Artificial Intelligence in Organizations: Current State and Future Opportunities

Full text

Artificial Intelligence in Organizations: Current State and Future Opportunities
Hind Benbya, Deakin University and Oxford Internet Institute (Australia and U.K.)
Thomas H. Davenport, Babson College and Oxford Saïd Business School (U.S.)
Stella Pachidi, Cambridge University (U.K.)
Recommended citation: Benbya, Hind; Davenport, Thomas H.; and Pachidi, Stella (2020) "Artificial Intelligence
in Organizations: Current State and Future Opportunities," MIS Quarterly Executive: 19 (4).
Available at: https://aisel.aisnet.org/misqe/vol19/iss4/4
Introduction
Artificial intelligence (AI) is typically defined as the ability of machines to perform human-like cognitive
tasks. These can include automation of physical processes such as manipulating and moving objects,
sensing, perceiving, problem-solving, decision-making and innovation
1
. AI is currently viewed as the
most important and disruptive new technology for large organizations.
2
However, the technology is
still in a relatively early state in large enterprises, and largely absent from smaller ones other than
technology startups. Surveys
3
suggest that fewer than half of large organizations have meaningful AI
initiatives underway, although the percentage is increasing over time.
For most organizations, AI projects remain somewhat experimental—undertaken as a pilot or proof of
concept. Relatively few organizations have deployed AI on a production basis, a problem that we
describe in greater detail below. The experimental use, of course, means that many organizations have
achieved little or no economic return on their AI investments. However, some analysts
4
suggest that
AI adoption will eventually have considerable positive impact on company growth and profitability.
AI is being applied in organizations for diverse objectives
5
: to make processes more efficient (28% as
one of top two), to enhance existing products and services (25%), to create new products and services
(23%), to improve decision-making (21%), and to lower costs (20%). Although a common theme in the
AI-oriented press is related to reducing headcount, this objective got the lowest number of mentions
at 11%.
Executives initially focused on using AI technologies to automate specific workflow processes and
repetitive work. Such processes were linear, stepwise, sequential and repeatable. But now, firms are
1
Innovation is defined here as the design, creation, development and/or implementation of new or altered
products, services, systems, organizational structures, management practices and processes, or business models;
see Benbya, H. and Leidner, D. (2018) "How Allianz UK Used an Idea Management Platform to Harness Employee
Innovation," MIS Quarterly Executive (17:2), 2018, pp ; Yan, J. Leidner, D. and Benbya, H. "Differential
Innovativeness Outcomes of User and Employee Participation in an Online User Innovation Community," Journal
of Management Information Systems (35:3), pp. 900-933.
2
NewVantage (2019) “Big data and AI executive survey 2019, executive summary of findings,” NewVantage
Partners, https://newvantage.com/wp-content/uploads/2018/12/Big-Data-Executive-Survey-2019-Findings-
Updated-010219-1.pdf
3
Genpact, 2020. “AI 360: Hold, fold, or double down,” https://www.genpact.com/uploads/files/ai-360-research-
2020.pdf
4
McKinsey & Co., 2018. “Notes from the AI frontier: modeling the impact of AI on the world economy”
https://www.mckinsey.com/featured-insights/artificial-intelligence/notes-from-the-AI-frontier-modeling-the-
impact-of-ai-on-the-world-economy#
5
Deloitte (2020) “Thriving in the era of pervasive AI: Deloitte’s state of AI in the enterprise, 3rd edition,” Deloitte
Insights, https://www2.deloitte.com/us/en/insights/focus/cognitive-technologies/state-of-ai-and-intelligent-
automation-in-business-survey.html

moving toward nonsystematic cognitive tasks that include decision-making, problem-solving and
creativity, which until recently seemed beyond the scope of automation. AI technologies are also
progressively enabling people and machines to work collaboratively in novel ways. In manufacturing,
for example, in order to fulfill customized orders and handle fluctuations in demand, employees are
partnering with robots to perform new tasks without having to manually overhaul any processes. AI
technologies are also performing certain tasks autonomously, though complex ones like driving a car
in all conditions remain tantalizingly out of reach.
We are beginning, however, to see autonomous systems that can perform tasks without any human
involvement at all, as the system can train itself and adjust to new training data. Consider automated
financial trading. Because it depends entirely on algorithms, companies can complete transactions
much faster with it than with systems relying on humans. In a similar fashion, robots are performing
narrow tasks autonomously in manufacturing settings.
6
Some companies, such as Amazon.com and Google, have attempted to create highly ambitious
applications of AI, including autonomous vehicles, unattended retail checkout, and drone delivery.
Some of these “moon shots” have been successful, but some highly ambitious projects, including
cancer treatment, have been largely unsuccessful thus far despite considerable expenditures
7
. Less
ambitious “low hanging fruit” projects have been more successful in most firms and are perhaps more
consistent with the narrow intelligence possessed by AI systems at the moment.
Likewise, most autonomous AI applications remain limited to low risk areas where the cost of failure
is limited. Although many AI systems can do certain things better than humans, workers’ trust in AI
technology is still limited due to the issues such technology might raise like algorithmic bias,
unexplainable outcomes, invaded privacy and/or lack of accountability. Consumers are also skeptical
about AI, and surveys suggest that most or many would not want autonomous vehicles, do not like
dealing with chatbots, and so forth.
This December 2020 Special Issue of the MIS Quarterly Executive is titled “AI in organizations: current
state and future opportunities.” It details current challenges and implications that might arise from AI
applications, and the ways to overcome such challenges to realize the potential of this emerging
technology. The collection of papers in this issue (December) combined with a forthcoming (March)
article will be insightful to managers who are currently running digital transformation initiatives driven
by AI technology, to practitioners who are considering implementing AI in their businesses, and to
research-oriented faculty and students. In this editorial, we first provide a brief history of AI and an
overview of AI typologies. We discuss current challenges, implications and future opportunities
regarding AI to enable readers to better understand the five papers in the special issue. Finally, we
summarize the special issue articles and highlight the contributions each makes.
Brief History of AI
AI as an academic field date back to the 1950s. The term AI was first introduced during a
multidisciplinary program presented at Dartmouth in 1956. The program aimed to study the possibility
that machine intelligence could imitate humans and involved researchers from various fields including
scientists, mathematicians, and philosophers.
6
Davenport, T. H., & Ronanki, R. (2018). Artificial intelligence for the real world. Harvard Business Review,
96(1), 108–116.

Despite early promises of the practical usefulness of AI, it largely failed to deliver and faced several
obstacles during the 1960s and 1970s. The biggest of which was the lack of computational power to
do anything substantial. Research funding gradually stalled and the field lost momentum. During the
1980s and 1990s, governments and firms made significant investments in research on expert systems,
which rejuvenated interest in AI. Machine learning and neural networks began to flourish as its
practitioners integrated statistics and probability into their applications. At the same time, the personal
computing revolution began. Over the next decade, digital systems, sensors, and the Internet would
become more common, providing all kinds of data for machine-learning experts to use when training
adaptive systems. Although the growth of AI and machine learning has been intermittent over the
decades, unprecedented computing capacity and growing volumes of data have given momentum to
the recent development of artificial intelligence applications.
AI Types and technologies
There are many types of AI systems. One typology differentiates AI systems based on the kind of
intelligence they display. A second typology distinguishes AI applications based on the type of
technology embedded into the AI system, whereas a third is based on the function performed by the
AI.
Based on intelligence: Philosophical debates on AI are centered on the notion of intelligent machines,
that is machines that can learn, adapt and think like people
8
. AI types based on such a notion fall in
general into three categories: artificial narrow intelligence, artificial general intelligence and artificial
super intelligence.
While narrow (or weak) AI is usually able to solve only one specific problem and is unable to transfer
skills from domain to domain, general AI aims for a human-level skill set. Once general AI is achieved,
it is believed that it might lead to superintelligence that exceeds the cognitive performance of humans
in virtually all domains of interest.
9
This type of superintelligence can emerge following evolutionary
and complex adaptive systems principles
10
. The argument states that if humans could create AI
intelligence at a roughly human level, then this creation could, in turn, create yet higher intelligence
and eventually evolve further
11
. AI enthusiasts are providing estimates and outline scenarios for when
technological growth will reach the point of singularity, where machine intelligence will surpass human
intelligence. This raises philosophical arguments about the mind and the ethics of creating artificial
beings endowed with human-like intelligence. Although the futuristic literature assumes that AI
systems will be able to perform all tasks just as well as, or even better than, humans, this type of
artificial general intelligence does not exist yet. There are, however, some AI programs, such as the
GPT-3 language prediction application, that are beginning to exhibit some aspects of general
intelligence.
12
8
Lake, B., Ullman, T., Tenenbaum, J. and Gershman, J. “Building machines that learn and think like people,”
Behavioral and Brain Sciences, 2017
9
Bostrom, N. (2014). Superintelligence Paths, Dangers, Strategies. Oxford, Oxford University Press.
10
See Benbya, H. Nan, N., Tanriverdi, H. and Yoo, Y. (2020), “Complexity and information systems research in the
emerging digital world,” MIS Quarterly (44: 1), pp. 1-18, for a recent article on evolutionary principles, and
Benbya, H. and McKelvey, B. “Using coevolutionary and complexity theories to improve IS alignment: a multi-
level Approach,” Journal of Information Technology (21:4), 2006, pp. 284-298 for an elaboration of such principles
in IT management.
11
Hawking, S., Russell, S., Tegmark, M., & Wilczek, F. (2014). Transcendence looks at the implications of
artificial intelligence - but are we taking AI seriously enough? The Independent, 01.05.2014.
12
GPT-3 stands for generative pre-trained transformer version three. It is a powerful machine-learning system
that can rapidly generate text with minimal human input. After an initial prompt, it can recognise and replicate
patterns of words to work out what comes next, see Thierry, 2020. New AI can write like a human but don’t
mistake that for thinking, The Conversation, Sept. 17., 2020, https://theconversation.com/gpt-3-new-ai-can-
write-like-a-human-but-dont-mistake-that-for-thinking-neuroscientist-146082

Based on technology: A second typology differentiates between the technologies embedded into the
AI systems which include machine learning, (its subclasses deep learning and reinforcement learning),
natural language processing, robots, various automation technologies (including robotic process
automation), and rule-based expert systems (still in broad use although not considered a state-of-the-
art technology). One recent survey
13
suggests that all the contemporary AI technologies (machine
learning, deep learning, natural language processing) are either currently being used or will be used
within a year by 95% or more of large adopters of AI. Table 1 below provides brief definitions and
domain of applications of AI technologies.
Technology
Brief Description
Example Application
Machine learning
• Reinforcement learning
• Supervised learning
• Unsupervised learning
Learns from experience
Learns from a set of training data
- Detects patterns in data that aren’t
labeled and for which the result isn’t
known
Highly granular marketing
analyses on big data
Deep Learning
A class of machine learning that
learns without human supervision,
drawing from data that is both
labeled and unlabeled.
Image and voice recognition,
self-driving cars
Neural Networks
Algorithms that endeavor to
recognize underlying relationships in
a set of data through a process that
mimics the way the human brain
operates.
Credit and loan application
evaluation, weather prediction
Natural Language Processing
A computer program able to
understand human language as it is
written or spoken
Speech recognition, text
analysis, translation, generation
Rule-based expert systems
A set of logical rules derives from
human experts
Insurance underwriting, credit
approval
Robotic process automation
Systems that automate structured
digital tasks and interfaces
Credit card replacement,
validating online credentials
Robots
Automatically operated machines
that automate physical activity,
manipulate and pick up objects
Factory and warehouse tasks
Table 1: AI technologies and domains of application
Based on function: This distinction differentiates between four types of AI: conversational, biometric,
algorithmic, and robotic. These categories overlap somewhat; for example, conversational and
biometric AI already make extensive use of algorithmic AI models, and robotic AI is increasingly doing
so as well.
Conversational AI refers to the general capability of computers to understand and respond with natural
human language. Such systems include both voice and text-based technologies and vary largely based
on their capability, domain and level of embodiment. Simple conversational AI are mainly used to
handle repetitive client queries whereas smart conversational AI, enabled by machine learning and
natural language processing, have the potential to undertake more complex tasks that involve greater
interaction, reasoning, prediction, and accuracy. Conversational AI have been used in many different
fields, including finance, commerce, marketing, retail, and healthcare. Although the technology behind
13
Deloitte (2020) “Thriving in the era of pervasive AI: Deloitte’s state of AI in the enterprise, 3rd edition,”
Deloitte Insights, https://www2.deloitte.com/us/en/insights/focus/cognitive-technologies/state-of-ai-and-
intelligent-automation-in-business-survey.html

smart conversational agents is continuously under development, they currently do not have full
human-level language abilities, sometimes resulting in misunderstanding and users’ dissatisfaction.
14
Biometric AI: Biometrics relies on techniques to measure a person’s physiological (fingerprints, hand
geometry, retinas, iris, facial image) or behavioral traits (signature, voice, keystroke rhythms).
AI powered biometrics uses applications such as facial recognition, speech recognition and computer
vision for identification, authentication and security objectives in computer devices, workplace, home
security among others. While fingerprints have the longest history as a marker of identity and continue
to be used in a number of applications across the world
15
, other bodily markers like face, voice, and
iris or retina are proliferating, with significant research exploring their potential large-scale application.
Meanwhile, the ubiquity of face images and voice recordings tagged with people’s names on the
Internet alongside algorithms to transform such data into biometric recognition systems has
accelerated their use at a larger scale. Examples include identifying suspects, monitoring large events
and surveilling protests. Such large-scale use has triggered calls for regulation to introduce new laws,
reform existing laws, or ban their use in some contexts.
Algorithmic AI revolves around the use of machine learning (ML) algorithms— a set of unambiguous
instructions that a mechanical computer can execute. Some ML algorithms can be trained on
structured data and are specific to narrow task domains, such as speech recognition and image
classification. Other algorithms, especially deep learning neural networks, are able to learn from large
volumes of labeled data, enhance themselves by learning, and accomplish a variety of tasks such as
classification, prediction and recognition. For example, a neural network can analyze the parameters
of bank clients such as age, solvency, and credit history, and decide whether to approve a loan request.
It can use face recognition to only let authorized people into a building. And it can predict outcomes
such as the rise or fall of a stock based on past patterns and current data. Despite the potential of ML
algorithms, there are concerns that in some cases it may not be possible to explain how a system has
reached its output. They may also be susceptible to introducing or perpetuating discriminatory bias.
Robotic AI: Physical robots have been used for many years to perform dedicated tasks in factory
automation. Recently, AI including ML and NLP, has become increasingly present in robotic solutions
enabling robots to move past automation and tackle more complex and high-level tasks. AI-enabled
robots are equipped with the ability to sense their environment, comprehend, act, and learn. This
helps robots do a lot of tasks from successfully navigating their surroundings, to identifying objects
around the robot or assisting humans with various tasks such robotic-assisted surgeries.
Current Challenges
AI’s Deployment Problem
One of the major concerns with AI in organizations at present is that many systems are only
experimental, and never deployed in production. A pilot AI project is relatively easy to develop and is
only meant to demonstrate that the technology is feasible in concept. Deployment, on the other hand,
requires a variety of tasks and capabilities that may be in short supply. These can include, for example,
integration with existing technology architectures and legacy infrastructure, change in business
processes and organizational culture, reskilling or upskilling of employees, substantial data
14
What is a Chatbot? All You Need to Know About Chatbots!. Botpress: Open-Source Conversational AI Platform.
2018. URL: https://botpress.io/learn/what-and-why/
15
Amba Kak, ed., “Regulating Biometrics: Global Approaches and Urgent Questions” AI Now Institute, September
1 2020, https://ainowinstitute.org/regulatingbiometrics.html.

engineering, and approaches to organizational change management. Full production deployment
tends to take much longer than pilots and cost substantially more.
Surveys of organizations and market research reports in the US and globally suggest that deployment
challenges are widespread with big data and AI. A survey16 (of large financial services and life sciences
firms found that firms were actively embracing AI technologies and solutions, with 91.5% of firms
reporting ongoing investment in AI. But only 14.6% of firms reported that they have deployed AI
capabilities into widespread production. In a 2019 global McKinsey survey with the headline “AI
adoption proves its worth, but few scale impact,” between 12% (in consumer packaged goods) and
54% (in high tech firms) had at least one machine learning application implemented in a process or
product. Only 30% of respondents overall were using AI in products or processes across multiple
business units and functions.17
In order to address these deployment concerns, companies need to plan for the possibility of
deployment from the beginning. Some companies, such as Farmers Insurance, have a well-defined
process that seeks to move projects when appropriate from the pilot phase to full deployment
18
. In a
survey of US early adopter organizations, 54% of executives said their organization has a process for
moving prototypes into production; and 52% had an implementation road map. These organizational
approaches would seem to be an aid to getting more AI systems into deployment, but they may only
be in the early stages.
AI Talent Issues
Securing a sufficient volume and level of human AI talent is a challenge for many organizations—
particularly those that are not in the technology sector. Data scientists and AI engineers are still scarce,
although many university programs have arisen to train them. Firms that can’t pay high levels of
compensation and aren’t based in technology centers are likely to have difficulty hiring the desired
number of skilled employees. Many companies should attempt not only to hire new employees with
AI skills, but to retrain existing employees to the degree possible.
Even when companies do manage to hire data scientists and other types of analytical and artificial
intelligence talent, there is little consensus within and across companies about the qualifications for
such roles. The term “data scientist” might mean a job with a heavy emphasis on statistics, open-source
coding, or working with executives to solve business problems with data and analysis. Some view the
role only as developing models, others as extending to deployment of the models in production. The
idea of data scientist “unicorns” who possess all these skills at high levels was never very realistic
19
.
Within university programs to train AI-oriented workers, the skills taught in such programs vary widely,
and some universities offer multiple programs with different emphases. For both newly hired and
experienced employees, titles such as data scientist and AI engineer are not likely to be a good guide
to their actual capabilities. Further, activities involved in deployment of AI systems and related
16
NewVantage (2019) “Big data and AI executive survey 2019, executive summary of findings,” NewVantage
Partners, https://newvantage.com/wp-content/uploads/2018/12/Big-Data-Executive-Survey-2019-Findings-
Updated-010219-1.pdf
17
Deloitte (2018) “State of AI in the enterprise, 2nd edition,” Deloitte Insights,
https://www2.deloitte.com/content/dam/insights/us/articles/4780_State-of-AI-in-the-enterprise/DI_State-of-
AI-in-the-enterprise-2nd-ed.pdf
18
Davenport, T. and Bean, R., 2018. “Farmers accelerates its time to Impact with AI.” Forbes, August 1.
https://www.forbes.com/sites/tomdavenport/2018/08/01/farmers-accelerates-its-time-to-impact-with-
ai/#51430150b672
19
Davenport, T., (2020). “Beyond unicorns: educating, classifying, and certifying data science talent,” Harvard
Data Science Review, May 19, https://hdsr.mitpress.mit.edu/pub/t37qjoi7/release/2

organizational change issues may not be taught at all by many technically focused programs. There is
an increasing need of a new type of professionals who can understand the business problems and
translate them into algorithmic problems, and vice versa to explain the technical insights to business
managers.
20
There are initiatives
21
in the early stages to standardize the different types of data, analytics, and AI
roles and requisite skills across organizations. This is an excellent idea, but developing new standards
typically takes many years.
In the meantime, companies need to devote considerable attention to classifying and certifying the
different types of AI and data science jobs needed in their organizations. Companies also would benefit
from expanding their talent pool by working with universities directly on educational programs, and
by building and nurturing communities within their organizations for employees on their data teams.
These steps are essential for companies looking to use AI to improve both current operations and
opportunities for digital innovation.
AI and social dysfunctions
Aside from the deployment and talent challenges, there are a few other potential dysfunctions from
AI that managers need to be aware of and plan to avoid.
Algorithmic bias: The employment of AI systems in classification or prediction tasks often comes with
the risk of algorithmic bias, which means that the outcomes of the machine learning algorithm can put
certain groups at a disadvantage
22
. This has already been observed in various cases, including
algorithms that are used to score job applicants and appear to be racist, or algorithms that recommend
sentences to judges and appear to propagate the preconceptions infused in past sentencing decisions
that were used as training data Algorithmic bias can also have consequences distributed across large
subsections of society by affecting the type of information that people are exposed to. This happens,
for example when the machine learning algorithms behind social media propagate fake news or enable
the targeting of individuals for political campaigns. To reduce any potential algorithmic bias, managers
will need to be proactive by performing small-scale experiments and simulations before implementing
such algorithms; regularly evaluate the dataset used for training; and involve human reviewers who
regularly provide feedback to the system designers. In politically and socially sensitive domains like
judicial sentencing, firms may find it necessary to publish their algorithms to preclude accusations of
bias.
Unexplainable decision outcomes: The possible social dysfunctions from AI implementation can
increase if one considers the fact that the decision outcomes of some machine learning algorithms—
deep learning in particular—cannot be easily explained due to the vast amount of feature layers
involved in their production. This could lead to problematic situations, such as unexplainable
evaluations of high school teachers, or parole decisions that cannot be justified and may cause rage
when they also appear unfair
23
. Organizations need to respond to regulators’ calls for explainability by
avoiding “black box” AI applications and by choosing algorithms whose outcomes can be explained.
Being open about the data that is used and explaining how the model works in non-technical terms is
20
Henke, N., Levine, J., and McInerny, P., (2018) “You don’t have to be a data scientist to fill this must-have
role,” Harvard Business Review, Feb. 5, https://hbr.org/2018/02/you-dont-have-to-be-a-data-scientist-to-fill-
this-must-have-analytics-role
21
See for example, https://www.iadss.org/
22
Davenport, T. H. (2018). The AI advantage: How to put the artificial intelligence revolution to work. MIT Press.
23
O’Neil (2016) Weapons of math destruction: How big data increases inequality and threatens democracy.
Broadway Books.

also necessary to ensure customers’ trust and to avoid potential dysfunctions triggered by lack of
transparency. In some industries such as banking, regulators sometimes force firms to use explainable
algorithms.
Blurring accountability boundaries: As AI is used to enhance or even automate decision making
procedures, the issue of accountability arises. Who is responsible in the case of a traffic accident with
a driverless car? Who is responsible for approving parole to a criminal who eventually commits another
crime? Who is responsible for a big financial loss in algorithmic trading? These are only few of the cases
where the accountability boundaries are blurred. Managers will need to proactively focus on the
reasons and processes that may lead to potential harm. They also need to carefully consider how they
engage the different actors that directly or indirectly interact with the outcomes produced by the AI
system (AI developers and designers, business users, institutions), and clarify responsibility and legal
liability upfront
24
.
Invaded privacy: Ethical issues arise even before any action is recommended or performed by the AI
system, with privacy being reported as one of the main ethical considerations behind AI
implementation
25
. Data is the primary resource that is fed into the AI systems, and quite often is seen
as a source of competitive advantage. AI’s need to process increasingly large amounts of data thus
conflicts with people’s right to maintain control over their data and its use in order to preserve their
privacy and autonomy. Organizations need to ensure that their data practices comply with the relevant
policies on the use of personal data (e.g. GDPR in EU countries) and avoid any possible privacy violation.
Developing auditable algorithms and performing algorithmic audits on them to identify what data is
used and what variables feed into a decision-making procedure are helpful solutions to increase
transparency of how consumers’ data are processed and used
26
. Overall, being open about how the
data is handled is necessary to ensure customers’ trust.
Implications
Despite existing challenges, AI has the potential to dramatically change how the workforce is
structured, how jobs are designed, how knowledge is managed, and how decisions are made.
These changes will have broader implications on organizations and societies, many of which have
yet to be understood or realized. But the most common effects are likely to be on how work is
conducted in the future.
AI and the future of work: Recent developments in AI are already affecting the workplace in different
ways:
Automating work tasks: AI will have significant impact on several occupations, by automating mundane
tasks and rendering various human skills obsolete. Given that AI can perform tasks that previously
required human judgment, the effects of AI-enabled automation differ from those of past
technologies, as for the first time they get to affect knowledge workers
27
. Professions such as doctors,
lawyers, consultants or architects, whose expertise, judgment and creativity have thus far been highly
24
Dourish, P. (2016). Algorithms and their others: Algorithmic culture in context. Big Data & Society, 3(2), 1-11.
25
Kinni (2017) Ethics Should Precede Action in Machine Intelligence. MIT Sloan Management Review
26
Mittelstadt, B. (2016). Automation, algorithms, and politics| auditing for transparency in content
personalization systems. International Journal of Communication, 10, 12.
27
See Davenport, T. (2005), Thinking for a Living: How to Get Better Performance and Results from Knowledge
Workers, Harvard Business School Press ; Benbya, H. (2008), Knowledge Management Systems Implementation :
Lessons From the Silicon Valley, Neal-Schuman Publishers, and Faraj, S., Pachidi, S., and Sayegh, K. 2018. "Working
and Organizing in the Age of the Learning Algorithm," Information and Organization (28:1), pp. 62-70.

valued and considered irreplaceable, for the first time in history appear threatened. And while the end
of those professions is not for the near future, the changing nature of their work is already a reality.
There are many predictions about how much job loss from AI will take place, but thus far it has been
relatively small.
28
Changing expertise: AI technology that is able to automate some of the workers’ tasks is already in the
workplace. In law firms for example, a plethora of applications have been developed for automating
the due diligence and contract review tasks that were previously performed by junior lawyers. In sales,
conversational AI can now automate various tasks that previously had to be carried out by account
managers. While such automations can increase efficiency of operations and decrease labor costs, they
leave professionals with voids in the processes they used to acquire knowledge about their subjects or
customers, or the ways through which they would develop their expertise. This will eventually lead to
changes in the knowledge of the affected occupations and could potentially even trigger their
restructuring. For example, in the legal profession there is already the tendency for various law
graduates to develop data science skills and engage with legal tech instead of following the traditional
career path of a lawyer.
Augmenting professionals: In several of the cases, AI systems are not yet able to replace human
experts, but they can augment their work by supporting experts’ judgment and decision-making
processes. For example, the debate has now moved away from the “end of radiologists” focus, and
acknowledges that radiologists will not be replaced by AI tools any time soon, but they will be
augmented by them.
29
Yet, as AI systems are introduced in the radiology profession to support the
radiologists’ diagnosis process, we begin to see several unintended consequences on their everyday
work: from having to overcome communication barriers in their unavoidable interactions with data
scientists, to even doubting the prediction of the AI system or questioning their own diagnosis
30
. This
becomes even more complicated if we consider that most often, the way in which a machine learning
algorithm functions and comes to render a specific outcome cannot be easily traced or explained.
Thus, the nature of work is changing dramatically, and while many observers predict that the
combination of human and machine intelligence will always be the winning one
31
, we have yet to see
how the “augmented professionals” will carry out their work, and with what further implications for
the workplace, the organization and institutions.
Organizational implications
The introduction of AI is associated with significant changes in how organizations are managed.
Changing authority arrangements: Unavoidably, as we have discussed above, expertise is redefined
and the knowledge and skills of technology practitioners such as machine learning experts, data
scientists or data analysts become increasingly valued in the workplace. This can lead to restructured
authority arrangements across all levels of hierarchy. On a tactical level, technology practitioners will
28
For one prominent prediction, see Carl Benedikt Frey and Michael Osborne (2013), The future of employment:
how susceptible are jobs to computerization, Oxford Martin School working paper,
https://www.oxfordmartin.ox.ac.uk/downloads/academic/future-of-employment.pdf
29
Davenport, T. and Dreyer K. (2018), “AI will change radiology, but it won’t replace radiologists”, Harvard
Business Review, March 27. https://hbr.org/2018/03/ai-will-change-radiology-but-it-wont-replace-radiologists
30
Lebovitz, S. Lifshitz-Assaf, H. and Levina, N. 2020, To Incorporate or Not to Incorporate AI for Critical Judgments:
The Importance of Ambiguity in Professionals’ Judgment Process (January 15, 2020). NYU Stern School of
Business, Available at SSRN: https://ssrn.com/abstract=3480593
31
Brynjolfsson, E., & McAfee, A. (2014). The second machine age: Work, progress, and prosperity in a time of
brilliant technologies. WW Norton & Company.

gain authority and control over the work design and decision-making procedures, given that they have
the ability to prescribe how the introduced AI systems will affect the operations and work. But even
on a more strategic level, new roles get to join the board, triggering questions towards the established
regime: e.g. where does the jurisdiction of the CIO end and where does the jurisdiction of the CDO
start when it comes to planning a major digital transformation with the implementation of AI
technology?
Changing coordination: The use of AI to manage work algorithmically leads to fundamental changes
in organizational design and coordination. Work tasks are redefined so that they can be broken down
into smaller sub-tasks and are then algorithmically assigned to workers of digital labor platforms such
as UpWork or Amazon MTurk
32
. Machine learning algorithms can be used to coordinate more on a
proactive basis, by analyzing historical data to predict the needs in skills and expertise for future
projects. Furthermore, practitioners and managers need to collaborate with new experts who enter
the workplace with expertise in data processing, algorithm development, data visualisation, etc.
Collaboration amongst people with different types of expertise can make work coordination more
challenging. Changing coordination results in a substantially different execution of an organization’s
operations and services
Changing valuation schemes: The way in which performance is evaluated is also changing substantially,
as employees are assessed by machine learning algorithms most often without knowing what variables
are included in the inherent model, or the extent to which a specific variable contributes to the
production of a specific outcome. Even the quality check of products becomes an automated task itself
with the use of robots
33
. Those fundamental changes in the values that matter in the organization
substantially impact how firms manage their employees but can also lead to counter performances
from the employees’ side. For example, delivery drivers hanging their phones on trees outside Whole
Foods in order to get more rides was a fascinating yet ironic demonstration of how people try to game
the system in their effort to maintain some control over their work.
34
Industrial transformations: AI technology is currently enabling significant digital transformations that
are not only redefining what an organization does, but even result in blurring industry’s boundaries. A
lot of traditional manufacturing organizations for example are taking advantage of machine learning
technology to transform their focus, from the production of goods to the provision of services. GE’s
digital transformation effort is a popular example of such an attempt, with AI being the driving force
behind their predictive maintenance services. Who are the new competitors of such a digitally
transformed organization? How should it be regulated? How do the relationships with their customers
change? Who are their new partners? A lot of new questions arise that need to be addressed.
Future Opportunities
As companies continue to use AI, they will explore a variety of different directions. We suggest
several of them below.
Management and governance mechanisms—Leading companies using AI already have management
and governance mechanisms in place. We’ve already mentioned those related to deployment. In
addition, organizations in various states of adoption, have put in place a wide range of internal
32
Faraj, S., Pachidi, S., and Sayegh, K. 2018. "Working and Organizing in the Age of the Learning Algorithm,"
Information and Organization (28:1), pp. 62-70.
33
Mahdawi A. (2019). The Domino’s ‘pizza checker’ is just the beginning - workplace surveillance is coming for
you. The Guardian, October 15 2019.
34
Soper S. (2020). Amazon Drivers Are Hanging Smartphones in Trees to Get More Work. Bloomberg,
September 1, 2020.

organizational structures and roles to manage and govern AI projects. A survey
35
suggested that the
governance mechanisms used to manage AI projects include appointing AI champions, which 45% of
respondents said their firms had already done; creating an AI center of excellence
36
, which 37% had
done; and developing a comprehensive strategy for AI
37
, which 37% had also adopted.
Democratization of data science and AI—Tools like automated machine learning
38
can structure and
automate the workflow of creating and implementing a machine learning model. These can be
employed to improve the productivity of professional data scientists, or to enable less highly educated
“citizen data scientists” to complete data science and AI projects. Several startups and large cloud
vendors have made such capabilities available, and it seems likely that the democratization of data
science and AI development, the notion that anyone, with little to no expertise, can do data science if
provided ample data and user-friendly analytics tools, will continue to advance.
Ongoing model improvement—Companies that are heavily committed to AI often find that they have
many models and algorithms in place, some of which in production processes and systems. Since their
business is dependent in part on the accuracy of these models, it’s important to monitor them for
“drift”—inaccuracy of predictions—and improve them over time. Vendors are developing tools to
support this process under the banner of “MLOps”—machine learning operations—and they are most
widely used in data- and analytics-dependent industries like financial services.
AI explainability and transparency—As outlined above, it is now widely known that AI models can be
biased against certain groups and individuals
39
. Some firms have established AI ethics organizations
40
or “algorithm review boards” to assess transparency issues. Complex models, such as those in deep
learning neural networks, may be impossible to interpret or explain
41
. Some vendors provide
“prediction explanations” that point out influential variables or features and their direction of
influence, but this isn’t yet possible for the most complex models. Many organizations and researchers
are now working on new approaches to explainability, but we are only in the early stages of addressing
the issue successfully.
Reduced requirements for data—Many AI models, particularly deep learning neural networks, require
large amounts of data to be trained effectively. A new deep learning-based natural language
generation model called GPT-3, for example, used billions of words to train the model and has 175
billion variables and parameters. Some researchers
42
have argued that the trend to such volumes of
data is unsustainable, and that new approaches to AI can use less data. This is another trend in the
early stages, however.
35
Davenport, T. H. (2018). The AI advantage: How to put the artificial intelligence revolution to work. MIT Press.
36
Davenport, T. and Dasgupta, S. (2019) “How to set up an AI centre of excellence,” Harvard Business Review,
January 16, https://hbr.org/2019/01/how-to-set-up-an-ai-center-of-excellence
37
Davenport, T. and Mahidhar, V. (2018) “What’s your cognitive strategy?” MIT Sloan Management Review,
Summer. https://sloanreview.mit.edu/article/whats-your-cognitive-strategy/
38
Sharma, M. (2020) “Navigating the new landscape of AI platforms,” Harvard Business Review, March 10.
https://hbr.org/2020/03/navigating-the-new-landscape-of-ai-platforms
39
Li, M. (2019) “Are your algorithms upholding your standards of fairness? Harvard Business Review, Nov. 5.
https://hbr.org/2019/11/are-your-algorithms-upholding-your-standards-of-fairness
40
Davenport, T., 2019. “What does an AI ethicist do?” MIT Sloan Management Review, June 24.
https://sloanreview.mit.edu/article/what-does-an-ai-ethicist-do/
41
Royal Society (2019). Explainable AI: the basics; policy briefing. https://royalsociety.org/-
/media/policy/projects/explainable-ai/AI-and-interpretability-policy-briefing.pdf
42
Wilson, J., Daugherty, P., and Davenport, C. (2019) “The future of AI will be about less data, not more.”
Harvard Business Review, Jan. 24. https://hbr.org/2019/01/the-future-of-ai-will-be-about-less-data-not-more

Special issue papers
This special issue started as a conversation between the guest senior editors and the editors in chief
of two journals: the MISQ Executive (MISQE) and the Journal of the Association of Information Systems
(JAIS) on the need to create concerted efforts to contribute to both IS theory and practice. This special
issue is the outcome of such dialogue, a dialogue started at the pre-ICIS special issue Workshop held
in Munich. We received over 50 extended abstracts, 30 submissions were selected for discussion and
received early feedback from the special issue editorial board and the participating senior editors from
both journals.
The special issue received a total of 50 submissions. About half of the submissions were sent out to
review after the initial screening and after three rounds, five articles were accepted for publication in
the MISQE special issue. The first four papers appear in the December 2020 issue; The last paper will
be published in the March 2021 issue.
Table 2 maps the contributions each paper makes to the special issue along with the type of AI
technology it covers. We then briefly discuss each of the papers and outline the challenges firms faced
while adopting AI technologies and guidelines to manage such challenges.
Paper
Authors
AI technology
Industry
Contribution
1
Zhang,
Nandhakumar,
Hummel and
Waardenburg
Machine Learning
Algorithmic AI
Legal
Covers challenges related to
developing machine learning
systems
2
Mayer, Strich,
and Fiedler
Machine learning
Algorithmic AI
Banking
Discusses intended and unintended
consequences of introducing an
autonomous AI system
3
Asatiani, Malo,
Nagbøl,
Penttinen,
Rinta-Kahila,
and Salovaara
Machine learning
Algorithmic AI
Government
Offers ways to address explainability
issues
4
Reis, Maier;
Mattke;
Creutzenberg,
and Weitzel
Machine Learning,
Natural Language
Programming
Healthcare
Explains physician’s resistance to an
AI virtual agent
5
Schuetzler,
Grimes,
Rosser, and
Giboney
Conversational AI
Multiple
examples
Offers guidelines to design
conversational AI systems
Table 2: Special issue papers focus and contribution
The first paper in the special issue, “Key challenges of developing machine learning AI systems for
knowledge intensive work” is by Zhewei Zhang, Joe Nandhakumar, Jochem Thomas Hummel, and
Lauren Waardenburg. The paper discusses how a machine learning AI for a legal practice firm
(LegalTechCo) was developed to help legal professionals make faster and better informed decisions.
The authors studied the development of the AI system at LegalTechCo over a couple of years. They
identified three challenges involved in developing machine learning systems. The challenges are
related to how to define ML problems, how to manage the training of ML models and how to evaluate

ML AI performance. The authors proposed three guidelines (and twelve recommendations) for
executives to address the various challenges. The guidelines include: 1) Co-formulate the appropriate
machine learning AI problems, 2) Develop machine learning AI through iterative refinement; 3) Go
beyond the numeric measurements and ask for clues.
The second paper, “Well-meant is not well done: unintended consequences of introducing AI” is
by Anne-Sophie Mayer, Franz Strich, and Marina Fiedler. This paper focuses on the unintended
consequences of introducing an autonomous AI system in the banking industry. It draws on a case
study from one of the largest banks in Germany (Main Finance)
43
. Main Finance confronted several
issues in the small loan segment including: (1) increased competition from new market participants
due to digitization, (2) mismatched personnel resources, (3) high default rates, and (4) a decline in
profitability. To address the issues faced, the firm introduced an AI system based on ML to make
decisions about who is qualified for a loan. The authors document the implementation of the AI system
and its consequences from the perspective of both front-line workers and senior management. While
the introduction of the AI system enhanced profitability and helped address the main challenges faced
with loan management; it also resulted in employees’ perceived loss of competence and reputation,
and unpredictability of decisions. From senior management’s perspective, the AI system resulted in
employees’ loss of critical thinking and expertise and in the misuse of the system. The authors offer
several guidelines to prevent related consequences: 1) Maintain employees’ abilities to reflect and
understand underlying processes; 2) Understand and guide the shift of employees’ roles; 3) Make the
AI system as transparent and explainable as possible; 4) Reconsider customer groups excluded from
the AI.
The third paper, “Implementing Black-Box Artificial Intelligence: Lessons from Tackling Explainability
Issues at the Danish Business Authority” is by Aleksandre Asatiani, Pekka Malo, Per Rådberg Nagbøl,
Esko Penttinen, Tapani Rinta-Kahila, and Antti Salovaara. The authors document ways used by The
Danish Business Authority (DBA)—an agency under Denmark’s Ministry of Industry, Business, and
Financial Affairs—to deal with challenges associated with explainability. Availability of large volumes
of data enabled DBA to pursue machine learning for such core tasks such as supporting companies’
legal compliance, checking annual reports for signs of fraud, and identifying companies early enough
in their route to distress that timely support can be given.
The organization has been able to implement AI responsibly and legally even though the inner workings
are not always entirely explainable. The authors build on a six-dimensional framework of an intelligent
agent to discuss explainability challenges at DBA: (1) the model, (2) goal, (3) training data, (4) input
data, (5) output data, (6) environment. They further offer guidelines for managers to address
explainability issues: 1) Use modular design to increase AI explainability, 2) Avoid online learning if
explainability is a priority, 3) Facilitate continuous open discussion between stakeholders.
The fourth paper, Resistance to AI: A case study exploring the implementation failure of cognitive
agents in healthcare is by Lea Reis; Christian Maier; Jens Mattke; Marcus Creutzenberg, Tim Weitzel.
The authors discuss a case of AI implementation failure in a German hospital. The hospital has decided
to integrate AI to improve their anamnesis-diagnosis-treatment-documentation process with the
intent of giving physicians more time to care for patients and reducing process costs. A virtual agent
based on machine learning and natural language processing was developed to support different
activities: 1) the cognitive agent engages with patients to perform anamnesis, collects data and
provides structured documentation; 2) the cognitive agent applies decision support algorithms to
43
A pseudomym

suggest a diagnosis based on the structured recorded data; and 3) the cognitive agent engages with
the physician to provide treatment options.
However, after nine months of developing the use case and the test version and six months of
technological testing, the project team realized that the hospital’s physicians do not want to use the
system. While the physicians acknowledge that complementary knowledge supporting the diagnosis
decision is valuable to themselves and the patients, they refuse to approve the project. The team
decided to postpone the project indefinitely until they can better understand the reasons for the
physicians’ rejection and what steps to take to ensure future project success. The authors document
the reasons behind the physician’s rejection of the cognitive agent and offer recommendations to
address them.
The fifth paper, Your Agent is Ready: Guidance for Designing Conversational agents, is by Ryan
Schuetzler, Mark Grimes, Holly Rosser, and Justin Giboney. The paper focuses on chatbot design.
Chatbots are used by organizations to improve business processes, automate routine interactions, or
provide an automated social touchpoint for customers. The authors build on their experience with
chatbot design and use examples of chatbot across industries to offer a decision guide about when
and how chatbots should be deployed. The framework presented in the paper provides questions and
considerations that should be discussed early in the bot development process. And offers a number of
implicit signals bots can use to create natural, humanlike conversations.
Conclusion
The five papers selected for this special issue along with this editorial provide a variety of examples of
AI applications across industries, challenges and implications for organizations. Table 1 summarizes
the AI technology and industry covered by each paper, as well as the main contribution each paper
makes. We then briefly discuss each of the papers and outline the challenges firms faced while
adopting AI technologies and the recommendations offered to manage such challenges.
As AI technology is still maturing, awareness regarding the new management challenges it poses and
the implications it raises for the workplace and the organization are still emerging. But the most
common effect is likely to be on how work is conducted in the future. Thus, companies need to begin
work now on developing AI applications that create economic value and that lead to new ways of
orchestrating work by humans and machines. And leaders will have to understand how AI will impact
their workforce, then get them prepared : upskill some workers to do existing jobs with AI, and retrain
and hire others for the new roles that AI will demand.
About the Special Issue Guest Editors
Hind Benbya
Hind Benbya is Professor and Head of IS and Business Analytics at Deakin University and Visiting Policy
Fellow at the Oxford Internet Institute at the University of Oxford. Hind’s research expertise includes
digital innovation, IT-enabled transformation and Artificial Intelligence. Her work has appeared in MIS
Quarterly, the Journal of Management Information Systems, MIT Sloan Management Review, MISQ
Executive, and Decision Support Systems, among others. She is currently senior editor of the MISQ
Executive, guest senior editor for the Journal of the Association of Information Systems and a member
of the editorial board of the Journal of Strategic Information Systems. Hind has been a visiting professor
at Cambridge Judge Business School, UCLA Anderson School, and the London School of Economics. She
received several best paper awards, regularly works with leading firms in Europe, UK and the US and
presents her research at premier academic and practitioner venues.

Thomas H. Davenport
Tom Davenport is the President's Distinguished Professor of Information Technology and Management
at Babson College, Visiting Professor at the Oxford Saïd Business School, Fellow at the MIT Initiative on
the Digital Economy, and Senior Advisor to Deloitte Analytics. He teaches analytics/big data in
executive programs at Babson, Harvard Business School and School of Public Health, and MIT Sloan
School.
Davenport pioneered the concept of competing on analytics with his best-selling 2006 Harvard
Business Review article and 2007 book. His most recent book is The AI Advantage: How to Put the
Artificial Intelligence Revolution to Work. He wrote or edited nineteen other books and over 200
articles for Harvard Business Review, MIT Sloan Management Review, The Financial Times, and many
other publications. He is a regular contributor to the Wall Street Journal and Forbes. He has been
named one of the top 25 consultants by Consulting News, one of the 100 most influential people in
the IT industry by Ziff-Davis, and one of the world's top fifty business school professors by Fortune.
Stella Pachidi
Dr. Stella Pachidi is a Lecturer in Information Systems at Judge Business School, University of
Cambridge. Her research interests lie in the intersection of technology, work and organizing. Currently,
her research projects include the introduction of artificial intelligence technologies in organizations,
managing challenges in the workplace during digital transformation, and practices of knowledge
collaboration across boundaries. She holds a PhD in Business Administration from VU University
Amsterdam, a MSc in Business Informatics from Utrecht University, and a MSc in Electrical and
Computer Engineering from National Technical University of Athens. Dr. Pachidi has articles in
information systems and organization journals and books including Organization Science, Information
and Organization, and Computers in Human Behavior. She has presented her work in various major
conferences in the fields of technology and organizations including the Academy of Management
Meeting, the International Conference on Information Systems, the European Group for
Organizational Studies Colloquium, the Process Symposium and other.