Telecommunications

Generative AI Study
September 2023

1
Contents

Introduction & Background 03

About The Study 06

What Is Generative AI? 08

Key Insights & Findings 10

Overall Telecommunications Perspective 13

Generative AI Telecommunications-Specific Use Cases 19

Constraints & Complexities 29

Deployment & Implementation 34

Implications & Outlooks 41

Appendix 45

Glossary 46

Definitions 47

Endnotes 50

About Altman Solon 51

About Amazon Web Services (AWS) 52

Acknowledgments 53
01 Introduction
& Background
 01 l Introduction & Background

In the ever-evolving telecommunications industry landscape, continuous innovation is imperative

to meet the demands of a digitally-driven world. As technology paves the way for groundbreaking
advancements, generative artificial intelligence (AI) captures attention and interest across the telecom
industry. From transforming network management to humanizing customer support and crafting
personalized content, this study explores where possibilities meet reality.

The telecommunications sector presents one of the richest case studies of generative AI usage.
Leveraging vast amounts of available data while navigating regulatory and technology challenges
makes it a compelling candidate for exploration and implementation:

1. Vast Amounts of Rich Data: The telecommunications industry generates enormous

volumes of different types of data daily, including call records, network performance data,
customer interactions, and more.

2. Significant Risk with Data: Because communication service providers handle a vast
amount of sensitive data, compounded by regulations that impose strict data protection
requirements, the telecommunications industry has been risk-averse with technology
investments compared to other sectors despite an increasingly data-driven landscape.

3. Legacy Technology Stack: Communication service providers still face significant

obstacles when dealing with outdated technology systems. The burden of addressing
technical debt and navigating complex integrations has led to a slower modernization
pace than in other industries. Gartner predicts that, by 2025, technical debt (the implied
cost incurred when businesses do not fix problems that will affect them in the future) will
consume more than 40% of operators’ current IT budget.

While the telecommunications industry is not the sole beneficiary, generative AI thrives on large
datasets, making the sector well-suited to process, interact, and derive insights from this data. At the
same time, generative AI can also help communication service providers (CSPs) navigate the goals of
risk, cost-efficiency, and innovation by facilitating use cases that can enhance customer experience,
spur innovation, and improve operational efficiency.

4
 01 l Introduction & Background

To date, few early-adopter communication service providers have embarked on plans to experiment,
test, and explore the potential of generative AI across multiple business functions and use cases.
A major U.S. telecommunication services provider noted exploring generative AI for network
optimization1, while a European provider is feeding anonymized transcripts to create summaries
and analyze customer conversations;2 an Asia-Pacific provider uses a large language model for
a chatroom experience where customers can have conversations with a generated AI character;
multiple major Asian providers have announced plans to develop a multilingual large language model
customized for global telecommunications firms.3,4

However, as the industry gravitates toward integrating generative AI, it faces its share of challenges.
Telcos seek high-performance models trained on their data for specific purposes, which can be
resource-intensive and time-consuming. Compliance with regulations is a priority, encompassing
ethical use and user data privacy. Legacy system integration, data security, scalability, infrastructure
costs, talent availability, and expertise are also areas of careful consideration.

Informed by in-depth discussions and research with telecom executives worldwide, this whitepaper
delves into the myriad applications and benefits of generative AI within the telecommunications
sector, how companies are approaching deployment, and the challenges that lie ahead.

5
02 About The
Study
 02 l About The Study

Exhibit 1: Respondent Demographics

Exhibit 1: Respondent Demographics​ AWS engaged Altman Solon, the largest global
% of total respondents, N = 102
% of total respondents, N = 102​
strategy consulting firm exclusively working in
Geography Business SizeMedia, and Technology
the Telecommunications,
industries, to
26%
offer insight into generative
27% AI within
23%
the telecommunications industry. Specifically, this
North America 56% 15%
39% Europe
research
9% seeks to understand the perspectives of
communication service providers trialing and adopting
17%
this fast-evolving,
Less than $1B $1B to <$10B
innovative
$10B to <$20B
technology.
$20B to <$50B Greater than
or equal to This report
$50B

Asia Pacific / Oceania aims to share new learnings and analyses on the

Communication Service Provider Type1

drivers, concerns, and
Title considerations that will impact

generative AI adoption in the telecom industry.

Director / Senior
77% 56%
Director Equivalent

49% 9% Head / Vice

41% To assess
Presidenttelecommunication’s
Equivalent
26% sentiments on
9%
generative AI, Altman Solon
18%
surveyed over 100 senior
C-Suite Executive
business leaders at Tier 1 communication service
Incumbent
Telephone
Fiberoptics
Network
Mobile
Network
Fixed Wireless
Operator
Cable
Operator
providers headquartered in the U.S., Western Europe,
Company Operator Operator
(i.e. LTE/5G) and Asia Pacific. More than 40% were C-level or VP/
SVP/EVP-level executives, with the balance being

Non-Mobile
Directors and Senior Managers.
23% Operator

phics Mobile
Additionally, Altman Solon drew on the experience
77%
Operator
of 25 industry experts through live, anonymized, in-
depth interviews to understand the relevant use cases
Business Size
and top concerns regarding generative AI tools of
27%
26%
MNO = Mobile Operator or CSP’s who (in addition to other
1
global wireless, wireline, and cable companies.
23% services) mobile network (e.g. LTE/5G) Non-MNO = Non-Mobile Operator or CSP’s who are not
mobile network operators but may provide other services such as fixed wireless, fiberoptics, cable, and/or telephony
15%
Europe
9%
Altman Solon combines output from this new survey
with proprietary insights from its recent study, Putting
Less than $1B $1B to <$10B $10B to <$20B $20B to <$50B Greater than
or equal to
$50B Generative AI to Work, from Spring 2023, which
surveyed 292 senior business leaders and interviewed
Title
der Type 1
21 industry experts to understand the adoption of
Director / Senior
Director Equivalent
56% generative AI tools for specific high-impact enterprise

% Head / Vice
use cases, key considerations for development
26%
President Equivalent
and deployment, and implications to organizations’
9%

C-Suite Executive 18% underlying infrastructure.

Wireless Cable
erator Operator
MNO = Mobile Operator or CSP’s who (in addition to other services) mobile network (e.g. LTE/5G) Non-MNO = Non-Mobile Operator or CSP’s who are not mobile
1

network operators but may provide other services such as fixed wireless, fiberoptics, cable, and/or telephony

Non-Mobile 7
Operator
03 What Is
Generative
AI?
 03 l What Is Generative AI?

Altman Solon defines deep generative models (i.e., generative AI) as a subset of artificial intelligence
distinguished by its ability to create new content based on a user’s prompt. While other AI models
have been around longer, these existing, traditional artificial intelligence or machine learning tools (AI/
ML) analyze data to enable decisions, classifications, or predictions rather than generate new content.

Exhibit 2: Artificial Intelligence Taxonomy

Exhibit 2: Artificial Intelligence Taxonomy
For example, machine learning models
Artificial Intelligence (AI)

Rule-Based AI Models
have long been used to predict purchasing
likelihood, classify images into specific
Machine Learning (ML)

Deep Learning (DL)

categories, discern the intent or sensitivity
Deep Reinforcement Learning (DRL) of a customer’s conversation with a call
In Scope
Deep Generative Models operator, or provide pre-set responses
Language
Models
Image
Models
Audio
Models
Video
Models
based on pre-defined questions. These
models all use data to “learn” and improve
their output, but none aims to generate new
content. Generative AI does exactly this:
generate content, imitate human “chattiness” and preferences, or, even, produce quality visuals based
on user prompts.

Altman Solon categorizes generative AI models by the type of content they create, including (but not
limited to) language, image, audio, video, or 3D. These five different model types can be characterized
by their capabilities. For example, the core capabilities of language generation include retrieving
responses through search, summarization, text translation (from one language to another, including
code), and autocomplete functions. For image generation, users can convert text to images or alter
existing images to produce a new image. Language and image capabilities are more mature today
than other generative AI models like video or 3D images. Future generative AI models are expected to
be multimodal, combining various data types.

9
04 Key Insights
& Findings
 04 l Key Insights & Findings

We identified several key insights as part of the study:

1. Communication Service Providers (CSPs) see generative AI as distinct, significant

incremental value on top of existing AI/ML
a. Seventy percent of CSPs see the incremental value served by generative AI as
distinct and significant.
b. Sixty-four percent of CSPs agree that many of the generative AI use cases being
considered are new applications, not yet served by existing non-generative AI
applications and processes.

2. CSPs have kick-started their generative AI journey and while still early in their
journey, some have already implemented use cases, which are now in production
a. The current adoption rate for the tested use cases averages at 19%. However, this
number is poised to increase significantly to 48% in the next two years, indicating
telcos implementing generative AI across more and more use cases.
b. North American CSPs marginally lead adoption (average use case adoption is 21%),
with Europe only marginally behind (19% average use case adoption); APAC slightly
lags at 16%, given the perception of limited capabilities in non-English languages of
existing generative AI models, and lower perceived data capabilities.
c. Survey results indicate CSPs will rapidly increase their generative AI spend as much
as 6x in the next two years; for instance, 45% anticipate their generative AI spend to
be 2-6% of total technology spend in two years, up from less than 1% today.

3. CSPs are placing strong emphasis on using generative AI to enhance customer-

centricity, but they primarily see it as a productivity tool
a. Customer chatbots are the most widely adopted use case, with 63% of respondents
who selected the use case indicating it was already in production.
b. Other use cases often focus on employee assistance, such as contact center
documentation and network operations knowledge management, and are considered
high business value with relatively easier implementation (Clear Wins); these are also
of immediate focus, especially since these drive operational efficiencies.
c. CSPs perceive network planning use cases such as generative route configuration
and synthetic data generation for security testing as high value but are considered to
have more complexity in implementation (Big Bets).
d. More data-capable CSPs are pursuing revenue-generating use cases such as
personalization and new product feature generations.

11
 04 l Key Insights & Findings

4. Data security and governance (i.e., data management principles, policies, processes,
and practices) are top challenges and critical enablers when implementing
generative AI solutions
a. Universally, CSPs see data governance and ownership as the greatest gaps,
especially given the importance of security, privacy, and regulation compliance.
b. While data governance is a greater concern than data infrastructure, CSPs will still
encounter data siloes during implementation of generative AI. Hence, CSPs will need
to integrate these data siloes, which will prove to be more critical than building new
infrastructure.
c. Organizations ranking in the top 30% for data proficiency tend to perceive revenue-
generating use cases as less complex to implement, signifying a higher level of
readiness.
d. More data capable CSPs leading in generative AI adoption organizationally share
certain characteristics: a dedicated AI center of excellence, pervasive use of
advanced data analytics, and modern data infrastructure (e.g., cloud).

5. Most CSPs expect to use off-the-shelf models rather than develop foundation models
in-house, but best practices are still being debated as CSPs learn cost-benefits of
generative AI
a. Fifteen percent of CSPs indicated an inclination to build foundation models in-house,
with the rest indicating purchasing and using off-the-shelf models.
b. Over 65% of CSPs anticipate training off-the-shelf models primarily using their
proprietary internal data to tailor them for their specific needs.
c. The jury is still out on a universal approach to training off-the-shelf models, with most
CSPs evaluating both context learning/prompt engineering (i.e., inference) and fine-
tuning (i.e., training) approaches.
d. The telecommunication sector seems to be leaning towards context learning; this
technique is becoming increasingly popular as it is less resource-intensive and
provides ‘good enough’ performance for many use cases.
e. CSPs note the unknown cost of implementing generative AI as a top business
concern; the availability of 3rd party generative AI solutions and platforms, including
model-as-a-service (MaaS), are critical accelerators to adoption.

12
05
Overall
Telecommunications
Perspective
 05 l Overall Telecommunications Perspective

Generative AI is a new wave of artificial intelligence, and in our study, communication service
providers (CSPs) affirmed interest in and adoption of this innovative technology. A majority of CSPs’
executives noted using traditional AI/ML technology within the organization. Even so,
telecommunications companies see generative AI as a means to create distinct, significant
incremental value, even for use cases that might be utilizing traditional AI/ML. Similarly, 70% of CSPs
see generative AI’s incremental value as distinct and significant. Sixty-four percent of CSPs indicate
these use cases as net-new and not simply a replacement for existing machine learning applications.

Exhibit 3:3:
Exhibit Communication Service
Communication Providers’
Service SentimentsSentiments
Providers’ on Existing AI/ML & Generative
on Existing AI & Generative AI1
AI/ML
1

% of
% oftotal
totalrespondents, N = 102N = 102​
respondents,

Agree Neutral Disagree

Many of the generative AI use cases being

considered are already served by existing 36% 23% 41%
non-generative AI applications and processes

The incremental value served by generative AI

for the use cases which already use
30% 14% 56%
non-generative AI applications and processes
is marginal

Integrating generative AI on top of existing

non-generative AI applications and processes 61% 21% 18%
will be challenging

Please indicate your organization’s sentiment on statements regarding your organization’s pre-existing non-generative AI solution
1

“ Customer support and self-care is a specific context where generative

AI can be seen as an improvement and an additional layer on top of the
existing AI and Machine Learning. We’ve been using bots for several years
now, and generative AI is a step further to help our customer care pre-
assess the requests from our customers.
Head of IT Architecture
Wireless CSP, Western Europe

14
 05 l Overall Telecommunications Perspective

Although still in the early stages, CSPs have kick-started their generative AI journey, piloting and
actively implementing use cases in production. Average use case adoption is 19% among surveyed
CSP respondents, and our survey indicates this will grow to 34% in the next year and 48% in the next
two years.

Exhibit
Exhibit 4: Communication
4: Communication Service Provider(s)
Service Provider(s) Generative
Generative AI Average AI Average
Use Case Adoption1 Use Case Adoption1
% of
oftotal
totalrespondents
respondents who
who have
have implemented
implemented or areorimplementing
are implementing a generative
a generative AI use
AI use case, case,by
N varies N region
varies and
by region and
business business
type type respectively
respectively

19% 21% 19% 16%

Global N.A. Europe APAC

37% 38%
of communication service providers of communication service providers
anticipate their spending on generative anticipate their spend on generative AI
AI use cases to be <4% of total use cases to be 4-8% of total
technology spend in 2 years technology spend in 2 years

1
Please indicate the current stage of adoption of the following generative AI use cases within each business domain.

North American CSPs marginally lead adoption (current average of 21% adoption of tested use cases)
compared to peers, potentially benefiting from from regional proximity to generative AI investments.
The Bay Area has been the main hub for generative AI technology investment, attracting over $18B
from venture capital and investors in the last four years, sixteen times the largest global hub outside
the US (i.e., Tel Aviv $433M).5

European CSPs are more cautious about adopting generative AI given regional data residency
restrictions such as GDPR. Particularly for CSPs outside North America, present and future
governmental regulation on AI usage and data restrictions and residency continues to be a key
consideration. Countries such as China and the EU are tightening AI regulation and monitoring for AI,
while the U.S. and India seek a more passive approach.

15
 05 l Overall Telecommunications Perspective

Exhibit5:5:Key
Exhibit KeyAIAI Global
Global Regulatory
Regulatory Considerations
Considerations

US lawmakers are considering potential reforms to Section 230 of the Communications

United Decency Act, which could require tech companies to identify and label AI-generated content.
States There is a proposal to establish a National AI Commission in the US to manage AI policy, as US
tech companies are concerned with the strict regulations implemented by the EU and seek a
more favorable regulatory environment to compete with China.

The EU’s General Data Protection Regulation (GDPR) (2016) requires consent before data
European transfer can occur and asks that tools disclose content generated by AI. Models must also
Union prevent illegal content generation, and AI deemed a threat to safety and rights will be
assessed before widespread usage.1

The EU AI Act (2023) restricts AI use in products, requires safe and ethical
European implementation, and ranks AI systems by risk. High-risk systems require impact
Union assessments and audits, and unacceptable risk AI, like real-time biometric monitoring and
social scoring systems, being forbidden in the EU.

India currently does not plan to regulate the growth of AI, considering it a "significant and
India strategic" sector for the nation, focusing instead on fostering a robust AI ecosystem and
promoting entrepreneurship and innovation in the field.

AI regulation in Australia currently is based on a voluntary AI Ethics Framework, and

existing laws covering privacy, consumer protection, anti-discrimination, and intellectual
Australia property are applied to AI. In the future, Australia is likely to align its AI policy with the EU
to facilitate international collaboration, data sharing, and cross-border AI deployment.
This also is likely to help Australian companies comply with EU regulation.

China is rolling out comprehensive regulations (2023) governing artificial intelligence (AI),
including measures on recommendation algorithms and synthetically generated content
China (deep synthesis), and strict draft rules on generative AI (AI chatbots like ChatGPT), with a
goal of information control and requiring developers to make filings to the algorithm registry.

UK abides by EU GDPR policy but not the EU AI Act (currently creating similar policy for AI Use)
1

16
 05 l Overall Telecommunications Perspective

“The most significant pace determinant is government regulation, which

could potentially make it illegal to build a model using customer data.
This is particularly relevant in regions with strict data regulations like the
UK and EU.
Strategy Director
Wireless CSP, USA

“ Especially for European telcos, we are highly cognizant of how certain

data types are deemed confidential and therefore prohibited from being
used to train AI models, and the legal restrictions on data movement
under GDPR.
Head of Data & Personalization
Wireless CSP, Western Europe

APAC lags slightly at 16% current average use case adoption despite featuring many leading CSPs
looking to implement more innovative and top-line impacting use cases. According to our survey,
APAC CSPs ranked their data capabilities much lower than their North American and European
peers. This suggests that CSPs in the Asia-Pacific region believe they need to do more in cleaning
up their data assets before they can fully realize benefits from generative AI. Additionally, APAC
telecommunications executives perceive current generative AI models to be limited in their ability to
handle non-English languages6. These perceptions may be easing through efforts such as AWS’s LLM
development program7, working to help Japanese organizations planning to build their own LLMs.

“Local languages are a factor to consider, especially in Asia. For instance,

ChatGPT is less effective in Thai than in English. Companies operating
in smaller countries might need to train parts of the language model
themselves to get the same effectiveness.
Head of AI
Wireless CSP, APAC

“ Adoption of generative AI in the APAC & MENA region has been

reported to lag due to language barriers. Generative AI needs to be
localized to understand our language and its different dialects.
General Manager Advanced Analytics
Wireless CSP, APAC

17
 05 l Overall Telecommunications Perspective

Our study reveals CSPs anticipate a sustained increase in investment in generative AI. CSPs already
adopting generative AI are allocating less than 1% of their total technology expenditure towards these
capabilities. However, 45% of CSPs expect this spending on generative AI use cases to surge to 2-6%
within the next two years. This suggests that CSP spending on generative AI could grow by as much
as six times in the near future as CSPs scale implementations and explore newer use cases, especially
as familiarity with the technology grows.

“I assume generative AI spend as a percentage of total technology budget

is small, maximum 1% today. However, I can see it growing through to
3-6% for big telcos.
SVP Global Systems
Wireline CSP, Western Europe

“Of our total technology budget, around 30% is allocated to AI and data.
Within that, generative AI accounts for about 1 to 2%. This is primarily
due to many use cases being in the proof-of-concept stage and not yet
scaled. Over time, I predict that data and AI will become 50 to 60% of
the overall spend, and generative AI will become 30 to 40% of the overall
spend.
Director of AI and Product Management
Wireless CSP, USA

Organizationally, CSPs need to navigate several stakeholders when adopting generative AI. Both
CSP business domain and IT stakeholders are critical in evaluating and implementing generative AI
solutions. Leaders of business function departments (e.g., customer service, marketing, network, etc.)
are involved in the ideation processes for generative AI and shepherding proof of concept with other
stakeholders such as legal or risk/compliance. On the other hand, IT departments have less of a role
in evaluation. They are often the key owner of the generative AI use case implementation, especially
around deployment choice, data implications, and maintenance and governance of the data and
model. Together, these two key stakeholders determine whether generative AI has a place in the
organization, and which use cases to deploy.

18
06
Generative AI
Telecommunications-
Specific Use Cases
 06 l Generative AI Telecommunications-Specific Use Cases

CSPs are optimistic about the potential for generative AI to create significant and incremental value.
They see the technology as a means to streamline existing processes, drive innovation, create new
opportunities, and unlock new sources of value in the telecommunications industry. In many of our
executive discussions, the most frequently mentioned KPIs for generative AI were reduction in call
volume, decreased time to resolution, and headcount reduction due to efficiency gains, though CSP
executives acknowledged realization of these benefits is still developing.

“ The primary focus is on reducing the volume of incoming calls, which

number in the tens of millions and are largely avoidable. The goal is to
resolve these issues with a quick chatbot interaction, which would result
in significant cost savings.
VP of Strategy
Cable CSP, USA

As part of the study, we tested 17 use cases across four domains (Product & Marketing, Customer
Service, IT, and Network). Unanimously across varying global CSPs, the foremost use case being
actively implemented was a customer chatbot, emphasizing the paramount importance of customer-
centricity.

Nonetheless, during our conversations, it became evident that these near-term implementations are
limited in their transformative potential. Numerous CSPs have articulated their reliance on rule-based
chatbots, previously utilizing AI/ML technologies. Executives expect to use generative AI to provide
a more conversational, “chatty,” human-like interaction to extend existing capabilities and features
to provide information to customers, thereby reducing call volume. However, in the longer term,
executives envision using generative AI to drive real-time customer action and decisions without
human intervention.

20
 06 l Generative AI Telecommunications-Specific Use Cases

Exhibit 6: CSP Generative AI Use Case by Stage of Adoption1

% of total respondents, N varies by respective domain

1
% of total respondents who noted they are currently implementing & currently evaluating with high likelihood to adopt the use case

“ The primary use case [for generative AI] is providing information to

customers, enabling them to find answers independently, reducing the
need for call center interactions. The next step is to use generative AI to
initiate tariff changes, upgrades, or contract renewals; however, these
require real-time actions and integrations. The complexity of these
integrations, especially in telecom companies with many legacy systems,
can be a significant hurdle. Therefore, most early use cases are more
focused on general information.
Former Head of Data

“
Wireless CSP, Western Europe

We're exploring how much intelligence generative AI can inject into these
interactions and how customers might engage with it. It's not just about
pushing information to the customer; we want to facilitate a dialogue where
we can notify them of important issues, potentially ask them to make a
decision and answer any questions they might have before they make that
decision. We just don't know enough as to how that's going to go.
Former VP Managed Services
Cable CSP, USA

21
 06 l Generative AI Telecommunications-Specific Use Cases

Other use cases seeing immediate traction often enable operational efficiency, such as guided
employee assistance and documentation in customer service and IT. In fact, improving efficiency and
productivity by using generative AI is by far the highest value driver. Forty-one percent of CSPs see
generative AI as assisting employees in producing quicker and higher quality output.

Exhibit
Exhibit 7:
7: Communication
Communication Service
Service Providers’ Data Capabilities
Providers’ Data Capabilities1
% of unique responses that chose driver for use cases selected as adopted, highly likely, or interested; N varies by respective domain
Weighted average on scale 1 (less proficient) to 5 (more proficient), N = varies by respective domain
Please choose the top driver for selection for each use case your organization is interested in and in the process of adoption

Majority
Increased Effi ciency Internal Employee Customer
Competitive Advantage Experience Experience
Cost Savings & Productivity
Providing a more Providing a more # of Respondents
Automating tasks Assisting employees Producing more revenue and
user-friendly experience user-friendly experience for a Selecting Use Cases As
and replacing human in producing quicker and more offering a better end-product
for a given process given element of the customer Adopted, Highly Likely,
labor / other tools quality output as opposed to competitors
for employees journey Or Interested

Product & Marketing 8% 33% 37% 4% 18% 63

Customer Service 28% 41% 3% 15% 13% 55

Network Planning 29% 54% 13% 4% 7% 31

Network Operations 22% 49% 2% 20% 3% 41

IT Operations 27% 26% 0% 44% 1% 52

IT Software Development 22% 62% 8% 7% 1% 50

% of Unique Use Case

Responses that 20% 41% 18% 14% 13%
Selected Top Driver

1
Please choose the top driver for selection for each use case your organization is interested and in the process of adoption

Many of these productivity use cases are considered easy to implement and lower risk because they
can largely leverage existing pre-trained, off-the-shelf models alongside existing knowledgebase and
documentation to enable summarization and accessible search functions.

“
Our decision to use a generative AI tool is based purely on productivity.
While it's beneficial if the representatives perform better with the tool, the
primary goal is to improve their performance.
CEO
Wireline CSP, USA

22
 06 l Generative AI Telecommunications-Specific Use Cases

The healthy tension between business value and implementation complexity helps CSPs understand
which use cases to prioritize and how fast can these be adopted. Our study highlights the nine
specific high-impact use cases for CSPs, which we qualify as Clear Wins and Big Bets.

Exhibit 8: Communication Service Providers’ Generative AI Use Cases Priority Segmentation1

Respondents based on median scores, N varies by use case

1
Please rate each use case your organization is in the process of adopting by ease of implementation and incremental business benefit on a scale of 1 to 5

23
 06 l Generative AI Telecommunications-Specific Use Cases

Clear Telecommunications generative AI use cases with high incremental business benefit and less
difficult implementation effort.

Wins
These use cases often focus on employee assistance, such as contact center documentation
and network operations knowledge management. Of the “clear win” use cases, nearly 41% of

Network Operations
Guided employee assistance for installation
troubleshooting, and maintenance
Retrieve and respond to network service engineers
with information to aid in installation,
troubleshooting, and maintenance network devices
and infrastructure like routers, switches, and
firewalls

Customer Service
Contact center documentation
Generate transcription of calls,
summarize customer interactions and
suggest follow-up actions

Customer Service
Customer chatbot
Interpreting and responding to customer
queries and requests in a human-like
conversational manner

IT Software Development
Automated code documentation
The use of automated tools to generate
documentation for code that is similar to existing
documentation, often using natural language
processing or non-generative AI

Product
Insight generation
Generating new customer insights to
inform product enhancements and
service improvements based on customer
information

24
 06 l Generative AI Telecommunications-Specific Use Cases

Big
Telecommunications generative AI use cases with high incremental business benefit but more
difficult implementation effort.
As noted, many of our discussions with CSP executives mentioned the potential for generative AI

Bets high value but more complex in implementation, and hence, not use cases for immediate adoption.

Marketing
Personalization
Generating tailored marketing messages
and content based on customer information
(e.g., demographic, history, behavior)

Network Planning
Generative route/network design and network
configuration
Generate a variety of possible network designs
optimized for key parameters (e.g., topography,
costs, power, etc.)

Network Planning
Synthetic data generation for security testing
Artificial data that mimics real-world data, which
is then used in testing security systems and identify
vulnerabilities without needing to use real customer
information

IT Software Development
Automated code generation, debugging, testing
The use of automated tools to create code,
find and fix errors in code, and test code

25
 06 l Generative AI Telecommunications-Specific Use Cases

Outside of guided assistance for network technicians, CSPs are hesitant to roll out network planning
use cases since they involve sensitive network data. In fact, while generative AI holds promise for
optimizing network planning and management, networks’ complex and critical nature drives a long-
term outlook on this front. One CSP executive noted using generative AI for asset management of field
technicians, dispatching them before an outage occurs while en route to a scheduled preventative
maintenance job.

Although network planning use cases are often regarded as significant strategic investments (“Big
Bets”), our extensive discussions with CSPs have revealed a notable degree of uncertainty regarding
the perceived value of these investments compared to existing machine learning technologies already
deployed. For example, a large CSP in Europe mentioned that AI is already being used in network
design generation. While acknowledging the potential of generative AI, they do not foresee generative
AI delivering significant incremental value in this particular context.

“ We avoided network data due to its sensitive nature. Instead, we

considered integrating less sensitive internal data, such as internal generic
data. This approach allows us to experiment without risking significant
consequences, which happen to be often the knowledge management or
productivity use cases..
Strategy Director

“
Wireless CSP, USA

There are numerous ideas in network planning, although none are

concrete at this point. Some examples include network design, where
network designs and configurations are fed to generative AI to create
more optimal or innovative designs.
Head of AI
Wireless CSP, USA

Furthermore, while not classified as a “Clear Win” or “Big Bet” in our analysis, CSP in-depth
discussions also highlighted their interest in the concept of a “customer service employee coach.”
These executives recognize the potential of generative AI-powered employee coaches, which can
leverage data from customer interactions to offer coaching and performance enhancement guidance
to customer service representatives.

“I see generative AI also for customer service and sales AI coach. We have
also experimented with AI tools through a private company for real-time
sales coaching, primarily for commercial use. These tools listen to calls
and provide post-call coaching to the representative, including directions
for follow-up.
CEO
Wireline CSP, USA

26
 06 l Generative AI Telecommunications-Specific Use Cases

“ We have been on a journey to implement data-driven coaching and

providing data-driven work for front-end employees. We are giving sales
representatives access to performance data on a bunch of KPIs, so you
know exactly where they need to improve.
Director of AI
Wireless CSP, USA

Many CSPs noted piloting and using generative AI for software code development. For example,
a major European CSP described using the technology for writing code, leading to a 30-45%
productivity gain in a trial with around 250 developers.2 Interestingly, we also heard some reservations
regarding tools’ security and ability to handle complex code. For example, some CSPs indicated
concerns about using existing offerings, given the need to keep code in-house and secure.
Additionally, some perceive generative AI to be more suitable for less complex code but not as well-
suited for more complex, unique, telecommunications-specific code.

“ Our software development team [has] developed our own generative AI

coding function to keep our coding in-house and prevent it from being
released externally.
Strategy Director
Wireless CSP, USA

“
In terms of coding, there is a debate. I believe that a significant portion of
code, perhaps 80 percent, is common, and only 20 percent is unique. The
common code, such as expanding a database by a column, adding a field,
or moving it, has already been written. We need human developers for the
unique code that is specific to us. Generative AI is not awesome but also
not horrible at creating code.
Former SVP Technology Architecture
Wireless CSP, APAC

We also examined how the perception of use cases differs among CSPs based on their data
capabilities. Among the top 30% of CSPs with advanced data capabilities (based on evaluation of 8
distinct capabilities covered in the next section), there is a shift in focus from only productivity tools to
product and marketing use cases that directly impact revenue generation and competitive advantage.
Unlike their peers, these CSPs see product and marketing use cases like search term generation and
insight generation as Clear Wins that enable a competitive advantage.

27
 06 l Generative AI Telecommunications-Specific Use Cases

Exhibit 9: Clear
Exhibit 9: Clear Win andWin and
Big Bet UseBig Bet
Cases Use Cases
for Leading for
& Lagging Leading
Data & Lagging Data Capable CSPs
Capable CSPs

Leading Data Capabilities Lagging Data Capabilities

High Data Readiness, Top 30% Low Data Readiness, Bottom 70%

1. Customer Service: Contact center

1. Product & Marketing: Search term generation documentation
Clear Win Use Cases 2. Customer Service: Customer chatbot 2. Customer Service: Customer chatbot
(High business value and 3. Product & Marketing: Insight generation 3. Network Operations: Guided employee
less difficult implementation) 4. Customer Service: Guided employee assistance for installation, troubleshooting,
assistance for knowledge management and and maintenance
troubleshooting 4. IT Operations: Employee self-service
desk

1. Product & Marketing: Personalization 1. Network Planning: Synthetic data

2. Network Operations: Guided employee generation for security testing
assistance for installation, troubleshooting, and 2. Product & Marketing: Personalization
Big Bet Use Cases maintenance
3. IT Software Development: Automated
(High business value and 3. Network Planning: Generative route/network code generation, debugging, and testing
more difficult implementation) design and network configuration
4. Network Planning: Generative
4. Network Planning: Synthetic data generation route/network design and network
for security testing configuration

Product & Marketing Network Planning Network Operations Customer Service IT Operations IT Software Development

Our conversations with larger CSPs also noted interest in product-related use cases. Recently,
a European CSP announced using generative AI to create new offerings in the sports market,
expanding their capabilities and understanding of how generative AI can be monetized (within the
media environment).8 Another example is a CSP from the Asia-Pacific region, which has developed
a consumer-focused application that enables users to interact with its digital services such as music
streaming, payments, and actions like sending texts and managing schedules.9 This is in line with our
survey findings, where APAC CSPs noted product and marketing use cases such as personalization
and new future ideation as high-value opportunities.

“ The goal is to transition from being a pure infrastructure player to

offering applications and service, and generative AI can be key to
unlocking those product offerings.
Head of Data Science
Wireless CSP, APAC

Meanwhile, European CSPs notably view customer service use cases as primary focus areas
and Clear Wins. One large European CSP explained that cost pressure, while universal in the
telecommunications industry, is particularly palatable in Europe. Thus, the focus is acutely on lowering
service costs, making generative AI use cases for customer service even more attractive.

Generative AI is an emerging technology with significant potential across a multitude of use cases in
the telecom industry. Certain use cases are poised for earlier implementation, promising immediate
benefits. In contrast, others remain firmly rooted in the longer-term horizon, requiring comprehensive
planning and adaptation to fully realize their potential.

28
07
Constraints &
Complexities
 07 l Constraints & Complexities

As CSPs delve deeper into deployment of generative AI, realizing these opportunities is not without
challenges and complexities. The biggest concern many CSPs share is the current state of data
practices and processes within the organization. Specifically, data security, privacy, and governance
considerations are noted as top concerns.

Exhibit
Exhibit 10: 10:
TopTop 3 Technical
3 Technical Concerns
Concerns1 1

% of
% total respondents
of total responding
respondents responding in in
thethe
toptop
3; N3;= N = 102
102

61% “We are also considering how to ensure data security when
working with providers, whose servers are based in the U.S.
Security, privacy, and data We need to ensure that our data is secure and not used by
governance concerns of data other actors.”
used by generative AI models
Head of IT Architecture and Infrastructure Strategy,
Wireless CSP, Western Europe

“Many telcos do not have the in-house expertise to

51% build their own LLM. Are they willing to take the risk
and invest in building a large LLM? This involves cost,
Lack of internal technical manpower, and potentially changing the company’s
expertise to implement direction.”
and manage generative AI
CTO, Wireless CSP, APAC

“Ensuring reliability is key to accelerating the adoption

46%
Concerns about accuracy
of generative AI. The tool is not fully mature yet, so
the quality of the output must be reliable enough for
production use without fear of errors.”
& reliability of generative
AI output Head of IT Architecture and Infrastructure Strategy,
Wireless CSP, Western Europe

Choose and rank the top 3 technical concerns/challenges related to adoption and scaling of generative AI-related use cases at your organization.
1

Informed by our discussions with executives, we analyzed data dependency concerning generative
AI, focusing on three specific dimensions of data capabilities: the what of data (i.e., data infrastructure),
the who of data (i.e., data fluency), and the how of data (i.e., data processes).

30
 07 l Constraints & Complexities

Exhibit11:
Exhibit 11:Communication
CommunicationService
ServiceProviders’
Providers’Data
Data Capabilities
Capabilities

Data Infrastructure Data Fluency Data Processes

The What of Data: The Who of Data: The How of Data:

“I need to invest in things” “I need to hire or train people” “I need to change my organization”
• Centralized Data Management • Data Analytics • Machine Learning (ML) Operations
• Data Quality & Integration • In-House Technical Expertise • Data Governance
• Scalability of Data Infrastructure
• Data Infrastructure Modernization

Using the evaluation of these distinct data capabilities, most CSPs view their state of data as overall
sufficient but lacking specifically in data governance. This remains a top concern even for the top 30%
of data capable CSPs.

Exhibit 12: Communication Service Providers’ Data Capabilities1

Weighted average on scale 1 (less proficient) to 5 (more proficient), N varies by respective domain

1
For each of the following components of your data environment, please rate your organization’s proficiency from 1 to 5; 2) Per market experts, weights are
60% data infrastructure, 15% data fluency, and 25% data processes

“ A lack of data governance and data ownership procedures are slowing the
pace of generative AI adoption at telcos.
Senior Executive Engineering
Wireless CSP, USA

31
 07 l Constraints & Complexities

“Data is important from two main standpoints: data availability and data
governance, including processes around who can access that data.
Director of Product Management
Wireless CSP, USA

It is important to note that while CSPs consider data infrastructure to be more “ready” for generative
AI and relatively more proficient, they also acknowledge, at times, data siloes are inevitable. Forty-six
percent of CSPs agreed their internal data is highly siloed and not well integrated compared to the
34% that disagreed.

Exhibit 13: CSP’s Data “Readiness” Sentiments 1 1

Exhibit 13: CSPs’ “Readiness” Sentiments
% of respondents, N = 102
% of responses, N = 102

Agree Neutral Disagree

My organization’s data is highly siloed and not

46% 20% 34%
well integrated across multiple sources

My organization lacks a well defined data

governance framework and policies which is 33% 21% 46%
consistently followed and complied with

My organization’s data is largely on-premise

(not on cloud) which will prevent faster 31% 21% 48%
adoption of generative AI

Data ownership within my organization is 30% 15% 55%

unclear

My organization’s data platform is not in a 36% 18% 46%

current state to support generative AI

My organization does not have a clear

understanding of generative AI data 39% 20% 41%
dependencies

1
Please indicate your organization’s sentiment on the following statements regarding your organization’s data capabilities.
Please indicate your organization’s sentiment on the following statements regarding your organization’s data capabilities.
1

“ Integrating messy data across multiple vector databases is critical. We

need to know where in the cloud the data is and where it’s moving.
SVP Global Systems
Wireline CSP, Western Europe

Given that the current state of data capabilities is top of mind for telecommunications companies, we
took a closer look at CSPs that are more data capable and lead in generative AI adoption. As noted
earlier, these data capable CSPs exhibit greater emphasis on near-term use cases focused on revenue
generation. They also consider these use cases to be, on average, 8% less complex to implement on a
data readiness rating scale.

32
 07 l Constraints & Complexities

Data-capable CSPs exhibit common organizational characteristics:

• Dedicated AI Center of Excellence (COE) – Many CSPs with data-leading capabilities

often have a COE for AI and generative AI. Where there are centralized AI/ML
departments, these COEs become AI product management functions to the business
domains, helping to streamline evaluation and implementation.

• Pervasive Use of Advanced Analytics – CSPs with data-leading capabilities often use
advanced analytics throughout the organization, not just certain business domains or
specialists. These CSPs make it their mission to enable users within the organization to
access data on a self-service basis and use it for data-driven decision making. Such
companies encourage a data-driven culture, have strong governance practices in place,
and foster greater adoption of new AI technologies.

• Cloud-Based Modern Data Infrastructure – CSPs with modern, cloud-based

data infrastructure are likely to have data architectures that can integrate data from
disparate sources. These CSPs are better prepared to meet the security and regulatory
requirements for handling sensitive data, and scale to handle the computational
requirements to train and deploy generative AI models.

“ The generative AI will likely accept many different forms of data, as

opposed to requiring multiple machine AI engineers. It requires a lot of
manipulation to understand different facets as I'm taking data sources
from network data, contact centers, and my respective ML learning
models. This process is time-consuming. That's why I advocate for more
centralization.
Director of AI and Product Management
Wireless CSP, USA

Telecom companies’ data capabilities and maturity play a pivotal role in shaping their adoption of
generative AI. Companies with advanced data capabilities tend to be more proactive in embracing
generative AI, leveraging their proficiency to experiment and adopt applications that address both
efficiency and revenue generation objectives.

33
08
Deployment &
Implementation
 08 l Deployment & Implementation

The discussion about improving data capabilities and proficiency is crucial because CSPs expect to
primarily use pre-trained, off-the-shelf foundation models, but train these models on CSP-owned,
internal data. Across all business domains, more than 65% of CSPs intend to use internal proprietary
data for generative AI use cases. This reliance on internal data is consistent with CSPs’ objectives
of customizing pre-trained models, tailoring them to handle industry-specific terminology and
context, and ultimately enhancing output quality. Hence, having a robust data infrastructure and
governance is essential to ensure that the foundation model is trained on high-quality and relevant
telecommunications data.

“ Many telcos do not have the in-house expertise to build their own
LLM, so they might opt for a pre-trained model.
model The challenge lies in
the expertise and the company's vision. Are they willing to take the risk
and invest in building a large LLM? This involves cost, manpower, and
potentially changing the company's direction.
CTO
Wireless CSP, APAC

“ Every model will eventually need to be fine-tuned in some way. I don’t

believe generative AI models should be plug-and-play. To take the models
to the next step, we should look at the architecture as fit. If the data is not
understood and missing confidence, then it will be inaccurate.
EVP
Wireless CSP, USA

CSPs are broadly considering two relatively different approaches for adapting large language
models (LLMs) for specific tasks: fine-tuning or prompt and context learning. These techniques vary
based on whether the organization will be customizing the training of the foundation model versus
inference. Fine-tuning trains the foundation LLM by adjusting its parameters and feeding the model
additional data. On the other hand, prompt engineering and context learning is used to customize the”
inference,” i.e., the output of the generative AI model to make it relevant to the domain context while
keeping the foundation model parameters frozen.

35
 08 l Deployment & Implementation

Exhibit 14: Fine Tuning vs. Context Learning & Prompt Engineering Comparison
Exhibit
% of total14: Fine Tuning
respondents; vs. Context Learning & Prompt Engineering Comparison
N = 102

Inference Training

An external process that utilizes any base model A modification made to the model that
deployed (fine-tuned or not) and enhances input, fundamentally alters its parameters and underlying
output, or both for a specific use case weights into a new model

Prompt Engineering & In-Context Learning Fine-Tuning

The ability of generative LLMs to adapt their Taking the weights of an already trained model
responses based on the context provided in the and using them as initial weights towards the training
prompt. They maintain a record of prompts in internal of a new model with new training data.
memory and can reference other data provided with
the prompts.

Reinforcement Learning
A method by which a model can automatically
modify its framework over time based on maximizing
rewards and minimizing penalties associated with its
responses to previous tasks.

Lower Cost and resources/data/compute needed Higher

Performance (e.g., accuracy and quality of output)

Fine-tuning can improve model performance on specific tasks by adjusting a sub-set of model
parameters and training the model on a curated task-specific dataset. In contrast, prompt engineering
and context learning involve providing the model with a small number of task-specific examples as
input. The model then uses these examples to quickly adapt to the task at hand and enhance input,
output, or even both in specific cases. This approach can be more resource efficient than fine-tuning,
as it requires less training data and computational resources with an output noted as ”good enough.”

CSPs are concerned about the costs required when implementing generative AI. In fact, 31% of CSPs
note concerns around significant operating costs to maintain the model and infrastructure (i.e., opex).
Since in-context learning is the cheaper technique as it does not involve any form of training, it is
attractive to CSPs given its low resource intensity, only referencing other data provided with prompts
for a sufficient-enough output.

But the jury is still out on trade-offs between prompt engineering/context learning and fine-tuning,
and their suitability for the various use cases. Both are being explored: 38% of CSPs preferred using
prompt and context learning and 40% for fine-tuning to train off-the-shelf pre-trained LLMs. In fact,
one CSP executive noted investing in both fine-tuning and prompt/context learning, with the former
showing only slightly better results for some use cases. They also stated that fine-tuning required
more time, expensive talent, and higher compute resource costs since they were fine-tuning a multi-
billion parameter model. In contrast, prompt and context learning was less costly and allowed them to
experiment with various open-source models.

36
 08 l Deployment & Implementation

“ Right now we’re indexing our data in a real-time manner and improving
prompts and pre-prompts. We’re vectoring data for fast results, but we
intend to look more into fine-tuning once GPT3 & 4 allow fine-tuning.
VP Data & AI
Wireline CSP, USA

Exhibit15:
Exhibit 15: Communication
Communication Service
Service Providers’
Providers’ Preferred
Preferred Generative
Generative AI Implementation
AI Model Implementation Method1Method1
%ofoftotal
% responses, NN
respondents, varies bybyrespective
= varies domain
respective domain

Purchase off-the-shelf model with no 7%

additional training

Purchase off-the-shelf model and utilize

prompt engineering OR provide data for 38%
context learning

Purchase off-the-shelf model with fine-tuning

40%

Build in-house model 15%

Please indicate the likely implementation method for generative AI models for the business domain(s) seslected.
1

Please indicate the likely implementation method for generative AI models for the business domain(s) selected.
1

“
We invested in both fine-tuning and prompt engineering and found similar
results for a number of use cases. There were few use cases where fine-
tuning was slightly better. However, the fine-tuning team required more
time, more expensive talent, and significantly more money on GPUs. As a
result, we shifted our focus to prompt engineering.
Director of Technology
Wireless CSP, USA

It is also worthwhile to note that there are additional techniques for customizing pre-trained LLMs,
such as Retrieval-Augmented Generation (RAG) and reinforcement learning through human feedback.
RAG works at the inference level, improving output of the LLM by allowing access to additional data
sources (including organization’s own data) and retrieve up-to-date information without modifying
the underlying model. Reinforcement learning, on the other hand, is a method by which a model
can automatically modify its framework dynamically over time based on maximizing “rewards” with
its responses to previous asks. However, reinforcement learning can also be expensive if it utilizes
feedback from humans (e.g., ranking answers). The associated cost can be a function of how many
answers are provided and how thorough the human feedback process is.

37
 08 l Deployment & Implementation

While a notable 15% of CSPs indicated interest in building foundation models in-house, this number
seems somewhat inflated. Upon closer examination and further discussions with CSPs, an emerging
trend becomes apparent: the in-house development of small custom models tailored to telecom-
specific needs. These smaller models utilize the LLMs as a framework and are then customized for a
specific purpose. For instance, one CSP remarked crafting a custom language model with fewer than
15,000 parameters—a significant contrast to the billion-parameter scale of open-source GPT models.
They integrated this model with the LLM to ensure that the CSPs’ distinctive voice and communication
guidelines were reflected in the results. Similarly, a major CSP is developing an in-house model with
over 200 million parameters, which will be used to fine-tune and integrate with third-party LLMs.
By using these smaller models to tailor LLMs, telecom providers can improve their performance in
specific tasks and improve capabilities such as multilingualism and interpretability. This aligns with
our survey results, where some CSPs mentioned developing their in-house models, but often on a
more modest scale. Forty-three percent of CSPs also indicated a need to integrate both a large and a
smaller language model into their systems.

“ I’m seeing a lot more teacher-student models where a larger, more

complex model (the teacher) is used to train a smaller, simplified model
(the student). The student model is usually more efficient/less expensive.
Head of Strategy and Analytics
Wireless CSP, APAC

“ Our solution takes GPT and adds a smaller customized LLM on top of
GPT to ensure it has our telco ‘voice’ and/or communication guidelines
come through in the output versus general English.
Strategy Director
Cable CSP, USA

“ We took the initiative to build our own large language models (LLMs)
for specific use cases and domains because there was nothing available
in our country and we wanted to be first movers in that space. From an
efficiency and latency perspective, we believe there are smaller models that
we can train and use for certain applications.
Senior Director of Technology and Business Development
Wireless CSP, APAC

Supporting telecommunication endeavors to customize LLM, multiple enabling platforms and

capabilities are emerging to help CSPs explore models & approaches to leveraging their internal data
for foundation models. For instance, 65% of CSPs noted context-learning tools and services as the
most important generative AI product offering to CSPs, with a high 54% citing fine-tuning tools and
services as the most important.

38
 08 l Deployment & Implementation

Exhibit
Exhibit16:16:
Generative AI Product
Generative Offerings
AI Product Most Important
Offerings to CSPs1 to CSPs1
Most Important
% of total respondents ranking in the top 3, N = 102
% of total respondents ranking in the top 3, N = 102

Providing context-learning
65%
tools and services

Providing fine-tuning tools and 54%

services

Provide off-the-shelf
applications
40%

Provide off-the-shelf models 34%

None of the above 3%

1What product offerings are most important to you when selecting a generative AI solution or provider?
1
What product offerings are most important to you when selecting a generative AI solution or provider

An emerging, notable development is the rise of generative AI-managed service platforms and
Model-as-a-Service (MaaS) providers. Managed service providers, for instance, can offer licenses to
multiple pre-trained models, enabling companies to access cutting-edge generative AI models via the
cloud. This allows organizations to generate innovative content without starting from scratch, akin to a
subscription-based software model.

Generative AI fully managed service platforms enable enterprises to integrate multiple models across
various premises without being locked into a single provider. Forty-four percent of CSPs plan to use
fully managed service platforms like Hugging Face or Lang Chain for their generative AI models.
These emerging business models (i.e., Managed Services Platforms and MaaS) offer end-to-end
service solutions without the infrastructure and cost demands of a fully custom, in-house build.

“ I strongly advocate for using third-party vendors who are experts in

this domain. Many agile leaders share this view. Some leaders at other
companies believe they can build it themselves, but we [CSPs] are not in
the business of making models – generative AI and foundation models are
not our core business.
SVP of Customer Service
Wireline CSP, USA

39
 08 l Deployment & Implementation

“ The real issue [for CSPs] is the lack of [technical] resources. Many
partners and large language model providers understand that
telecommunications companies may not have the necessary resources. And
vendors provide their own studio and professional services and help with
fine-tuning or help CSPs to fine-tune themselves. I personally believe that
this will be the key accelerator for generative AI in telcos.
General Manager of Advanced Analytics
Wireless CSP, APAC

While much attention is understandably devoted to the complexity of model customization and
deployment, it is essential to consider the application layer. Interestingly, 69% of CSPs prefer
purchasing off-the-shelf end-user application interfaces for their generative AI solutions rather
than building them. This trend spans business domain use cases, particularly focusing on product
& marketing and customer service. The rationale behind this preference lies in the convenience of
ready-to-use applications, which accelerates time to market and minimizes maintenance.

“ For example, in the case of conversational AI, specifically chatbots, the

customer interface is vendor-based.
vendor-based Every time a customer interacts with
the chatbot on our website or app, it launches a code base. We then
integrate all the conversational AI components.
Technology & Innovation Director
Wireless CSP, USA

40
09
Implications &
Outlooks
 09 l Implications & Outlooks

Generative AI adoption within the telecommunications industry is only beginning and will continue to
evolve dynamically. This study underscores:

• “Clear Win” use cases, such as customer chatbots and employee assistance, are ripe for
implementation. Many of these use cases require a lighter implementation touch and can
use an off-the-shelf model with some adjustments.

• More transformative high-value use cases, or “Big Bets,” such as network planning and
optimization, are of interest but dependent on high-quality and sensitive internal data.
Establishing strong data governance practices becomes imperative for CSPs seeking to
prepare for potential use cases, especially those related to these “Big Bets.”

• CSPs should explore various implementation and deployment options and be aware that
the future of generative AI in the telecommunications space will be shaped by technology
developments, such as managed service platforms and tools enabling easy customization
and development of LLMs (e.g., Amazon Bedrock10 enabling enterprise developers to
evaluate and fine-tune foundation models of choice using custom data or build generative
AI apps from scratch), and AI regulations.

Telecommunication companies are confronted with a multitude of technical and operational

challenges when harnessing the potential of generative AI. A pivotal aspect of this journey involves
a review of data governance practices to ensure the data infrastructure supports compliance to data
security, privacy, and sovereignty requirements. CSPs also need to formulate effective strategies for
managing data scattered across multiple systems for seamless access to high-quality data.

The significance of high-quality data cannot be overstated, especially when it comes to training
generative AI models for telecom-specific tasks and contexts. The reliability of these models, as well
as the transparency and trust they inspire through their outputs, hinges on the quality of the data they
are trained on.

While many CSPs prefer off-the-shelf models to expedite their generative AI initiatives, the off-the-
shelf approach comes with its own set of considerations. Telecom companies are acutely aware of the
importance of model customization to suit their unique requirements. This necessitates external
vendor platforms that not only provide access to pre-trained models, but also support various
methods for model fine-tuning, model training, and model development.

42
 09 l Implications & Outlooks

Moreover, the choice of generative AI deployment heavily weighs the security and privacy concerns
of data used by generative AI models. As discussed in the previous section, 61% of CSPs noted
security as the top technical concerns when considering generative AI.

However, there are increasingly more 3rd party solutions to ease these concerns for CSPs who
are considering implementing generative AI. These external vendor platforms are seen as critical
accelerators in the telecommunications industry’s AI journey. They not only bridge the expertise
gap but also empower telecom companies to leverage the advantages of generative AI without
compromising on the customization and control they seek. AWS, for instance, is democratizing
access to this technology and providing flexible approaches to meet enterprises where they need -
whether that be providing infrastructure to build their own foundational models, leveraging pre-trained
foundational models to build their applications, or fine-tuning available models with additional data.

Amazon Bedrock allows CSPs the flexibility to work with publicly available models or build their own
foundation models based on custom silicon (Inferentia and Tranium chips), in addition to instances
built using leading chip providers, to offer price performance infrastructure. With platforms such as
Amazon Bedrock, CSPs can strike a delicate balance between expediency and precision, ensuring
they stay at the forefront of innovation in their highly competitive landscape.

Regardless, the fact remains that generative AI is already making waves in the telecommunications
industry with multiple use cases being tested and implemented in the near future. For communication
service providers who are looking to better understand the relative risk and technical complexity
when adopting certain generative AI use cases, here are tactical questions to consider:

Exhibit
Exhibit 17:17:
KeyKey Questions
Questions to Understand
to Understand Relative
Relative Risk and TechnicalRisk and Technical
Complexity of Generative Complexity
AI Use Cases of Generative AI Use Cases

Relative Tech
Question Options Relative Risk
Complexity
Common General Purpose (Knowledge search, guided
Lower Lower
What type of use case is under assistance)
considerations?
Highly domain specific & niche Higher Higher

Customers (external-facing) Higher Neutral

Who does the use case target?
Employees (internal-facing) Lower Neutral

Customize & adjust off-the-shelf, pre-trained models Higher Lower

What foundational model will be used?
Build in-house Lower Higher

Prompt/context learning Neutral Lower

What methodology will be relevant to make
the off-the-shelf model fit?
Fine-tuning Lower Higher

Internal proprietary, own data Higher Higher

What data will be used to train/build the
foundational model?
Externally available data or synthetic data Lower Lower

In databases, which are available and accessible by the team Lower Lower
Where is the internal data stored that will be
used to train/build the model?
In multiple siloed databases, which are not well integrated Neutral Higher

Sensitive data such as customer interactions, network data Higher Higher

What type of internal data will be used?
Documentation, knowledge databases Lower Lower

Lack clarity on data governance or data governance

Are data governance, policies, and Higher Higher
policies not rigorously followed
ownership, which can impact the use case,
clearly defined? Data governance clear and well-defined Lower Lower

43
 09 l Implications & Outlooks

The fusion of generative AI and telecommunication isn’t just innovative; it’s imperative. Having clear
answers to use case types, deployment preferences, and data dependencies, communication service
providers can ready themselves for the transformative journey generative AI can enable. To thrive
in this dynamic landscape, companies must embrace weaving generative AI into the fabric of their
operations, and this study can arm them with the insights needed to navigate concerns.

44
10 Appendix
 10 l Appendix

Glossary

Artificial Intelligence (AI) Latency

Artificial intelligence (AI) is the simulation of human The amount of time it takes for a packet of data to travel
intelligence in machines that are programmed to think from its source to its destination, often measured in
and learn. milliseconds (ms).

Context Learning Machine Learning (ML)

A process of using internal data to provide prompts Machine learning (ML) is a subfield of AI that involves
to a pre-trained model to enable it to understand the using algorithms and statistical models to enable
problem or task at hand and provide telecom-relevant systems to improve their performance with experience
responses; requires a medium amount of data to be without being explicitly programmed. A form of non-
trained. generative AI.

Data Lake/Warehouse Model-As-A-Service (MaaS)

A centralized repository allowing the storage and Cloud-based end-to-end solutions providing
management of vast amounts of structured and companies with a way to deploy and maintain machine
unstructured data. learning and generative AI models in a production
environment.
Fine-Tuning
A process of updating and/or adding parameters to a Parameters
pre-trained model using a new dataset, which results Weights and connections of a generative AI model
in the model to learn and adapt to nuances; requires that determine the model’s behavior, capabilities, and
a large amount of labeled data to be trained, though ability to generate new data or content.
smaller than the data set the model was initially trained
on. Prompt Engineering
A process of manually feeding the model prompts to
Generative AI guide model output; requires a small amount of data to
Artificial intelligence technology that can create be trained.
original content (e.g., text, image, audio, video) based
on a prompt. Generative AI differs from non-generative Rule-Based AI
AI because it can create new content rather than for AI that follows a set of pre-defined rules to make
example, follow a predefined set of rules or make a decisions or perform specific tasks; form of non-
prediction or classification.. generative AI.

Generative AI Application Structured Data

User experience platform in which the user interacts Tabular data (e.g., tables, databases, or spreadsheets)
with the generative AI tool to produce output from the that can be used to train some machine learning
underlying model. models effectively.

Generative AI Foundation Models Training

Artificial intelligence models trained on large datasets The process in which model parameters are learned
that allow AI models to learn the underlying patterns during which the model adjusts its parameters to
and structures of data and generate new content that is minimize the difference between its predictions and
both realistic and contextual. the actual target data/output.

Inference Unstructured Data

The operational phase of generative AI model where Data that lacks a consistent format or structure (e.g.,
it applies the knowledge gained during training to text, images, and audio files) and typically require more
generate outputs that are usable, contextually relevant, advanced techniques to extract insights and analytics.
and aligned to patterns used.

Large Language Models (LLMs)

A class of foundation models built using deep learning
techniques and trained on massive amounts of data
to understand, generate, and manipulate human
language.

46
 10 l Appendix

Definitions

Business Domains

1. Marketing: Responsible for creating awareness of a company’s products or services and

generating leads.
2. Customer Service: Responsible for providing support to customers before, during, and
after they make a purchase.
3. Network Planning: Responsible for designing and implementing the network
infrastructure.
4. Network Operations: Responsible for managing & maintaining the day-to-day operations
of the network infrastructure.
5. IT Operations: Responsible for managing the organization’s IT infrastructure, including
servers, storage, and networks.
6. IT Software Development: Responsible for developing and maintaining the
organization’s software applications.

Use Cases
1. Marketing
• Content Generation (Text, Image, Video): Creating valuable, relevant, and
consistent marketing content for various channels (e.g., ads, email, social media, web
pages, etc.); includes copy generation and web page content generation.
• Search Term Generation: Generating the most optimal relevant keywords and
phrases for search engine optimization (SEO).
• Personalization: Generating tailored marketing messages and content based on
customer information (e.g., demographic, history, behavior).

Product
• Insight Generation: Generating new customer insights to inform product
enhancements and service improvements based on customer information.
• New Feature Ideation: Generating the most optimal relevant keywords and phrases
for search engine optimization (SEO).

2. Customer Service
• Customer Chatbot: Interpreting and responding to customer queries and requests in
a human-like conversational manner.
• Employee Coach: Uses data on customer interactions to provide coaching and
performance improvement guidance to customer service representatives.
• Guided Employee Assistance for Knowledge Management and Troubleshooting:
Retrieve and respond to employee queries using knowledge base of articles, FAQs,
customer case history and other documentation for issue resolution.

47
 • Contact Center Documentation: Generate transcription of calls, summarize
customer interactions and suggest follow-up actions.

3. Network Planning
• Generative Route/Network Design and Network Configuration: Generate a variety
of possible network designs optimized for key parameters (e.g., topography, costs,
power, etc.).
• Network Image Generation: Generate designs such as CAD of physical aspects of a
network project scheme including but not limited to maps, blueprints, etc.
• Synthetic Data Generation for Security Testing: Artificial data that mimics real-
world data, which is then used in testing security systems and identify vulnerabilities
without needing to use real customer information.

4. Network Operations
• Guided Employee Assistance for Installation, Troubleshooting, and Maintenance:
Retrieve and respond to network service engineers with information to aid in
installation, troubleshooting, and maintenance network devices and infrastructure like
routers, switches, and firewalls.

5. IT Operations
• Guided Employee Assistance for Installation, Troubleshooting, and Maintenance:
Generate responses or summarized information when employees install, troubleshoot,
and maintain software and hardware systems within an organization, specifically
those encompassing end-user devices and applications like computers, tablets, and
CRM software.
• Employee Self-Service Desk: Generative chatbot where internal employees can
self-service IT issues

6. IT Software Development
• Automated Code Generation, Debugging, Testing: The use of automated tools to
create code, find and fix errors in code, and test code.
• Automated Code Documentation: The use of automated tools to generate
documentation for code that is similar to existing documentation often using natural
language processing or non-generative AI.

Process Terminology
• Evaluation: The process by which an organization assesses a use case for technical
performance and capabilities to decide whether to implement or not
• Implementation: The process by which an organization deploys and integrates the
use case into the day-to-day operations of the company
• Adoption: The process of the evaluation and implementation of a use case at an
organization; in this study, we define current adoption as if the organization has
fully implemented/implementing the generative AI solution while interest is if the
organization has implemented OR has a high likelihood of implementing the solution

48
Data Capabilities
• Centralized Data Management: Degree to which data is collected and stored in a
centralized manner, avoiding data siloes to ensure data is easily accessible and usage
across the organization (e.g., data lake, data warehouse, etc.)
• Data Quality and Integration: Degree to which organization has implemented tools
and practices for master data management to maintain accurate data as well as
integration capabilities to enable data flow between various systems and multiple
vector bases.
• Scalability of Data Infrastructure: Degree to which data platform is designed to
handle large volumes of structured and unstructured data.
• Data Analytics: Degree to which the organization uses advanced analytics, AI/ ML to
gain insights from data, improve business operations and drive decision making.
• Data Infrastructure Modernization: Degree to which organization data resides on
virtual private and/or public cloud (includes isolated instances, etc.)
• Data Governance: Establishment of practices for data documentation, collecting,
storing, using and deleting data AND clearly defined data ownership responsibilities.
• MLOps Practices: Degree to which practices and processes are established for
building, deploying, monitoring and maintaining machine learning models.
• In-House Technical Expertise: In-house talent and technical expertise to maintain
data infrastructure and architecture for existing systems and emerging technology
solutions (e.g., data scientists, scrum masters, etc.)

Adoption Methods
• Off-The-Shelf: Model that is pre-trained on a large dataset and available for purchase
or free use.
• Prompt Engineering: A process of manually feeding the model prompts to guide
model output; requires a small amount of data to be trained.
• Context Learning: A process of using internal data to provide prompts to pre-trained
model to enable it to understand the problem or task at hand and provide telecom
relevant responses; requires a medium amount of data to be trained.
• Fine-Tuning: A process of updating and/or adding parameters to a pre-trained model
using a new dataset; requires a large amount of data to be trained.
• Build In-house: Model that is developed and trained on a custom dataset and is
owned and operated internally.

Data Sources
• First Party Data: Data the company collects directly from customers, website visitors,
or social media followers.
• Third Party Data: Data the company obtains from external sources or organizations
that are often specialized in collecting and aggregating data from diverse sources.

Use Case Contexts

• External Use Case: Interactions between an internal system and 3rd party external
factors such as customers & stakeholders outside the business.
• Internal Use Case: Interactions between a system and its internal actors like a
business’s own processes, employees, and data.

49
 10 l Appendix

End Notes

1. Kusterer Ziser, Kelsey, “Verizon’s Debika Bhattacharya on the endless possibilities for generative
AI,” Light Reading, May 22, 2023
2. Morris, Iain, “Google and Microsoft loom large in Vodafone’s gen AI plans,” Light Reading, August
4, 2023.
3. Park, Kate, “AI Startup Anthropic raises $100M from Korean telco giant SK Telecom,” TechCrunch,
August 14, 2023
4. Bilan, Maryna, “Improving CX with Generative AI Chatbot in Telecom: Success Stories and
Potential Use Cases,” Master of Code, September 22, 2023.
5. “Guide: Generative AI,” Dealroom.co, 2023
6. Dave, Paresh, “ChatGPT Is Cutting Non-English Languages Out of the AI Revolution,” Wired, May
31, 2023
7. Sivasubramanian, Swami, “Announcing New Tools for Building with Generative AI on AWS,” AWS
Amazon, April 13, 2023
8. “AI, Sport, and Cyberdefense: Orange at Viva Technology 2023,” Orange Press Release, June 13,
2023
9. “SKT Introduces New AI Service ‘A’”, SK Telecom Press Release, May 16, 2022
10. https://aws.amazon.com/bedrock/

50
 10 l Appendix

About Altman Solon

Altman Solon is the largest global strategy consulting firm exclusively working in the
Telecommunications, Media, and Technology industries. With thirteen offices across the Americas,
Europe, and Asia Pacific, our team of consultants possess a diverse range of expertise in strategy,
commercial, operational, technical, and capital markets. We work with market leaders, challenger
brands, and investors worldwide to help them identify, develop, and implement effective company
strategies, new market entry approaches, digital transformation, and global M&A. We also provide
comprehensive market understanding to conduct due diligence and enable high-stakes, confident
decision making for investors.

Altman Solon combines this new Survey with proprietary insights from its recent study, Putting
Generative AI to Work, from Spring 2023, which analyzed emerging enterprise use cases by
examining generative AI companies and capabilities. Our team surveyed 292 senior business leaders
and spoke to 21 industry experts to understand the adoption of generative AI tools for specific high-
impact enterprise use cases. The surveyed decision-makers’ roles covered a range of strategic and
technical functions, and the findings highlight adoption and adoption reasons for generative AI. All
survey respondents were from the United States of varying business sizes. Altman Solon focused on
four areas with strong momentum: software development, marketing, customer service, and product
design, and found the highest adoption in software development and marketing and emerging
adoption in customer service and product design.

51
 10 l Appendix

About Amazon Web Services (AWS)

Amazon Web Services (AWS) is the world’s most comprehensive and broadly adopted cloud, offering
over 200 fully featured services from data centers globally. Millions of customers are using AWS to
lower costs, become more agile, and innovate faster.

AWS for Telecom is raising the bar on what’s possible for CSPs. Harnessing the robust, mature, proven,
and smart platform, to deliver innovative use cases that are digitizing industries, transforming telcos,
and reimagining the consumer experience. With the power of the AWS Global Cloud Infrastructure,
that’s architected to be flexible and secure, AWS for Telecom help CSPs take their business to the next
level. Through simplifying operations, cloudifying networks, monetizing assets, and innovating at scale
using AIML and generative AI services.

Visit our website to learn more https://aws.amazon.com/telecom.

52
 10 l Appendix

Acknowledgements

The research underpinning this report was led by Swope Fleming, Partner – Boston; Priya Mehra,
Director – New York; and Elisabeth Sum, Senior Consultant – San Francisco. The Altman Solon project
team also included Danny Malloy, Analyst - New York.

We are grateful to the following AWS advisers who provided their knowledge and added new insights:
Ishwar Parulkar – AWS Telecom CTO; Ross McWalter – Head of AWS Telecom Business Applications;
Dafna Yanay – Head of Telecom Marketing; Allison Bishop – Telecom PR Lead; and Jeff Dillon –
Telecom AR Lead.

This research benefited immensely from the expertise and perspective of many Altman Solon
colleagues, including:

• Mary Yarborough, Senior Advisor

• Daniel Torras, Partner – New York
• Carli Esser, Partner – London
• David Borstein, Partner – Sydney
• Ryan Dean, Partner – Boston
• Daniel Mazor, Partner – Boston
• Sujit Jha, Partner – New York
• Philip Morse, Partner – London
• Don Sutherland, Director – San Francisco
• Riccardo Consani, Director – Milan
• Daniel Granoff, Director – Boston
• Dave Baldwin, Principal – Singapore
• Oussama Fadil, Senior Consultant – Boston

We also thank our Altman Solon marketing and communications colleagues: Kate Fagan, Madeline
Taylor, and Sebastian Denkiewicz.

53