 178–179
latency of, 212 domain expert, 178
machine learning algorithms compared with AI services IT auditor, 181
and, 113 model owners, 180
maintenance of, as a role for model owners, 180 model risk manager, 179
as a maturity dimension, 403 model validators, 181
parameters, 316–317 security engineer, 179–180
registration pipeline, 225 software engineer, 180
498
nonlinear data – pilot project
nonlinear data, 288 organizations, 168–170, 232

nonlinear dimensionality reduction, using manifold outliers, managing, 256–257, 288
learning, 300 out-of-the-box options, 445–446
nonlinear transformations, as a feature transformation overfitting, 319
technique, 263 oversight, data poisoning threats and, 334
normalization, 256, 257 owners, tracking, 382
NVIDIA StyleGAN, 462
NVIDIA’s TensorFlow, 112
P
O P4D, 112
packaged software, 445–446
object detection, 304, 306 pain points, 81–82
Objectives component, of AI transformation plan, 426 parameters, 316–317, 321
one-hot encoding, as a feature creation technique, 262 partnerships, managing, 60, 126
online sales, AWS Translate for, 115 pattern analysis, 293, 296
on-premises deployment, 349 PayPal, 93
open API service, Cognitive Services for Vision and, 143 people, 59, 400, 401
OpenAI API integrations, 461 people-centric approach, to org change management,
open-source libraries, 368 164–168
operational efficiency, 130, 346 performance
operational maturity, in Optimizing stage of maturity, 397 across training and production, 356
operationalizing governance, 381–386 Amazon Code Guru and, 127
Operationalizing stage, as a stage of maturity, 396, 400, Amazon DevOps Guru and, 127
402–404 ensuring optimal, 335
operations, 21, 59, 400–40 metrics for, 156
operations focus, 410–411 optimizing for models, 373–375
operations management, revolutionizing with AI/ML, 21 as a role for software engineers, 180
optimization algorithm, 319 performance metrics, 180, 366–367
optimization techniques, for choosing hyperparameters, performance monitoring
324–325 automating, 345
Optimizing stage, as a stage of maturity, 396–397, 400, as a key component of enterprise AI/ML healthcare
402–404 platforms, 206
Oracle Cloud AI, 147 managing artifacts and, 384
ordinal variables, 279 as a role for cloud engineers, 177
org change management performance optimization, 277
about, 163–164, 171 periodic reviews, conducting, 369
aligning organization around AI adoption to achieve personal protective equipment, 99
business outcomes, 168–170 personalization, 442
answer key, 172 personalized marketing, transforming retail with, 13
people-centric approach, 164–168 personalized patient care, as a use case for recommender
review questions, 171–172 systems, 300
workbook template, 170 personalized recommendation architecture, 193–195
Org Change Management Plan workbook template, 170 pilot project
organizational alignment, as a change management focus about, 152–153, 159
area, 166, 168–170 answer key, 159
organizational capabilities, leveraging, 57–59 business goal identification, 155
organizational change, process transformation as a data processing, 156
driver for, 26–27 Envision phase, 153–154
organizational data, protecting data privacy and following machine learning lifecycle, 155–158
security of, 449 launching, 152–159
organizational design, as a change management focus machine learning problem framing, 156
area, 166 model deployment, 157
organizational readiness assessment, conducting, 72 model development, 156
organizational silo changes, in Transforming stage of model monitoring, 157–158
maturity, 398 planning for launch of, 153
499
Pilot stage – responsible AI
planning for machine learning project, 154 Q

review questions, 159
workbook template, 158 quality control, transforming manufacturing with, 13
Pilot stage, as a stage of maturity, 395, 400, 401, 403, 404 quantizations, 470
pipelines, 223, 370, 374 quantum machine learning, 473
platform-related challenges, 33–34 question answering, as a use case for transformer models,
platforms, 59, 400–401 306, 307
playground parameters, 460
policy enforcement, as an Azure best practice, 233
population health, 131 R
post-processing, 258, 354
potential projects, in Pilot stage of maturity, 395 random forests, 260, 285–286
prebuilt models, Amazon Fraud Detector and, 124 random search, 324–325
precision, as a validation metric for classification ranking metrics, 367
problems, 328 ratio-based features, 260
Precision@k, 367 real estate industry, transforming, 14–15
prediction, in supervised learning, 274 real-life testing, as a characteristic of shadow deployment,
predictive algorithms, features and weights of, 270 352
predictive imputation, 264 real-time alerts, Amazon DevOps Guru and, 127
predictive maintenance, 13, 130 real-time customer engagement, 195–198
predictive modeling, 291 real-time data ingestion, 251
preprocessing, as a key component of enterprise AI/ML real-time inference, 352–353
healthcare platforms, 206 real-time monitoring, 367
pretraining models, 210, 445–446 real-time processing, edge deployment and, 350
price optimization, 277 recall, as a validation metric for classification problems,
price predictions, in real estate industry, 14–15 328
principal component analysis (PCA), 261, 262, 294–296 recommender systems, 299–300, 307
privacy risks, addressing, 37 recoverable endpoints, implementing, 370–371
private data, ethical handling of, 452 recurrent neural networks (RNNs), 123, 260, 305–306
proactive care, Amazon HealthLake and, 131 redundancy management, in multicloud architecture, 214
problem framing, in machine learning lifecycle, 156 reference architecture patterns, for typical use cases,
process automation, 8, 14–15 188–200
process flow use case, 80–87 regression
process optimization, clustering and, 294 metrics for, 366–367
process speed, as a pain point, 82 models for, 260
process transformation, 25–27 in supervised learning, 276
processes, integrating with existing, 452 as a technique for handling missing values, 256
product development, 82, 397 using random forests for, 285–286
product recommendations regularization, 323
personalized, 441 regulatory concerns, 37
as a use case for logistic regression, 280 reinforced learning, 300–301
as a use case for manifold learning, 300 reliability, real-time vs. batch inference and, 353
as a use case for recommender systems, 299 renewable solutions, 451
product-centric approach, to building AI solutions, representative validation data, 334
419–420 reproducibility, 346, 383
production, 356, 362 requirements gathering, as a role for domain experts, 178
productivity revolution, 469–471 requirements.txt, for creating training code containers,
products, 9–10, 60, 420–421 321
project implementation plan, developing, 481 resilient systems, building, 23
prompts, 447 resistance to change, overcoming, 37
proof of concepts (PoCs), 5, 37–39 resource allocation, 37, 352
Prophet, 260 resources, 235, 350
prototyping, Vertex AI Workbench and, 136 response quality, 447–448
pruning the weights, 470 responsibility, as a role for model owners, 180
public images, 232 responsible AI
purpose, as a feature of Azure Cognitive Services, 143 about, 41, 48
PyTorch, 112, 309–310 answer key, 49
500
responsible deployment – specialized AI services
pillars of, 41–48 Amazon Rekognition for, 118

review questions, 48–49 as a challenge of cloud-based data processing, 247
workbook template, 48 as a focus areas for building competitive advantage, 59
responsible deployment, as a role for model risk hybrid deployment and, 350
managers, 179 as a maturity dimension, 404
retail industry monitoring, as a role for cloud engineers, 177
Amazon Rekognition for, 119 in multicloud architecture, 214
AR and VR technology in, 466 on-premises deployment and, 349
AWS Panorama for, 119 protecting of organizational data, 449
transforming, 13 as a role for software engineers, 180
retrieval augmented generation (RAG), 447–448 security compliance, as benefit of using cloud AI/ML
reusability, managing artifacts and, 383 services, 108
revenue growth, as a success metric, 83 security controls, 179, 232
risk assessment, 11, 82, 277 security engineer, 179–180
risk management, 12, 73, 179, 300 security-related challenges, 34
risk mitigation, 448, 452–453 segmentation, 292–294, 351
risk reduction, as a success metric, 83 self-driving cars, as a use case for reinforcement learning,
risks, enabling awareness around, 482 301
Roadmap component, of AI transformation plan, 427 self-service catalog, as a feature of data science
robotic surgery, 12–13 experimentation platform, 211
robotics, as a use case for reinforcement learning, 301 self-supervised learning, 470–471
robust AI, 43 semi-structured data, 248
robustness, 286, 334–335 sentiment analysis, 114, 139, 280
roles, customizing to suit business goals/needs, 174–182 sentiment scores or topic features, 260
rollback, 350, 351, 384 sequential data, RNNs and, 306
Rolls Royce, 94 service development, in Transforming stage of maturity,
root mean squared error, 367 397
root-cause analysis (RCA), 82 service integration, cloud deployment and, 349
R-squared, 367 shadow deployment, 351, 352
rules-based filtering, 121, 122 shared data lake, 207–208
Runway ML, 462 short-term trends, ETS and, 123
Siemens, 94
signal processing, as a use case for singular value
S decomposition, 297
silhouette coefficient, 367
safety, as a role for model owners, 180 simple imputation, 264
sales forecasting, 122, 276 singular value decomposition (SVD), feature extraction
Salesforce Einstein AI, 147 and dimensionality reduction using, 296–297
sample size, as a characteristic of canary deployment, 351 skills, 35–36, 481–482
saturated activation functions, 319–320 smart farming, 14
scalability smart world, 466
about, 22–23 smoothing approach, ETS and, 123
cloud deployment and, 348–349 SOAR, 94
hybrid deployment and, 350 social media, as a use case for recommender systems, 299
as a model deployment challenge, 346 social unrest, navigating, 452
scaling AI to become an AI-first company, 392–394, software engineer, 180
408–409 source code control, as a feature of data science
scalable AI, 44–45 experimentation platform, 211
scheduler pipeline, 374–375 spatial relationships, as a use case for convolutional
Scikit, 311 neural networks (CNNs), 303
screen plot, 295 specialized AI services
searches, 144, 324–325 about, 121
seasonality and trends, ETS and, 123 Amazon Code Guru, 126–127
security Amazon DevOps Guru, 127–129
Amazon Code Guru and, 127 Amazon Kendra, 125–126
Amazon DevOps Guru and, 127 business processing services, 121–125
501
specific services – trained parameters
healthcare solutions, 130–133 security engineer, 179–180

industrial solutions, 129–130 software engineer, 180
specific services, as a feature of Azure Cognitive Services, support vector machines (SVMs), 260, 287–289
143 surveillance, balancing with ethical use of data, 452
speech, 117–118 sustainability, advancing in energy industry, 15
speech recognition, 306, 307 Swiss Re, 95
speech-to-text, 140, 142 Symantec, 94
speed, Amazon SageMaker Canvas and, 135 system bottlenecks, 319
splitting, as a feature creation technique, 263 systems, for model monitoring, 368–369
Splunk, 94
sponsor roadmaps, creating, 74
stakeholders, 37, 69–71, 480, 481
standardization, 256, 257
T
statistical relevance, as a characteristic of A/B testing, 351 tag management, 233–237
stock prices, DeepAR and, 123 tagging, controlling model costs using, 385
stop sequences, as a playground parameter, 460 talent management, in Envision phase, 153
storage, as a benefit of artifact repositories, 332 talent shortage, as a model monitoring challenge, 347
strategic change targeting, as a characteristic of A/B testing, 351
about, 66–67, 75 Team Building Matrix workbook template, 182
accelerating AI adoption with management of, 67–75 team-building
answer key, 76 about, 173, 182, 481–482
review questions, 75–76 adopting cloud and AI to transform current roles, 174
workbook template, 75 answer key, 183
Strategic Change Management Plan workbook template, building cross-functional team for AI transformation,
75 173–174
strategic value, aligning operating model with, 423 customizing roles to suit business goals/needs, 174–182
strategies review questions, 182–183
considerations for generative AI and ChatGPT, roles and responsibilities in ML projects, 173–182
449–450 workbook template, 182
deployment, 350–352 technical infrastructure, navigating, 37
design of, as a role for AI architects, 176 technology, 35–36, 132, 452
for fairness, 378–381 temperature, as a playground parameter, 460
for managing and mitigating risks, 452–453 TensorFlow, 112, 309–310, 454, 472
as a maturity dimension, 399–400 TensorFlow Extended (TFX), as a CI/CD tool, 358
for monitoring ML models, 363–366 testing, 285, 422–423
strategy focus, 410–411 text classification, 114, 288
structure, of models, 316–317 text generation, 307, 308
structured data, 248, 290 text recognition, Amazon Textract for, 115
studio domains, 233–234 text services, 114–116
success metrics, identifying, 83 text-related generative AI use cases, 100
superposition, 473 text-to-image, 444–445
supervised learning text-to-speech conversion, speech-to-text and, 140
about, 273, 274 text-to-task, 444–445
classification, 274–276 third-party tools, 368
regression problem, 276 thought leadership, 440–441
types of algorithms, 276–291 3D generative AI use cases, 103
using, 274–276 thresholds, for model monitoring, 369
supplier relationships, 442 time to market, as benefit of using cloud AI/ML services,
supply chain management, AWS Panorama for, 119 107
support roles time-series features, 260
data analyst, 178–179 tools, 367–368. See also specific tools
domain expert, 178 services, 108
IT auditor, 181 traceability, 335
model owners, 180 traffic management, transforming transportation industry
model risk manager, 179 with, 14
model validators, 181 trained parameters, 321
502
training – video-related generative AI use cases
training Use Case Identification Sheet workbook template, 104

AWS Translate for, 115 use cases
bias in, 356 about, 79–80, 104
models, 318–322, 327 Amazon Lookout for Equipment and, 130
performance across, 356 Amazon Lookout for Vision and, 130
selecting local, 336–337 Amazon SageMaker Canvas and, 135
in supervised learning, 274 Amazon Transcribe Medical and, 132
training code container, 320–321 answer key, 105
training job versioning, as an artifact, 230, 231 business analysts, 177
training stage, random forests and, 285 to choose from, 91–103
training strategy, creating, 73 continuous improvement through measurement and
train.py, for creating training code containers, 321 testing, 422
transactional SQL databases, 252 for decision tree algorithms, 281–282
transfer learning, risks from, 332–333 DeepAR and, 123
transformation plan. See AI operating model in Discovery stage of maturity, 394
transformational leadership, as a change management EDS and, 123
focus area, 166 as a feature of Azure Cognitive Services, 143
transformer models, 306–307, 472 healthcare system, 46
transformers, 444 identification process flow, 80–87
Transforming stage, as a stage of maturity, 397–398, 400, for logistic regression, 280
402–404 mapping to business problems, 87
transition, Vertex AI Workbench and, 136 organizing teams around products, 421
translation, 140, 306, 307 prioritizing, 87–91
Translation AI, 140, 142 reference architecture patterns for typical, 188–200
transparency, 179, 452 review questions, 104–105
transportation industry, transforming, 14 for support vector machines (SVMs), 288
travel industry, chatbots for, 116 workbook template, 104
travel recommendations, as a use case for recommender user data uploads step, in real-time customer engagement,
systems, 299 196
trigger automation, as a feature of data science user interactions, role of generative AI in, 448
experimentation platform, 210 user requests, Dialogflow and, 139
true negatives (TNs), 329–330 users, Amazon Transcribe Medical and, 132
true positives (TPs), 329–330
trust-building, through beta testing, 449
trusted data sources, 334 V
trustworthiness, of models, 334–335
trustworthy AI, 44 validation
tuning models, 322–325, 327 choosing techniques for, 325–326
ensuring, 334–335
holdout, 325–326
U K Cross, 326
leave-one-out cross, 326
unbias, 257–259 metrics, 326–327
uncorrelated data, clustering and, 294 metrics for classification problems, 327–330
underfitting, 319 models, 325–330
unified feature engineering, 334 value drivers, 68, 276–278
unified framework, as a model monitoring challenge, 347 vanishing gradients, 320
units, hidden, 324 version control, 32, 335, 382–383
University of San Francisco, 92 version management, managing artifacts and, 384
unlabeled data, 292–300 versioning, as an AWS best practice, 232
unstructured data, 144, 248 Vertex AI, 136, 137
unsupervised learning, 273, 292–300 Vertex AI Workbench, 136, 137
upskill, SageMaker Studio Lab and, 135 Video AI, 141
U.S. Government and Power of Humans and Machines video editing, as a use for generative adversarial networks
Working Together to Solve Problems at Scale case (GANs), 308
study, 19–21 video-related generative AI use cases, 102–103
503
virtual assistants – z-score normalization
virtual assistants, 116 workforce transformation, as a change management focus

virtuous cycle, 7 area, 166
Vision AI, 141, 142 workloads, 108, 213
Vision Alignment Worksheet, 55 workplace efficiency, 442
Vision component, of AI transformation plan, 424–426
vision services, 118–120
visual damage detection, Amazon Lookout for Vision and, X
130
voice activated systems, 117 XGBoost, 260
voice-enabled IoT devices, Amazon Polly for, 118
VR technology, 466
Y
Yamaha Corporation, 98
W YouTube, 98
warehouse management, AWS Panorama for, 119
warm starting, 338
Washington Post, 93
web scraping stage, as a data collection method, 250
Z
Wix, 92 z-score normalization, as a technique for handling
workflows, integrating with existing, 452 outliers, 256
504
WILEY END USER LICENSE AGREE-
MENT
Go to www.wiley.com/go/eula to access Wiley’s ebook EULA.

Enterprise AI

Uploaded by

Copyright:

Available Formats

Enterprise AI

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Enterprise AI

Uploaded by

Copyright:

Available Formats

ENTERPRISE

A Practical Guide to Deploying

Published by John Wiley & Sons, Inc., Hoboken, New Jersey.

Published simultaneously in Canada and the United Kingdom.

ISBNs: 9781394213054 (Hardback), 9781394213078 (epdf), 9781394213061 (epub)

Cover image: © CSA Images/Getty Images

I also dedicate this to my mom, Leela Jayachandar, who is

Finally, I dedicate this to you, dear reader, for picking up

How to Contact the Author

CHAPTER 1: ENTERPRISE TRANSFORMATION WITH AI IN THE CLOUD 3

CHAPTER 2: CASE STUDIES OF ENTERPRISE AI IN THE CLOUD 19

PART II: STRATEGIZING AND ASSESSING FOR AI

CHAPTER 3: ADDRESSING THE CHALLENGES WITH ENTERPRISE AI 31

CHAPTER 4: DESIGNING AI SYSTEMS RESPONSIBLY 41

CHAPTER 5: ENVISIONING AND ALIGNING YOUR AI STRATEGY 50

CHAPTER 6: DEVELOPING AN AI STRATEGY AND PORTFOLIO 57

CHAPTER 7: MANAGING STRATEGIC CHANGE 66

PART III: PLANNING AND LAUNCHING A PILOT PROJECT

CHAPTER 8: IDENTIFYING USE CASES FOR YOUR AI/ML PROJECT 79

CHAPTER 9: EVALUATING AI/ML PLATFORMS AND SERVICES 106

CHAPTER 10: LAUNCHING YOUR PILOT PROJECT 152

PART IV: BUILDING AND GOVERNING YOUR TEAM

CHAPTER 11: EMPOWERING YOUR PEOPLE THROUGH ORG CHANGE

CHAPTER 12: BUILDING YOUR TEAM 173

PART V: SETTING UP INFRASTRUCTURE AND MANAGING OPERATIONS

CHAPTER 13: SETTING UP AN ENTERPRISE AI CLOUD PLATFORM

CHAPTER 14: OPERATING YOUR AI PLATFORM WITH MLOPS

PART VI: PROCESSING DATA AND MODELING

CHAPTER 15: PROCESS DATA AND ENGINEER FEATURES IN

CHAPTER 16: CHOOSING YOUR AI/ML ALGORITHMS 268

CHAPTER 17: TRAINING, TUNING, AND EVALUATING MODELS 315

PART VII: DEPLOYING AND MONITORING MODELS

CHAPTER 18: DEPLOYING YOUR MODELS INTO PRODUCTION 345

CHAPTER 19: MONITORING MODELS 361

CHAPTER 20: GOVERNING MODELS FOR BIAS AND ETHICS 377

PART VIII: SCALING AND TRANSFORMING AI

CHAPTER 21: USING THE AI MATURITY FRAMEWORK TO

CHAPTER 22: SETTING UP YOUR AI COE 407

CHAPTER 23: BUILDING YOUR AI OPERATING MODEL AND

PART IX: EVOLVING AND MATURING AI

CHAPTER 24: IMPLEMENTING GENERATIVE AI USE CASES

CHAPTER 25: PLANNING FOR THE FUTURE OF AI 465

CHAPTER 26: CONTINUING YOUR AI JOURNEY 479

HOW THIS BOOK IS ORGANIZED

WHO SHOULD READ THIS BOOK?

Data Scientists and AI Teams

IT Leaders and Teams

Students and Academia

Consultants and Advisors

Business Strategists and Leaders

WHY YOU SHOULD READ THIS BOOK

Comprehensive Coverage of All Aspects of Enterprise-­wide AI

Case Study Approach

Coverage of All Major Cloud Platforms

Discussion of Nontechnical Aspects of AI

Best Practices for MLOps and AI Governance

UNDERSTANDING ENTERPRISE AI TRANSFORMATION

NOTE ML stands for machine learning and includes generative AI.

Why Some Companies Succeed at Implementing AI and ML While

AI Adoption by Leaders (%) AI Adoption by Others (%)

Comprehensive Coverage of All Aspects of Enterprise-wide AI

FIGURE 1.1: Comparison of enterprise-wide AI adoption by leaders and others

HANDS-ON CONCEPTUAL EXERCISE: UNDERSTANDING ENTERPRISE AI

Adopt AI-First to Become World-Class

Importance of an AI-First Strategy

FIGURE 1.2: An AI-first strategy leads to an enterprise transformation.

HANDS-ON EXERCISE: AI-FIRST STRATEGY FROM CONCEPTS TO APPLICATIONS

Enable One-to-One, Personalized,

The Network Effect: How AI-Powered

HANDS-ON RESEARCH PRESENTATION EXERCISE: EXPLORING THE CLOUD AND

Cross-selling and Companies can use AI algorithms such as recommendation systems to

Enhance Employee Collaboration and Decision-Making with AI-Powered Tools and

EMPLOYEE COLLABORATION AND DECISION-M AKING USE CASES

COMPLIANCE-R ELATED USE CASES

Revolutionizing Retail: AI-Powered Customer Service, Personalized Marketing, and

AI-Powered Personalization in Entertainment: Netflix, Amazon, and the Gaming

WORKBOOK TEMPLATE - ENTERPRISE AI TRANSFORMATION