MIS Quarterly Executive

Volume 20 Issue 1 Article 4

March 2021

Deciding Whether and How to Deploy Chatbots

Ryan M. Schuetzler

G. Mark Grimes

Justin Scott Giboney

Holly K. Rosser

Follow this and additional works at: https://aisel.aisnet.org/misqe

Recommended Citation
Schuetzler, Ryan M.; Grimes, G. Mark; Giboney, Justin Scott; and Rosser, Holly K. (2021) "Deciding Whether
and How to Deploy Chatbots," MIS Quarterly Executive: Vol. 20 : Iss. 1 , Article 4.
Available at: https://aisel.aisnet.org/misqe/vol20/iss1/4

This material is brought to you by the AIS Journals at AIS Electronic Library (AISeL). It has been accepted for
inclusion in MIS Quarterly Executive by an authorized administrator of AIS Electronic Library (AISeL). For more
information, please contact elibrary@aisnet.org.
 Deciding Whether and How to Deploy
Chatbots
In the past five years, chatbots have become more widely available and more popular
as a customer service solution. However, chatbots present unique challenges because
of their social nature. This article explores the strengths and weaknesses of chatbots
and identifies the fundamental questions organizations need to answer to guide their
development of chatbots. Based on years of experiments studying what makes a good
chatbot interaction, we provide recommended best practices for deciding whether and
how to deploy chatbots successfully.1 ,2

Ryan M. Schuetzler G. Mark Grimes

Brigham Young University (U.S.) University of Houston (U.S.)

Justin Scott Giboney Holly K. Rosser

Brigham Young University (U.S.) University of Nebraska Omaha (U.S.)

Deploying Chatbots Brings Many Challenges 2

“What I had not realized is that extremely short exposures to a relatively simple computer
program could induce powerful delusional thinking in quite normal people.” Joseph
Weizenbaum (1976), on the creation of the ELIZA chatbot3

For decades, text-based chatbots were mostly limited in their interactions to research
systems and Internet Relay Chats.4 However, recent advances in technology have enabled
businesses to interact with customers in ways never before possible through the use of
chatbots. The technology that has been a focus of science fiction writing for centuries and
part of legend for millennia5 has now become reality as computers that engage in humanlike
interactions are becoming increasingly prevalent in day-to-day interactions. The rise of
messenger services like Slack and Facebook Messenger has allowed companies to create
chatbots to interact with millions of users directly through popular messaging platforms for the
first time. Today, companies report that their use of chatbots is growing quickly. The chatbot
market is expected to increase from $2.6 billion in 2019 to $9.4 billion in 2024.6 According to
1 Hind Benbya is the accepting senior editor for this article.
2 The authors thank Hind Benbya and the reviewers for their suggestions and guidance through the review process.
3 One of the oldest chatbots is ELIZA, developed by Joseph Weizenbaum in the 1960s.
4 Internet Relay Chat is an application layer protocol that facilitates communication in the form of text. The chat process works on
a client/server networking model.
5 Talking automata were mentioned in the ancient Greek myths of Daedalus and Hephaestus, and in the ancient Chinese text of Lie
Zi from the 3rd century BC.
6 Nguyen, M.-H. “The Latest Market Research, Trends, and Landscape in the Growing AI Chatbot Industry,” Business Insider,
January 23, 2020, available at https://www.businessinsider.com/chatbot-market-stats-trends.

DOI: 10.17705/2msqe.00039
March 2021 (20:1) MIS Quarterly Executive 1
Deciding Whether and How to Deploy Chatbots

the 2019 Gartner CIO survey, 70% of white-collar failures of 2016.11 Both practice and research
workers will interact daily with some type of have shown that consumer-facing AI applications
conversational platform by 2022. are often met with suspicion and resistance.
Despite the rapid growth, there are many However, when chatbots are well-implemented
questions that are as of yet unanswered regarding and well-received, customers are more likely to
how companies can most effectively evaluate return to the website and rate the organization
and deploy chatbots. One intuitive approach favorably when compared to others.12 To gain
for assessing a chatbot’s effectiveness is to benefits from deploying chatbots, organizations
assess how well it is able to mimic real human must understand both the technological
communication, which is the goal of the “Turing components of chatbots and the unique
Test.”7 Each year, Loebner prize8 competitors challenges and expectations that come with a
get closer to passing the Turing Test by using conversational interaction.
pattern matching and machine learning to teach As a team of researchers, we have been
computers how to understand and respond to investigating chatbot systems for over seven
natural language. The aim of many conversational years. Like many others, we originally started
AI developers is to create a chatbot that can with the goal of creating better, more realistic
converse in a way indistinguishable from chatbot conversations. Over the course of many
conversing with a human. experiments, we have come to realize that the
Though significant progress has been made, intuitive view—that more humanlike chatbots
technology has not yet reached the long-sought- produce better results—is not true in all
after goal of parity with humans. Moreover, applications. In this article, we describe some of
some research on human-chatbot interactions the unique circumstances we have encountered
has shown that more humanlike conversation that will help guide decisions about when and
should not always be the goal. For example, how chatbots should be deployed. In addition
when the intent of the interaction is to elicit to discussing our own studies, we provide an
disclosure of sensitive information—as in a overview of research on chatbots with the goal of
medical interview—people tend to disclose less understanding when chatbots are best deployed
information when chatbots are more humanlike.9 and how their design can be matched to the
In addition to understanding what desired outcomes. We also discuss the benefits
applications are best suited to chatbots, there and pitfalls of creating more humanlike chatbots.
are both technical and social challenges to the
successful deployment of a chatbot. In 2017, A Decision Framework for
Facebook was forced to scale back chatbot Chatbot Deployment
and AI use due to a 70% failure rate with
users.10 Microsoft’s notorious Tay chatbot was Based on our experiences and research
manipulated by Twitter trolls to share racist program, this article provides guidance for
content, putting it on MIT’s list of top technology understanding how chatbots can be used most
effectively in an organization. The guidance is
structured around three central questions that
7 The Turing Test assesses the capability of an AI system by asking
human judges to distinguish between a human and a machine on the organizations need to answer: whether or not
other end of a conversation. The test was first proposed in Turing, A. to build a chatbot, what the technology options
M. “Computing Machinery and Intelligence,” Mind (59:236), 1950, for building a chatbot are, and how the chatbot
pp. 433-460.
8 The Loebner Prize is the largest and most widely recognized Tur- should interact with users. Figure 1 lists the
ing Test competition. questions and topics that provide a framework for
9 In a study of disclosure of sensitive information, people were
more likely to disclose to a less humanlike chatbot. For more infor- 11 Boyle, A. “Microsoft’s Chatbot Gone Bad, Tay, makes MIT’s
mation, see Schuetzler, R. M., Giboney, J. S., Grimes, G. M. and annual list of biggest technology fails,” GeekWire, December 27,
Nunamaker, J. F. “The Influence of Conversational Agent Embodi- 2016, available at https://www.geekwire.com/2016/microsoft-chat-
ment and Conversational Relevance on Socially Desirable Respond- bot-tay-mit-technology-fails/.
ing,” Decision Support Systems (114), October 2018, pp. 94-102, 12 Araujo, T. “Living up to the chatbot hype: The influence of
available at https://doi.org/10.1016/j.dss.2018.08.011. anthropomorphic design cues and communicative agency framing
10 Orlowski, A. “Facebook Scales Back AI Flagship After Chatbots on conversational agent and company perceptions,” Computers in
Hit 70% f-AI-lure Rate,” The Register, February 22, 2017, available Human Behavior (85), August 2018, pp. 183-189, available at https://
at https://www.theregister.com/2017/02/22/facebook_ai_fail/. doi.org/10.1016/j.chb.2018.03.051.

2 MIS Quarterly Executive | March 2021 (20:1) misqe.org | © 2021 University of Minnesota
 Deciding Whether and How to Deploy Chatbots

Figure 1: Questions Guiding Decisions About Chatbot Development

deciding whether and how to deploy a chatbot. Defining the Business Problem
In the sections that follow, we describe the steps Before deciding whether to create a chatbot,
organizations can take to answer these questions. an organization must first define the business
problem to be solved. Chatbots provide a text-
Deciding Whether to Build a based natural language interface that gives
Chatbot users a way to interact with a system. As a new
technology, developers and management might
First, an organization must consider whether want to build a chatbot just because of the
a chatbot development project is indeed the right novelty of the technology. However, the novelty
solution. As with many new technologies, there of the interface alone will not create a positive
is a lot of hype about chatbots because of their experience for users. For example, if the problem
novelty and potential for cost savings. However, is helping users find information, robust search
just because chatbot technology exists does not capabilities and a good index could provide the
make it the appropriate solution for all situations. same benefits without requiring the development
The BBC in-article chatbots case (see text box on of a chatbot. In short, instead of saying “We
next page) provides an example of a company have decided we want to build a chatbot,” the
that had clear goals and a clear purpose for using organization should spend time clarifying what
chatbots.

March 2021 (20:1) MIS Quarterly Executive 3

Deciding Whether and How to Deploy Chatbots

Case: BBC In-Article Chatbots

In 2017, the BBC began a pilot program to add in-article chatbots to its website. Certain articles were
selected to include a built-in chatbot that readers could interact with to get more detailed background
information on the article. Through a custom interface, reporters developed questions that readers
could click on and the corresponding answers were given by the chatbot. The chatbot provided
background or additional information about the topic of the article. The interface was selected
to appeal to a younger audience or audiences that may not have all of the background information
they might need to understand the article. Grant Heinrich, the developer of the BBC chatbot project,
said that “... if you have a really complicated, long piece and you want people to grasp the basics very
quickly, [chatbots] are very good.” 13

purpose the chatbot will serve and why a chatbot For internal applications, chatbots can save
is the right solution.13 time by automating routine interactions. For
example, a tech support helpdesk could use
Considering the Task-Technology Fit a chatbot to respond to user issues and walk
Once the organization has defined the users through some basic troubleshooting
business problem, it must then consider the scenarios, such as password resets or power
task-technology fit to help it decide if a chatbot is cycling computing devices. Chatbots could
the appropriate solution for the problem. Some also be developed to respond to common HR
business problems are well-suited for chatbots interactions like requesting payroll information
and others are not. A well-designed chatbot or vacation time. Automating simple interactions
can save time, guide users through simple steps with a chatbot can provide a personal touch while
to accomplish a task, gather information and saving human time for complicated or nonroutine
interact in a much more personal way than other tasks.
options such as web forms, search engines or Chatbots can also be used to provide
apps. information for teams located in a shared virtual
A good chatbot can create a social connection space. Examples include Slack’s popular Lunch
with its users, including current and potential Train chatbot,16 which helps coordinate team
customers, in a way that more traditional systems lunches, and the Conclude conversation app that
cannot.14 The conversational nature of chatbot provides a polling platform so employees can
technology can allow a more personal tone for share ideas and gather feedback.17 Interacting
customer-facing applications.15 Because chatbots with a chatbot in a shared chat enables teams
can be connected through Facebook, Twitter, SMS to connect, share and respond together. If an
and other platforms, they can reach customers organization is using Slack or a similar messaging
where they are without requiring them to install system, it should consider how chatbots can be
a new application. When that social connection used to help improve the user experience.
improves access to an organization’s application, There is no doubt that a poorly designed
a chatbot may be the right solution. chatbot will frustrate users, and may even result
in abusive language and negative feelings toward
13 Bilton, R. “With In-Article Chat Bots, BBC is Experimenting
with New Ways to Introduce Readers to Complex Topics,” Nie- the company. However, even a well-designed
manLab, February 23, 2018, available at https://www.niemanlab. chatbot can lead to frustration. Because of the
org/2018/02/with-in-article-chat-bots-bbc-is-experimenting-with- natural language interface, users tend to assume
new-ways-to-introduce-readers-to-complex-topics.
14 Schuetzler, R. M., Grimes, G. M. and Giboney, J. S. “The Im- the chatbot has a high level of skill, which may
pact of Chatbot Conversational Skill on Engagement and Perceived
Humanness,” Journal of Management Information Systems (37:3),
November 2020, pp. 875-900, available at https://doi.org/10.1080/07
421222.2020.1790204.
15 Ibars, O. “Breathing Life to Your chatbot: Branding, Personality, 16 For information on Lunch Train, see https://slack.com/apps/
and Tone,” Chatbots Magazine, April 3, 2018, available at https:// A1BES823B-lunch-train.
chatbotsmagazine.com/breathing-life-to-your-chatbot-branding- 17 For information on Conclude, see https://www.concludebot.
personality-and-tone-3170663d0fd1. com/.

4 MIS Quarterly Executive | March 2021 (20:1) misqe.org | © 2021 University of Minnesota
 Deciding Whether and How to Deploy Chatbots

not match its actual capability. If a chatbot does customizations can be made to fit a specific use
not live up to expectations, users get frustrated.18 case. The organization’s capabilities and goals will
An alternative to a chatbot is to use web forms, determine the appropriate answer to the “build
which allow users to fill in information and or buy” decision.
can provide additional information. Moreover, Users’ perceptions of an organization’s
web forms do not have to deal with natural integrity and reputation, and the decision-making
language understanding issues inherent in abilities of the system, may be influenced by their
chatbot technology. For example, when asking for experience with the chatbot. Organizations must
something simple such as a user’s name, a web therefore consider their capabilities to develop
form can provide a field for the name. However, a working, high-quality chatbot that can meet
a natural language chatbot first must learn to the needs of the business. Since users will rarely
recognize what a name is and how it might be be aware of the degree of autonomy the chatbot
written in text. If a chatbot asks “What’s your possesses, the ability of an organization to
name?” the user might reply with just with their recognize and take responsibility for a chatbot’s
name, “Anna,” or could give a longer response: decision-making will affect the perceived integrity
“My name is Anna.” A chatbot also needs to and reputation of the business.20 A poorly
recognize the name if it is uncapitalized or if it is implemented chatbot may cause reputational
a name with other meanings (“North” or “Apple,” damage to the business.
for example), or an uncommon name such as “X
Æ A-12” that may not be present in the training Deciding Which Technology to
data used to match names. For cases where the Use
information to be gathered is highly structured
and clearly defined, a web form might be a better Once an organization has decided that
choice than a chatbot. a chatbot is an appropriate solution for the
identified business problem, it must next
Considering the Expertise of the determine which kind of technology to use to
Organization’s Developers build the chatbot. There are three main types
Even if a chatbot is an appropriate solution for of chatbot technology: keyword matching,
the identified business problem, the organization conversational and machine learning, with the
must then consider if it has the development latter being subdivided into supervised and
expertise needed to properly handle a chatbot unsupervised learning.21 As with the decision
project. Developing a chatbot requires skills that on whether to build a chatbot or not, the
many software developers do not have, including selection of technology depends on the needs
natural language processing and natural language of the organization—there is not a one-size-
understanding. Several tools, such as DialogFlow fits-all “best” type of chatbot. Below, we discuss
and IBM’s Watson, aim to make the development some of the primary use cases for each chatbot
process easier and can smooth the transition technology. A summary comparison of the
from traditional software development to chatbot different technologies is shown in Table 1, and
development. each type is described in detail below.
Rather than developing a chatbot in-house, an
organization could choose to purchase an existing Keyword Matching Chatbots
solution. These solutions enable a chatbot to Some of the earliest chatbots used simple
be constructed from a template and designed keyword matching to accomplish their “natural”
to include domain-specific knowledge. IPsoft’s
20 Alaieri, F. and Vellino, A. “A Decision Making Model for Ethi-
Amelia19 is one such example, where the internals
cal (Ro)bots,” [Proceedings of] 2017 IEEE International Symposium
of the chatbot are already designed and small on Robotics and Intelligent Sensors (IRIS), October 2017, pp. 203-
207, available at https://doi.org/10.1109/IRIS.2017.8250122.
18 López, G., Quesada, L. and Guerrero, L. A. “Alexa vs. Siri 21 For an in-depth comparison of different types of chatbots, see
vs. Cortana vs. Google Assistant: A Comparison of Speech-Based Diederich, S., Brendel, A. B. and Kolbe, L. M. “Towards a Taxonomy
Natural User Interfaces,” Advances in Intelligent Systems and of Platforms for Conversational Agent Design,” [Proceedings of]
Computing (592), June 2017, pp. 241-250, available at https://doi. Human Practice. Digital Ecologies. Our Future. 14. Internationale
org/10.1007/978-3-319-60366-7_23. Tagung Wirtschaftsinformatik (WI 2019), February 2019, pp. 1100-
19 For information on Amelia, see https://www.ipsoft.com/amelia/. 1114.

March 2021 (20:1) MIS Quarterly Executive 5

Deciding Whether and How to Deploy Chatbots

Table 1: Comparison of Chatbot Technologies

Keyword Supervised Machine Unsupervised Machine
Technology Type Conversational
Matching Learning Learning
Tools Hubot, Errbot ChatScript, DialogFlow, Watson, DialogFlow’s Knowledge,
Pandorabots, Azure Language LoopAI
Cleverbot Understanding
Application type Command- Conversation- Intent-based interaction Question answering and
based based search
interaction interaction
Example IRC bots, Cleverbot, Starbucks’ My Barista, Loop Q
applications ChatOps Woebot Siri

understanding. For example, early Internet Relay amounts of data generated by users to attempt to
Chat chatbots used basic keyword matching to respond to every possible user message. Because
recognize that a command has been entered these products depend on users to create the
and to parse the instructions. Chatbots using dialog, they are vulnerable to abuse and can be
text matching technology can mimic a small manipulated by users to engage in conversation
level of understanding without a great deal of in ways not anticipated by the creators.23 Other
technical sophistication. Today, these primitive approaches for creating chatterbots include the
implementations are mainly used to create what use of languages like ChatScript24 and AIML,25 or
essentially amounts to a chat-based command- using a hosted platform like Pandorabots.26
line interface. Keyword matching chatbots are Conversational chatbots work best when
best used for internal applications with users the primary goal is to engage with users and
who will have the time or interest to become carry on a conversation without a specific end in
familiar with the commands and instructions the mind. Because they provide a social connection
chatbot can respond to. Example applications for and allow an organization to reach customers in
keyword matching chatbots include: new ways, this type of chatbot can be used for
• Carrying out actions in Internet Relay Chat customer contact and brand promotion.
systems such as ejecting users from a chat Examples of conversational chatbots are:
room, or telling jokes • Mitsuku,27 the five-time winner of the
• Creating a quiz bot where users respond Loebner Prize competition for best
by selecting an option from a list chatbot
• Customizing a chatbot platform (e.g., • Woebot,28 the robot therapist developed
GitHub’s Hubot) to perform actions like by Stanford researchers
checking on build status, reporting uptime • Interviewing chatbots, such as those used
and translating text for automated job interviews.29
• Appointment reminder bots.
As well as Hubot, commercially available tools 23 This vulnerability led to the downfall of Microsoft’s Tay chatbot,
as described in Wolf, M. J., Miller, K. and Grodzinsky, F. S. “Why
for building keyword matching chatbots include
We Should Have Seen That Coming: Comments on Microsoft’s Tay
Errbot and basic programmed text matching ‘Experiment,’ and Wider Implications,” ACM SIGCAS Computers
tools. and Society (47:3), September 2017, pp. 54-64.
24 For information on ChatScript, see https://github.com/
ChatScript/ChatScript.
Conversational Chatbots 25 For information on AIML, see http://callmom.pandorabots.com/
Conversational chatbots—also known as static/reference/.
chatterbots—are designed specifically to use 26 For information on Pandorabot, see https://home.pandorabots.
natural language, sometimes with no other com/home.html.
27 For information on Mitsuku, see https://www.pandorabots.com/
goal than to engage users in conversation. mitsuku/.
Some products, such as Cleverbot,22 use large 28 For information on Woebot, see https://woebot.io/.
29 Gecko, Impress and others are currently working to automate
22 For information on Cleverbot, see https://www.cleverbot.com/. parts of the recruiting process with AI, including chatbots.

6 MIS Quarterly Executive | March 2021 (20:1) misqe.org | © 2021 University of Minnesota
 Deciding Whether and How to Deploy Chatbots

Case: Poncho the Weather Cat

The weather chatbot Poncho was created in 2016 to join the first wave of Facebook Messenger
chatbots. Unfortunately, the design left something to be desired. Due to inadequate machine learning
training data, Poncho was unable to respond to many basic questions, such as “What is the seven-
day forecast for Brooklyn” and “What's the weather like this weekend?” It was also unable to respond
appropriately to follow-up questions, leaving users frustrated with their experience. Poncho CEO Sam
Mandel told reporters, “... tolerance for a mediocre [chat]bot is much less than for a mediocre app.”31

Tools for creating conversational chatbots If more information is required, supervised

include ChatScript, Cleverbot and Pandorabots. learning chatbots can use conversation to gather
the information needed to fully understand
Machine Learning Chatbots the intent of a user’s message. For example, to
Machine learning chatbots are growing in both complete a travel authorization, the required
adoption and capability. Competing offerings information would be the start and end dates,
from Google, Amazon, IBM and many others are destination and the purpose of the travel.
constantly improving the abilities of machine Using follow-up questions, the chatbot can ask
learning chatbots to respond appropriately and for additional data if needed. But it could also
manage more complex dialogs. These types recognize that all information has been provided
of chatbots are the underlying technology if the user says, “I need to authorize travel to
behind digital personal assistants such as Detroit from 3/15 to 3/19 to visit our office there
Alexa and Siri. Their primary goal is to identify and provide employee training.”31
what a user wants to do and take action or Examples of applications of machine learning
provide information in response. In contrast chatbots are:
to conversational chatbots, the interactions
• Personal digital assistants like Alexa or Siri
are typically short and clearly directed toward
• Flight scheduling chatbots
accomplishing a task.
• Starbucks’ My Barista
Supervised Machine Learning Chatbots.
Most machine learning chatbots use supervised Tools for creating supervised machine
machine learning30 to accomplish their task. Using learning chatbots include IBM’s Watson, Google’s
dozens, hundreds or even thousands of sample DialogFlow and Microsoft’s Azure Language
messages, the machine learning algorithm can be Understanding. The Poncho the Weather Cat
trained to distinguish different intents and allow chatbot (see text box above) demonstrates
it to properly understand the intent of a new the importance of adequate training data in
message. Supervised machine learning chatbots the creation of a supervised machine learning
are best used when a user has a specific purpose chatbot.
or intent. The algorithm parses the user messages Unsupervised Machine Learning Chatbots.
to identify the intent. For example, intents for an The latest chatbot technologies currently being
HR chatbot could be to find the date of the next developed use unsupervised machine learning.
paycheck, input a vacation request or complete Rather than using sample messages to train
travel authorization. Sample training messages the chatbot, unsupervised learning uses source
for the paycheck date intent could be “When is my documents like a company’s FAQ pages or other
next paycheck going to come,” “When’s payday” or documentation and attempts to automatically
“How long until I get paid?” With enough training recognize answers to user questions from the
data, the chatbot can accurately determine the information in those documents. Unsupervised
intent of the user’s question and provide the machine learning is not new, but its use in
preprogrammed response. chatbots is a relatively recent development.
Examples of tools for building unsupervised
30 Supervised machine learning uses training data where the intent
is known and is manually classified to train a machine learning 31 Orf, D. “Facebook Chatbots Are Frustrating and Useless,”
model. That model then attempts to classify the intent of new mes- Gizmodo, April 14, 2016, available at https://gizmodo.com/facebook-
sages based on patterns extracted from the training data. messenger-chatbots-are-more-frustrating-than-h-1770732045.

March 2021 (20:1) MIS Quarterly Executive 7

Deciding Whether and How to Deploy Chatbots

learning chatbots include DialogFlow’s the expectations you want users to have about the
Knowledge plugin and LoopAI’s Loop Q. chatbot and how those expectations influence the
Another technology-related decision is the interaction.
platform that customers or employees will use to People tend to anthropomorphize and apply
interact with the chatbot. Different chat platforms social context to technology they are interacting
have different capabilities, some of which may with. In fact, the Computers Are Social Actors
limit development options. For example, the (CASA) research paradigm has repeatedly shown
amount of multimedia content possible varies that people anthropomorphize even the most
by platform. Facebook Messenger allows chatbot non-humanlike computer systems. The opposite
responses to include an image or even a carousel holds true as well: there is a strong tendency for
of images, but that is not as simple for a text people to try to prove that a chatbot isn’t human.
message (SMS) chatbot. Some platforms allow The tension between the two tendencies is
the designer to suggest responses that users can common enough for researchers to have given it a
click to guide the conversation, while others do name: the oscillation effect.
not. Organizations must therefore consider if the The oscillation effect suggests that as users
limitations of the platform match the needs of the interact with a chatbot, they alternate between
use case. unthinkingly treating it as human and actively
probing to find its limits.32 This pivoting between
Deciding How Humanlike the the two effects often has negative consequences
Chatbot Should Be for the chatbot, especially when it is positioned
as more human than machine. In this type of
Organizations need to decide how they want situation, some users behave badly and are prone
users to perceive their chatbots. In some cases, to becoming abusive or even using sexually
they may want users to perceive and interact explicit language in their interactions.33Chatbot
with a chatbot just as they would with any other designers must therefore consider how they
computer system. This is frequently the case should explicitly and implicitly shape user
for procedural tasks where keyword matching expectations. This is especially true for customer-
chatbots are most appropriate. In other cases, facing chatbots, because customers will perceive
organizations will want users to feel a social the chatbot as a representative of the company.
connection with the chatbot, just as they would One mechanism for explicitly shaping
with a human agent. Such a connection may be expectations is for the chatbot (or the interface
enhanced if users anthropomorphize the chatbot. around it) to clearly communicate that it is a
Identity, verbal and nonverbal cues designed chatbot—for example, by telling users they are
into a chatbot will determine how humanlike chatting with an automated agent. Offering cues
users perceive it to be. Before describing these such as including “bot” in the chatbot’s name
cues, it is first necessary to understand how the (e.g., “Hi! I’m CompanyBot. How can I help?”)
concepts of anthropomorphism affect the degree can also help users understand that they are in
of humanlike-ness that users ascribe to a chatbot. an automated conversation. However, explicitly
telling users they are interacting with a chatbot
Understanding How could result in undesired behavior such as verbal
Anthropomorphism Concepts Affect abuse and profanity, so designers may need to
Chatbot Design build in responses to these types of user actions.
Anthropomorphism is the attribution of The chatbot interface could also explicitly
human characteristics like agency, emotion deceive users into believing they are interacting
and personality to non-human things. In the
case of chatbots, anthropomorphism is when
32 Brahnam, S. “Building Character for Artificial Conversational
people believe the chatbot to be a human or Agents: Ethos, Ethics, Believability, and Credibility,” PsychNology
subconsciously expect humanlike behavior from Journal (7:1), January 2009, pp. 947.
the chatbot. Users can have these expectations 33 Hill, J., Randolph Ford, W. and Farreras, I. G. “Real Conversa-
tions with Artificial Intelligence: A Comparison Between Human–
even if they know for certain they are talking with Human Online Conversations and Human–Chatbot Conversations,”
a chatbot. Anthropomorphism is concerned with Computers in Human Behavior (49), August 2015, pp. 245-250,
available at https://doi.org/10.1016/j.chb.2015.02.026.

8 MIS Quarterly Executive | March 2021 (20:1) misqe.org | © 2021 University of Minnesota
 Deciding Whether and How to Deploy Chatbots

with a human. But doing this will result in alcohol consumption when interacting with the
frustration and anger when the chatbot makes a skilled chatbot and disclosed higher levels to the
mistake. In our own studies, users had negative unskilled chatbot (presumably the more honest
reactions when they were told they would response, since people tend to under-report
be interacting with a human but were in fact socially undesirable behaviors). We concluded
interacting with a chatbot. Misleading users that the reason for this difference was that a
will make them more sensitive to errors in the skilled chatbot evoked the same feeling that
chatbot’s responses, resulting in probing and people get when interacting with a person: their
abusive behavior at the first indication they are interviewer might judge their responses and
not conversing with a human. For these reasons, therefore think less of them.
we do not recommend this type of deception. In In another study we investigated how typing
our own experiments, we have found that a single behavior is affected by lying, to determine if we
“bad” response—where the chatbot responds in a could detect deception by observing changes in
way that betrays it is not a human—is enough to typing patterns.37 When users interacted with an
convince users they are interacting with a poorly unskilled chatbot, we observed a clear difference
designed chatbot. in response timing characteristics when lying
A third option is to leave the nature of the as opposed to telling the truth—a feature that
chatbot ambiguous: neither telling users that could be used to develop automated interviewing
it is an automated system nor lying about its systems that are capable of detecting deception
humanness. The design of a chatbot and its in responses. However, when users interacted
interface can implicitly affect the way people view with a conversationally skilled chatbot, deception
it as they interact with it. Over the course of an and truthful typing behavior were too similar
interaction, users will continually evaluate the to distinguish truth from deception. As a result,
chatbot. More humanlike behavior makes users deception detection is easier when users interact
more likely to anthropomorphize the chatbot and with an unskilled chatbot, because their behavior
treat it as human. We have seen both positive and changes more. Once again, the “better” chatbot
negative effects of such anthropomorphism. Our created a worse result.
primary point of comparison has been between These results, along with others from research
“skilled” chatbots (designed to provide natural, literature,38 support the idea that, at least in some
conversational responses) and “unskilled” contexts, a less humanlike experience is better.
ones (designed to provide repetitive, robotic Later, we describe in more depth the identity,
responses). The skilled chatbots in each study verbal and nonverbal cues that can be designed
resulted in better user engagement, a reflection into chatbots to make their behavior seem more
of how well users felt the chatbot connected with humanlike, and identify the trade-offs associated
them.34 with these design decisions.
However, better conversation and engagement It is also important to assess whether the
doesn’t always mean the chatbot is better for its chatbot design and expectations are appropriate
intended purpose. In one study, we designed for vulnerable sections of the population or
a chatbot to interview users about sensitive for socially sensitive settings.39 To avoid undue
health information such as their alcohol exploitation, special care must be taken with
consumption.35 Our goal was to understand children, the elderly, persons suffering from
if chatbots could be used to help doctors get certain mental illnesses and lonely or isolated
accurate answers from patients to questions that
might be embarrassing.36 Users reported less 37 Schuetzler, R. M., Grimes, G. M. and Giboney, J. S. “The Effect
of Conversational Agent Skill on User Behavior During Deception,”
34 Schuetzler, R. M., Grimes, G. M. and Giboney, J. S., op. cit., Computers in Human Behavior (97), August 2019, pp. 250-259,
November 2020. available at https://doi.org/10.1016/j.chb.2019.03.033.
35 Schuetzler, R. M., Giboney, J. S., Grimes, G. M. and Nunamak- 38 See, for example, Kang, S.-H. and Gratch, J. “Virtual Humans
er, J. F., op. cit., October 2018. Elicit Socially Anxious Interactants’ Verbal Self-Disclosure,” Com-
36 Studies have shown that 60-80% of Americans have lied to their puter Animation and Virtual Worlds (21), May 2010, pp. 473-482.
doctors about information relevant to their healthcare. See Vogel, L. 39 Borenstein, J. and Arkin, R. C. “Nudging for Good: Robots and
“Why Do Patients Often Lie to Their Doctors?,” Canadian Medical the Ethical Appropriateness of Nurturing Empathy and Charitable
Association Journal (191:4), January 2019, E115, available at https:// Behavior,” AI & SOCIETY (32:4), November 2017, pp. 499-507,
doi.org/10.1503/cmaj.109-5705. available at https://doi.org/10.1007/s00146-016-0684-1.

March 2021 (20:1) MIS Quarterly Executive 9

Deciding Whether and How to Deploy Chatbots

individuals. All of these groups may be more chatbot implies that users are interacting with a
prone to behavioral manipulation. human, even if it is not said explicitly. But to avoid
In summary, more humanlike behavior from implying that the chatbot is human, a cartoon or
a chatbot will encourage people to treat it as anonymous avatar could be used. Alternatively,
if it were human. Depending on the context, and depending on the chat platform, no avatar
that perception may help or hinder the chatbot could be used. Similarly, giving the chatbot a
in achieving its goals. If the organization’s human name implies that it is human, so some
goals for the chatbot are social connection and chatbots are given a bot-oriented name, such as
engagement, humanlike behavior will help. Slackbot or Woebot, or are not given a name at all.
If social interaction is not a core goal, a less The use of self-references (such as “I like …”) can
humanlike chatbot would be better. However, create a more humanlike conversation, even if the
even with a humanlike chatbot, the goal should chatbot is explicitly nonhuman. However, some
not be to convince users they are chatting with a have argued that even this level of simulation
person, but rather to make the conversation feel of human behavior is unethical because it is a
natural and engaging. form of deception, even if users know they are
interacting with an automated system.41
Selecting the Cues to Design into a Nonverbal Cues. The second set of cues are
Chatbot nonverbal and determine the way a chatbot
If an organization decides to deploy a responds to users. With spoken communication,
humanlike (i.e., anthropomorphic) chatbot agent, nonverbal signals relate to the way things are
there are three types of cues that designers can said, while verbal cues (see below) focus on the
use to suggest humanness: identity cues such as words. With text conversations, nonverbal signals
a human name and photorealistic human avatar; take a variety of different forms. Adding these
nonverbal cues such as artificial reading and cues makes the conversation feel more natural
typing delays, typing indicators and use of emojis; because it mimics human behavior. We discuss
and verbal cues where the chatbot attempts to three types of nonverbal cues that can be built
mimic human language patterns.40 Even without into a chatbot: reading and typing delays; typing
the goal of humanlike behavior, designers might indicators; and emoticons.
include these cues to help users feel comfortable Reading and typing delays are inherent
in the conversation. Nevertheless, the chatbot’s in human conversations—the other person
communication style and cues should match needs time to see the message, process it and
the expectations users have and the goals of compose a reply. Chatbots are generally capable
the chatbot. If the goal is to create a social of responding nearly instantly to user input, but
and humanlike chatbot, it must avoid outright immediate responses can feel jarring to users.
deception. The purpose of using humanlike By injecting an artificial delay for the chatbot
cues in a chatbot is not to convince users they to “read and respond,” the conversation mimics
are talking to a human when they are not, but to human messaging conversations. However,
make them feel that the conversation is natural developers should be good stewards of users’
and comfortable. Figure 2 graphically summarizes time and not make the delay unnecessarily long.
the different types of cues (or signals) that can be We have found that a brief delay of a couple
used in a chatbot. of seconds for “reading” and a typing delay
Identity Cues. The first set of cues that can of around 50 milliseconds per character (an
create a more humanlike chatbot are identity extremely fast typing speed) makes the delay
cues, such as an avatar, a name and self- feel a bit more human while also keeping the
references as part of the responses. These cues conversation moving.
imply that the chatbot has a unique identity and A typing indicator is a related feature to
help to make it appear more humanlike. The accompany a typing delay. As with the typing
use of a photo of a human as an avatar for the
41 Porra, J., Lacity, M. and Parks, M. S. “Can Computer Based
40 Seeger, A.-M., Pfeiffer, J. and Heinzl, A. “Designing Anthropo- Human-Likeness Endanger Humanness? – A Philosophical and Ethi-
morphic Conversational Agents: Development and Empirical Evalu- cal Perspective on Digital Assistants Expressing Feelings They Can’t
ation of a Design Framework,” Proceedings of the 39th International Have,” Information Systems Frontiers (22), December 2019, pp. 553-
Conference on Information Systems, 2018. 547, available at https://doi.org/10.1007/s10796-019-09969-z.

10 MIS Quarterly Executive | March 2021 (20:1) misqe.org | © 2021 University of Minnesota
 Deciding Whether and How to Deploy Chatbots

Figure 2: Implicit Signals Can Create Natural, Humanlike Conversations with Chatbots

delay itself, the indicator mimics what users will be able to pass the Turing Test, there are
expect to see in a text chat. four types of verbal cues developers build in to
The use of emoticons, emojis, common texting make the interaction better for users: tailoring,
abbreviations and other forms of formal or conversational variety, past references and small
informal language can introduce a degree of talk.
humanlike behavior to a chatbot.42 Tailoring refers to the ability of a chatbot to
Verbal Cues. The final category of cues that respond in a way customized to user input. This
can make a chatbot more humanlike is verbal ability relies on understanding how people will
cues. These cues are provided through the speak to a chatbot and the types of conversations
words the chatbot uses to interact with users. they are likely to have. To provide tailored verbal
Simulating a human conversation is a daunting cues, developers can give a chatbot a variety of
task that has been one of the main goals of ways to respond. For our interview chatbots, we
conversational artificial intelligence research typically program 15 to 25 different responses
since Alan Turing first proposed the Turing Test. based on what we think users might say. This
However, even though it is unlikely that a chatbot number could be higher depending on how
constrained the interaction is.
42 Vivibot is an example of a chatbot that uses intentional gram-
matical errors and informal language in an attempt to appeal to a
Conversational variety refers to the number
younger audience. For more information, see Greer, S., Ramo, D., of ways a chatbot can say the same thing. Users
Chang, Y.-J., Fu, M., Moskowitz, J. and Haritatos, J. “Use of the expect variation from natural language, so giving
Chatbot ‘Vivibot’ to Deliver Positive Psychology Skills and Promote
a chatbot a variety of ways to respond (e.g., “It’s
Well-Being Among Young People After Cancer Treatment: Random-
ized Controlled Feasibility Trial” JMIR MHealth and UHealth (7:10), going to be 78 degrees today” or “Today it will be
October 2019, e15018, available at https://doi.org/10.2196/15018.

March 2021 (20:1) MIS Quarterly Executive 11

Deciding Whether and How to Deploy Chatbots

78 degrees outside”) makes the interaction feel companies use chatbots as the first line of contact
more natural. with customers, the reluctance to engage with
Past reference cues are a valuable way to a chatbot has decreased. Modern chatbots can
increase the conversational quality of a chatbot. answer a wide variety of questions in a variety
In general, people do not expect chatbots to of contexts and will save time and money as
remember things about their past conversations, customers get automated answers to their
so adding those references is likely to delight questions. Providing a good chatbot experience
users. is far more valuable than any temporary initial
Finally, providing a chatbot with some level of benefit that might be gained from telling
small talk can help users relate to even the most customers they are talking with a human when in
task-focused of chatbots. Handling small talk fact they are not.
and probing behaviors—when users try to find
the limitations or weaknesses in the chatbot— 2. Use Cues to Make the Conversation
is a signal to users that it is capable of handling Feel Natural
human behavior. The small talk responses can Subtle cues can help put customers at ease and
give the agent some personality, or they can be make them feel they are talking with a person,
straightforward—e.g., “I am just a bot, let’s stay even if they know they are not. Short typing
on topic.” Google’s DialogFlow tool includes delays—short enough to give users a break,
a small talk module that enables a chatbot to without being so long that it wastes their time—
answer simple questions about itself, respond to contribute to this. Giving the chatbot an identity
greetings and goodbyes, and answer other off- and personality also helps create the feeling of
topic user messages. a conversation. The identity can personify the
The examples given above of identity, company or be a distinct personality suited to the
nonverbal and verbal cues that can be used chatbot’s role.
to create a more humanlike conversation are
not exhaustive. The number and types of cues 3. Consider the Limitations of Chatbots
developers include in a chatbot will depend on Natural language processing technologies
the goals of the chatbot and the capabilities of the have come a long way since the 1960s when the
platform. If the chatbot’s role is to create a social first chatbots were created, and new levels of
connection with users, more cues can help the understanding are constantly possible. Chatbots
conversation feel natural and encourage users are well-suited for answering queries, directing
to connect with it (and hence the organization). users and connecting on a social level in a way
If the chatbot interaction is purely transactional, that other interfaces cannot. But they are still
users may prefer fewer cues. not capable of understanding complex, multi-
part queries. Nor can they come up with new
Recommended Best Practices responses by themselves, though this is an
active area of research, and these capabilities
for Deploying Chatbots are improving quickly.43 If the use case does not
From our research, we have identified six best match well with current chatbot technology, then
practices for planning, designing, developing a chatbot is not the right solution.
and deploying conversational AI (i.e., chatbots).
Following these best practices will help 4. Provide Prompts to Show Customers
organizations create the optimum experience for What the Chatbot Can Do
customers and users, and thereby increase their Both text and voice chatbots suffer from
engagement with the business. problems of “discoverability” (i.e., users find it
hard to find out what the chatbot can do). Unlike
1. Don't Deceive Customers traditional user interfaces, chatbots do not have
In addition to being ethically questionable,
lying to customers about the nature of your 43 For a review of the state-of-the-art in natural language genera-
tion, see Dušek, O., Novikova, J. and Rieser, V. “Evaluating the
organization’s chatbot will more likely frustrate State-of-the-Art of End-to-End Natural Language Generation: The
them rather than delight them. As more E2E NLG challenge,” Computer Speech & Language (59), July 2020,
pp. 123-156.

12 MIS Quarterly Executive | March 2021 (20:1) misqe.org | © 2021 University of Minnesota
 Deciding Whether and How to Deploy Chatbots

Figure 3: Example of Quick Response Buttons

menus to indicate what they can do, so, without 6. For Customer Service Chatbots,
experience, customers may be unsure how to Provide a Way to Connect to a Human
approach a chatbot. Many chat platforms include In chat technology, the ability to connect
options for getting quick replies by clicking on with a human is often called “human-in-the-
a button (see Figure 3). Providing a chatbot loop” (HITL). Even the best chatbots are bound
with similar features helps to guide customers, to fail at some point. Just as a good automated
teaching them about new capabilities and phone system provides a way to get to an agent
demonstrating the types of messages the chatbot or operator—hopefully without users shouting
is designed to answer. “REPRESENTATIVE!” into their phones—bots
should do the same. If a chatbot fails to provide
5. Provide the Chatbot with the Ability an adequate response after a couple of tries,
to Handle Small Talk and Probing the best HITL implementations automatically
Behaviors transfer the customer to a human agent. The
Many people want to probe the limitations of a automated transfer should allow the human agent
chatbot as soon as they discover they are talking to review the previous chatbot conversation so
to one.44 Providing a chatbot with some “canned” the customer does not have to repeat herself.
responses to these types of questions can give it The chatbot could even be designed to identify
some personality and increase customers’ long- queries it is unable answer and immediately
term engagement. connect the customer to a human.
An example of applying these six best practices
is provided by Apple’s customer service chatbot,
44 A study of a statistics tutor chatbot found that about 30% of as described in the text panel on the next page.
messages sent to it in the first week were small talk or probing. This
decreased over the semester, but the users who sent the small talk
messages were more engaged with the tutorbot throughout the semes- Concluding Comments
ter. See Hobert, S. and Berens, F. “Small Talk Conversations and the
Long-Term Use of Chatbots in Educational Settings – Experiences In this article, we have identified the three
from a Field Study,” In Følstad, A., Araujo, T., Papadopoulos, S., questions organizations need to answer when
Law, E. L.-C., Granmo, O.-C., Luger, E. and Brandtzaeg, P. B. (eds.) deciding whether and how to deploy chatbot
Chatbot Research and Design, SpringerLink, January 2020, pp. 260-
272, available at https://doi.org/10.1007/978-3-030-39540-7_18. technology. If they decide a chatbot is an

March 2021 (20:1) MIS Quarterly Executive 13

Deciding Whether and How to Deploy Chatbots

Case: Apple Customer Service Chatbot

Apple has introduced a chatbot as its first line of customer support for the most common types of
inquiries. This implementation is an excellent example of following our recommended best practices.
(1) From the first message, the chatbot discloses to customers that it is an automated system. (2)
Though it does not include all of the cues described earlier, the interface includes a one-second
delay after users send a message to make the conversation feel more natural, even though it is fully
transparent about its automation. (3) The chatbot focuses on answering commonly asked customer
service questions, which is an excellent use case for conversational technology. (4) In its first message,
the chatbot provides examples of the types of queries it can answer. (5) The chatbot does not handle
probing or small talk, but if users go off-topic, it gently nudges them back with an array of buttons
for the types of queries it can support. (6) At any point in the conversation, customers can request
to talk to a human specialist and are immediately put in touch with one. If the chatbot detects that it
has failed to respond correctly multiple times, it automatically pulls a specialist into the conversation.
app.”

appropriate solution for the identified business About the Authors

problem, they then need to consider which
chatbot technology to use and how humanlike
Ryan M. Schuetzler
the chatbot should be. Answering these questions
Dr. Ryan Schuetzler (ryan.schuetzler@byu.
will help organizations direct their chatbot efforts
edu) is an Assistant Professor of Information
toward their most effective use.
Systems in the Marriott School of Business at
The use of chatbots has grown massively in
Brigham Young University. He conducts research
the past five years. With the availability of new
studying the impact of chatbot design on user
tools from major suppliers like Google, Microsoft
behavior and acceptance. His main focus is on
and IBM, it has never been easier to develop and
making chatbot conversations feel natural and
deploy a chatbot to provide 24/7 support for
engaging, and on the effectiveness of chatbots
customers or employees. However, the growing
compared to traditional technology.
popularity of chatbots may lead companies to
deploy a chatbot solution just because it seems G. Mark Grimes
like the next cool thing to do. Don’t do this. Dr. Mark Grimes (gmgrimes@bauer.
If a chatbot is deployed inappropriately, the uh.edu) is an Assistant Professor of Decision
organization and its customers may form the and Information Sciences at the University of
impression that “bots are bad,” when in fact the Houston. His research focuses on conversational
problem was that the implementation and use agents, information systems security and
case were misaligned. To avoid this problem, the analysis of human-computer interaction
organizations should consider if the use case behaviors to detect changes in emotional and
fits with the unique strengths and weaknesses cognitive states. Mark received his Ph.D. in
of conversational technology before beginning Management Information Systems from the
a chatbot project. Chatbots are best used in University of Arizona.
situations where their ability to connect with
users on a social level is an advantage. Because Justin Scott Giboney
of the unique social nature of a chatbot, a poorly Dr. Justin Giboney (justin_giboney@byu.
guided effort to build one can result in a system edu) is an Associate Professor of Information
that is at best useless and frustrating. However, Technology and Cybersecurity at Brigham Young
a chatbot project that properly incorporates University. He received his Ph.D. in Management
the unique characteristics of conversational Information Systems from the University of
interfaces can delight users and create a better Arizona, and master’s and bachelor’s degrees
connection to your company. in Information Systems from Brigham Young
University. His research focuses on information,

14 MIS Quarterly Executive | March 2021 (20:1) misqe.org | © 2021 University of Minnesota
 Deciding Whether and How to Deploy Chatbots

security and forensic technologies, and deception-

detection and knowledge-based systems. His
work has led to a better understanding of how
security experts store knowledge, make decisions
and interact with technology.

Holly K. Rosser
Holly Rosser (hrosser@unomaha.edu) is a
doctoral student in the College of Information
Systems and Technology at the University of
Nebraska Omaha. Her research interests include
human-computer interaction, human-bot
interaction, instructional design and informal
STEM (science, technology, engineering and
math) learning. Her work has been published in
Information Systems Frontiers and Journal of Peer
Production, and in the proceedings of the Hawaii
International Conference on System Sciences
and the 2018 conference of the ACM on Human-
Computer Interaction (CSCW 2018).

March 2021 (20:1) MIS Quarterly Executive 15