PreprintPDF Available

Artificial Intelligence and Management: The Automation-Augmentation Paradox



Taking three recent business books on artificial intelligence (AI) as a starting point, we explore the automation and augmentation concepts in the management domain. Whereas automation implies that machines take over a human task, augmentation means that humans collaborate closely with machines to perform a task. Taking a normative stance, the three books advise organizations to prioritize augmentation, which they relate to superior performance. Using a more comprehensive paradox theory perspective, we argue that, in the management domain, augmentation cannot be neatly separated from automation. These dual AI applications are interdependent across time and space, creating a paradoxical tension. Over-emphasizing either augmentation or automation fuels reinforcing cycles with negative organizational and societal outcomes. However, if organizations adopt a broader perspective comprising both automation and augmentation, they could deal with the tension and achieve complementarities that benefit business and society. Drawing on our insights, we conclude that management scholars need to be involved in research on the use of AI in organizations. We also argue that a substantial change is required in how AI research is currently conducted in order to develop meaningful theory and to provide practice with sound advice.
Review Essay
Sebastian Raisch
University of Geneva
Sebastian Krakowski
University of Geneva
Stockholm School of Economics
Pre-print version
Article accepted for publication in the Academy of Management Review
The authors wish to thank associate editor Subramanian Rangan and two anonymous reviewers
for their guidance on developing the paper. Furthermore, they are grateful to Tina Ambos, Jean
Bartunek, Jonathan Schad, Katherine Tatarinov, Xena Welch Guerra, Udo Zander, and the
participants of research seminars at HEC Lausanne, Nova School of Business and Economics,
and the Geneva School of Economics and Management for their valuable comments on earlier
versions of the paper. This research received partial support from the Swiss National Science
Foundation [SNSF grants 181364 and 185164] and the European Union [EU Grant 856688].
Taking three recent business books on artificial intelligence (AI) as a starting point, we
explore the automation and augmentation concepts in the management domain. Whereas
automation implies that machines take over a human task, augmentation means that humans
collaborate closely with machines to perform a task. Taking a normative stance, the three books
advise organizations to prioritize augmentation, which they relate to superior performance.
Using a more comprehensive paradox theory perspective, we argue that, in the management
domain, augmentation cannot be neatly separated from automation. These dual AI applications
are interdependent across time and space, creating a paradoxical tension. Over-emphasizing
either augmentation or automation fuels reinforcing cycles with negative organizational and
societal outcomes. However, if organizations adopt a broader perspective comprising both
automation and augmentation, they could deal with the tension and achieve complementarities
that benefit business and society. Drawing on our insights, we conclude that management
scholars need to be involved in research on the use of AI in organizations. We also argue that
a substantial change is required in how AI research is currently conducted in order to develop
meaningful theory and to provide practice with sound advice.
Keywords: artificial intelligence, automation, augmentation, paradox theory, business and
The rise of powerful AI will be either the best or the worst thing ever
to happen to humanity. We do not yet know which.
Stephen Hawking, Theoretical physicist
What all of us have to do is to make sure we are using AI in a way
that is for the benefit of humanity, not to the detriment of humanity.
Tim Cook, CEO of Apple
Artificial intelligence (AI) refers to machines performing cognitive functions usually
associated with human minds, such as learning, interacting, and problem solving (Nilsson,
1971). Organizations have long used AI-based solutions to automate routine tasks in operations
and logistics. Recent advances in computational power, the exponential increase in data, and
new machine-learning techniques now allow organizations to also use AI-based solutions for
managerial tasks (Brynjolfsson & McAfee, 2017). For example, AI-based solutions now play
important roles in Unilever’s talent acquisition process (Marr, 2018), in Netflix’s decisions
regarding movie plots, directors, and actors (Westcott Grant, 2018), and in Pfizer’s drug
discovery and development activities (Fleming, 2018).
In the 1950s, pioneering research predicted that AI would become essential for
management (Newell, Shaw, & Simon, 1959; Newell & Simon, 1956). However, initial
technological progress was slow and the discussion of AI in management was “effectively
liquidated” in the 1960s (Cariani, 2010: 89). Scholars subsequently adopted a contingency
view: The routine operational tasks that machines could handle were separated from the
complex managerial tasks reserved for humans. Consequently, AI was researched in computer
science and operations research, whereas organization and management studies focused on
humans (Rahwan et al., 2019; Simon, 1987). Management scholars have therefore provided
very little insight into AI during the last two decades (Kellogg, Valentine, & Christin, in press;
Lindebaum, Vesa, & den Hond, in press). Nonetheless, these scholars’ understanding will be
required, because AI is becoming increasingly pervasive in managerial contexts.
In this review essay, we strive to reposition AI at the crux of the management debate.
Three highly influential business books on AI (Brynjolfsson & McAfee, 2014; Daugherty &
Wilson, 2018; Davenport & Kirby, 2016) serve as a source of inspiration to challenge our
thinking and spark new ideas in the management field (Bartunek & Ragins, 2015).
The three books develop a common AI narrative for practicing managers. The authors
distinguish two broad AI applications in organizations: automation and augmentation. Whereas
automation implies that machines take over a human task, augmentation means that humans
collaborate closely with machines to perform a task. Taking a normative stance, the authors
accentuate the benefits of augmentation while taking a more negative viewpoint on automation.
Their combined advice is that organizations should prioritize augmentation, which they relate
to superior performance. In addition, the two more recent books (Daugherty & Wilson, 2018;
Davenport & Kirby, 2016) provide managers with ample advice on how to develop and
implement such an augmentation strategy.
Assuming a more encompassing paradox theory perspective (Schad, Lewis, Raisch, &
Smith, 2016; Smith & Lewis, 2011), we argue that augmentation cannot be neatly separated
from automation in the management domain. These dual AI applications are interdependent
across time and space, which creates a paradoxical tension. Over-emphasizing either
augmentation or automation fuels reinforcing cycles that not only harm an organization’s
performance, but also have negative societal implications. However, organizations adopting a
broader perspective comprising both automation and augmentation are not only able to deal
with the tension, but also achieve complementarities that benefit business and society.
We conclude by discussing our insights’ implications for organization and management
research. The emergence of AI-based solutions and humans’ increasing interactions with them
creates a new managerial tension that requires research attention. Management scholars should
therefore play a more active role in the AI debate by reviewing prescriptions for managerial
practice and developing more comprehensive perspectives. They could do so by changing the
ways they currently conduct research in order to accurately analyze and describe AI’s
implications for managerial practice.
We started with a review of three recent business books on the use of AI in organizations.
While there are many other books on this topic, we selected the following three, which have
been widely influential in managerial practice, filling the void that the lack of scholarly
research caused. The New York Times bestseller The Second Machine Age by MIT Professors
Erik Brynjolfsson and Andrew McAfee (Brynjolfsson & McAfee, 2014) was called “the most
influential recent business book” in a memorandum that Harvard Business School’s Dean sent
to the senior faculty (The Economist, 2017). The much-debated Only Humans Need Apply
(Press, 2016) is the latest book on AI by Babson Professor Thomas H. Davenport and Harvard
Business Review contributing editor Julia Kirby (Davenport & Kirby, 2016). Finally, the
recently published Human + Machine by Accenture leaders Paul R. Daugherty and H. James
Wilson (Daugherty & Wilson, 2018) has had an immediate impact on both academia and
practice (Wladawsky-Berger, 2018).
Collectively, the three books suggest that we are on the cusp of a major transformation
in business, comparable to the industrial revolution in scope and impact. During this “first
machine age,” which started with the invention of the steam machine in the 18th century,
mechanical machines enabled mass production by taking over manual labor tasks at scale.
Today, we face an analogous inflection point of unprecedented progress in digital technology,
taking us toward the “second machine age” (Brynjolfsson & McAfee, 2014: 7). Instead of
performing mechanical work, machines now take on cognitive work, traditionally an
exclusively human domain. However, machines still have many limitations, which means we
are entering an era in which the human-machine relationship is no longer dichotomous, but
evolving into a machine “augmentation” of human capabilities. Rather than being adversaries,
humans and machines should combine their complementary strengths, enabling mutual
learning and multiplying their capabilities. Instead of fearing automation and its effects on the
labor market, managers should acknowledge that AI has the potential to augment, rather than
replace, humans in managerial tasks (Davenport & Kirby, 2016: 30–31).
Building on this analysis, the three books advise organizations to focus on augmentation
rather than on automation. The two more recent books explicitly relate such an augmentation
strategy to superior firm performance. For example, Daugherty and Wilson (2018: 214)
conclude that companies using “AI to augment their human talent (…) achieve step gains in
performance, propelling them to the forefront of their industries.” Conversely, companies
focusing on automation may “see some performance benefits, but those improvements will
eventually stall” (Daugherty & Wilson, 2018: 214). Similarly, Davenport and Kirby (2016:
214) predict that “a company whose strategy all along has emphasized augmentation, not
automation (…) will win big. Consequently, Davenport and Kirby advise companies to
prioritize augmentation (“don’t automate, augment”) (2016: 59), which they hail as “the only
path to sustainable competitive advantage” (2016: 204).
The two more recent books also provide managers with ample advice on how to develop
and implement such an augmentation strategy in their organizations. Davenport and Kirby
(2016: 89) describe five strategies for “post-automation” human work, all involving some form
of augmentation. In addition, they provide a seven-step process for planning and developing
an augmentation strategy (Davenport & Kirby, 2016: 201). Daugherty and Wilson (2018:
105ff) describe a range of new jobs that organizations could create and in which managers
complement machines and machines augment managers. The authors further describe how
augmentation could be implemented across domains, ranging from sales, marketing, and
customer service to research and development (Daugherty & Wilson, 2018: 67ff).
Consistent with the books’ recommendations, companies have started adopting an
augmentation strategy. For example, CEO Satya Nadella has announced that Microsoft will
“build intelligence that augments human abilities and experiences. Ultimately, it’s not going to
be about human vs. machine” (Nadella, 2016). Similarly, in the preamble to its AI guidelines,
Deutsche Telekom states that “AI is intended to extend and complement human abilities rather
than lessen or restrict them” (Deutsche Telekom, 2018). At IBM, the corporate principles
declare that the purpose of AI and cognitive systems developed and applied by the IBM
company is to augment human intelligence(IBM Think Blog, 2017). In her speech at the
World Economic Forum, IBM’s President and CEO Ginni Rometty suggested replacing the
term “artificial intelligence” with “augmented intelligence” (La Roche, 2017).
Taking the three books as a starting point, we use a paradox theory perspective (Schad et
al., 2016; Smith & Lewis, 2011) to explore organizations’ use of AI further. A paradox lens
allows us to elevate the level of analysis to study both automation and augmentation, which
reveals a paradoxical tension between these dual AI applications in management. Following
the Smith and Lewis (2011) paradox framework, we will analyze the paradoxical tension, the
management strategies to address it, and their outcomes.
Paradoxical Tension
The three books describe the relationship between automation and augmentation as a
trade-off decision: Organizations attempting to use AI have the choice of either automating the
task or using an augmentation approach. If they opt for automation, humans hand over the task
to a machine with little or no further involvement. The objective is to keep humans out of the
equation to allow more comprehensive, rational, and efficient processing (Davenport & Kirby,
2016: 21). In contrast, augmentation implies continued close interaction between humans and
machines. This approach allows for complementing a machine’s abilities with humans’ unique
capabilities, such as their intuition and common-sense reasoning (Daugherty & Wilson, 2018:
191f). The nature of the task determines whether organizations opt for one or the other
approach. Relatively routine and well-structured tasks can be automated, while more complex
and ambiguous tasks cannot, but can be addressed through augmentation (Brynjolfsson &
McAfee, 2014: 138ff; Daugherty & Wilson, 2018: 107ff; Davenport & Kirby, 2016: 34ff).
The arguments that the books provide are hard to refute, but their perspective is largely
limited to a given task at a specific point in time. Paradox theory, however, warns that such a
narrow trade-off perspective does not represent reality adequately (Smith & Lewis, 2011). A
paradox lens can help increase the scale or level of analysis for a more systemic perspective
(Schad & Bansal, 2018), which allows organizations to not only perceive the contradictions,
but also the interdependencies between automation and augmentation. A more comprehensive
paradox perspective (both/and) then replaces the traditional trade-off perspective (either/or).
The essence of paradox is that the dual elements are both contradictory and interdependent
forming a persistent tension (Schad et al., 2016).
Automation and augmentation are contradictory, because organizations choose either one
or the other approach to address a given task at a specific point in time. This choice creates a
tension, since these AI approaches rely on competing logics with different organizational
demands. For example, Lindebaum et al. (in press) maintain that automation instills a logic of
formal rationality in organizations that conflicts with the logic of substantive rationality, or the
human capacity for value-rational reflection, whereas augmentation preserves this substantive
rationality. The tension is further reinforced, because some organizational actors prefer
augmentation (e.g., managers at risk of losing their jobs to automation), while others prioritize
automation (e.g., owners interested in efficiencies) (Davenport & Kirby, 2016: 61).
While these contradictions are real, they only reveal a partial picture. If we increase our
analysis’s temporal scale (from one point in time to its evolution over time) and spatial scale
(from one to multiple tasks), we comprehend that, in the management domain, the two AI
applications are not only contradictory, but also interdependent.
Increasing the temporal scale. Taking a process view of paradox reveals a cyclical
relationship between opposing forces (Putnam, Fairhurst, & Banghart, 2016; Raisch, Hargrave,
& van de Ven, 2018). Engagement with one side of the tension may set the stage or even create
the conditions necessary for the other’s existence; in addition, over time there is often a mutual
influence between the opposing forces, with swings from one side to the other (Poole & van
de Ven, 1989). Elevating the temporal scale from one point in time to the process over time
allows for exploring this cyclical relationship between automation and augmentation.
As the books suggest, the process of using AI for a managerial task starts with a choice
between automation and augmentation. Organizations addressing a well-structured routine
task, such as completing invoices or expense claims, could opt for automation. They could do
so by drawing on codified domain expertise to program rules into the system in the form of
algorithms specifying the relationships between the conditions (“if”) and the consequences
(“then”) (Gillespie, 2014).1 Such rule-based automation requires an explicitly stated domain
model, which optimizes the chosen utility function (Russell & Norvig, 2009).2 With clear rules
in place, managers can relinquish the task to a machine.
However, most managerial tasks are more complex, and the rules and models are
therefore not fully known or readily available. In such cases, rule-based automation is
impossible, but managers could use an augmentation approach to explore the problem further
(Holzinger, 2016). This choice allows managers to remain involved and to collaborate closely
with machines on these tasks. It is a common misconception that this augmentation process
can be delegated to the IT department or external solution providers. While rule-based
automation allows such delegation, because the rules can be explicitly formulated, codified,
and passed on to data scientists, complex tasks’ augmented learning relies on domain experts’
tacit knowledge, which cannot be easily codified (Brynjolfsson & Mitchell, 2017). Data
scientists can provide technical support, but domain experts need to stay “in the loop” in
augmented learning (Holzinger, 2016: 119).
Augmentation is therefore a co-evolutionary process during which humans learn from
machines and machines learn from humans (Amershi, Cakmak, Knox, & Kulesza, 2014;
Rahwan et al., 2019). In this iterative process, managers and machines interact to learn new
rules or create models and improve them over time. The type and extent of human involvement
vary with the specific machine-learning solution (Russell & Norvig, 2009).3 Human domain
expertise is the starting point for supervised learning. Managers provide a machine with a set
of labeled training data specifying the inputs (or features) and the corresponding outputs. The
machine analyzes the training data and generates rules and/or models. In contrast, unsupervised
learning allows managers to induce patterns, of which they were not previously aware, directly
from the unlabeled data (Jordan & Mitchell, 2015).
In both applications, managers then use their domain expertise to evaluate, select, and
complement machine outputs. Spurious correlations or other statistical biases need to be
weeded out. For example, machines generally learn from large, noisy datasets containing
random errors. Overfitting is a key risk in this context, which means that a machine may learn
a complete model that also explains the errors, consequently failing to generalize appropriately
beyond its training data (Fan, Han, & Liu, 2014). The experts’ revision of the learned
knowledge is therefore an important part of the augmented learning process (Fails & Olsen,
2003).4 In each iteration, managers assess the current model’s quality, subsequently deciding
on how to proceed (Langley & Simon, 1995). The resulting tight coupling between humans
and machines, with the two influencing one another, makes it increasingly difficult, or even
impossible, to decouple their influence on the resulting model (Amershi et al., 2014).
Over time, this close collaboration with machines sometimes allows managers to identify
rules and/or models that either optimize the utility function or come sufficiently close to an
optimal solution to be practically useful.5 If these models are sufficiently robust, they can
subsequently be used to automate a task. Managers are taken “out of the loop,” which allows
them to focus on more demanding and valuable tasks. Augmented learning thus aims to provide
increasing levels of automation, replacing time-consuming human activity with automated
processes that improve accuracy, efficiency, and/or effectiveness (Langley & Simon, 1995).
Consequently, augmentation may enable a transition to automation over time.
To provide illustrations of such transitions from automation to augmentation, we briefly
discuss two examples from managerial practice. Organizations are increasingly employing AI-
based solutions in human resource (HR) management to acquire talent (Stephan, Brown, &
Erickson, 2017). For example, JP Morgan Chase chose an augmentation approach to assess
candidates. A team of experienced HR managers worked closely with an AI-based solution to
identify reliable, firm-specific predictors of candidates’ future job performance. It took a full
year of intensive interaction between the human experts and the AI-based solution to remove
statistically biased, or socially vexed predictors, and make the system robust. After the initial
augmentation stage, JP Morgan Chase decided to automate the candidate assessment task on
the basis of the identified criteria. By removing humans from this activity, the bank intends to
increase the candidate assessment’s fairness and consistency, while also making the process
faster and more efficient (Riley, 2018).
Product innovation is another key domain of AI application in management (Daugherty
& Wilson, 2018: 67f). For example, Symrise, a major global fragrance company, adopted an
augmentation approach to generate ideas. An AI-based solution helped the company’s master
perfumers identify correlations between specific customer demographics and different
combinations of ingredients based on the company’s database of 1.7 million fragrances.
Subsequently, Symrise’s master perfumers used their expertise to confirm or reject the possible
connections, create additional ones, and refine them further. After two years of close interaction
between the master perfumers and the machine, the resulting model was considered sufficiently
robust to automate the idea generation task. Based on a customer’s requirements, the AI-based
system now searches for possible new fragrance formulas far more rapidly and
comprehensively than humans can, which has helped increase these formulas’ novelty, while
simultaneously reducing the search cost and time significantly (Bergstein, 2019).
As the examples illustrate, organizations may initially choose augmentation to address a
complex task, but this advanced interaction between managers and AI-based solutions helps
them expand their understanding of the task over time, which sometimes allows subsequent
automation. While such a transition relaxes the tension temporarily, it resurfaces when
conditions change over time (Smith & Lewis, 2011). For example, digitalization is likely to
significantly alter the skills that JP Chase Morgan’s future talents need to be successful.
Bankers will need advanced data-science skills, which did not play a role in the extant
employee data. Such substantial changes therefore make the automated solutions function less
effectively (Davenport & Kirby, 2016: 72). Organizations should therefore, at least
temporarily, return to augmentation, which allows human and machines to jointly work through
the changing situation and adjust their models accordingly.
We conclude that the two AI applications in management are not only contradictory, but
also interdependent. Organizations may opt for one or the other application at a given point in
time, which softens the underlying tension temporarily, but fails to resolve it. Eventually,
organizations will face the same choice again, demonstrating the two applications’
interdependent nature and cyclical relationship.
Increasing the spatial scale. Paradox theory explores tensions not only over time, but
also across space. Paradoxical tensions are nested and interwoven across multiple levels of
analysis (Andriopoulos & Lewis, 2009). Addressing a tension at one level of analysis may
therefore simply reproduce the tension on a different level (Smith & Lewis, 2011). Elevating
the spatial scale from one task to multiple tasks allows us to explore automation and
augmentation’s nested interdependence across levels of analysis.
Focusing our attention on the use of either one (i.e., automation) or the other (i.e.,
augmentation) solution for a specific task sets artificial boundaries, fosters distinctions, and
fuels opposites (Smith & Tracey, 2016). However, in practice, managerial tasks rarely occur
in isolation, but are generally embedded in a managerial process. There are interdependencies
between the various tasks constituting this process. These interdependencies cause managerial
interventions in one task to have ripple effects throughout the process (Lüscher & Lewis, 2008).
If organizations automate a task hitherto reserved for humans, this change could affect other,
closely related human tasks, and lead managers to start interacting with machines. Such
interactions are often iterative, resulting in the augmentation of adjacent tasks.
For example, at Symrise, the automation of the “idea generation” task also affected the
preceding “objective setting” and the succeeding “idea selection” tasks in the product
innovation process. Consequently, in the initial, objective-setting stage, the company’s master
perfumers must now enter customers’ objectives and constraints into the AI-based system to
allow the automated generation of fragrance formulas matching these requirements in the
subsequent idea-generation stage (Goodwin et al., 2017). This is often an iterative process, with
the master perfumers circling back to adjust the objectives and the constraints according to the
system outputs. In the later idea-selection stage, the master perfumers continue using their
human senses, expertise, and intuition to select one of the formulas that the machine proposed.
They subsequently use the AI-based solution to further refine their chosen formula (Bergstein,
2019). For example, the master perfumers employ the machine to experiment with different
dosages of the selected formula’s ingredients. This refinement process can include hundreds of
iterations between the machine and the master perfumers. This iterative process involving close
human-machine interaction has led to the augmentation of the idea selection task.6
As this example illustrates, a task’s automation can lead to human-machine interaction
in the preceding and/or the succeeding tasks in the managerial process. Automation in one task
“spills over,” enabling adjacent tasks’ augmentation. These spillovers are particularly fast in
AI systems, which often rely on distributed computing and cloud-based solutions that make
the knowledge gained from a given insight immediately accessible across the system (Benlian,
Kettinger, Sunyaev, & Winkler, 2018). The two AI applications in management are therefore
not only interdependent across time, but also across space. While automation and augmentation
are distinct activities operating in different temporal or spatial spheres, they are nevertheless
intertwined at a higher level of analysis. Viewed as a paradox, automation and augmentation
are no longer separate, but are mutually enabling and constituent of one another.
Persistence of the tension. Paradox refers to a tension between interdependent elements,
but this tension is only considered paradoxical if it persists over time (Schad et al., 2016). We
argue that the emerging co-existence of interdependent automated and augmented tasks will
persist in the management domain. Sometimes, highly visible advancements driven by
machine-learning applications are misinterpreted and extrapolated to imply that we are on the
threshold of advancing toward artificial general intelligence.7 However, there is widespread
agreement among computer scientists that we are actually far from machines wholly surpassing
human intelligence (Brynjolfsson & Mitchell, 2017; Walsh, 2017). Technical and social
limitations make the full automation of complex managerial processes impossible in the
foreseeable future. Managers will therefore remain involved in these processes and interact
with machines on a wide range of tasks.
A few of machines’ limitations are worth pointing out here: First, machines have no sense
of self or purpose, which means managers need to define their objectives (Braga & Logan,
2017). Setting objectives is closely related to taking responsibility for the associated tasks and
outcomes; consequently, while organizations can extend accountability to machines,
responsibility requires intentionality, which is an exclusively human ability (Floridi, 2008). In
turn, humans can only take responsibility if they retain some level of involvement with and
control over the relevant tasks. In our example of product innovation at Symrise, the perfumers
set the objectives, remain involved throughout the innovation process, and take responsibility
for its outcomes. The same is true of HR managers in the talent acquisition process.
Second, in respect of complex managerial tasks, machines can only provide a range of
options that all relax certain real-life constraints.8 Managers need to use their intuition and
common-sense judgment reconciling the machine output with reality − to make a final
decision about the most desirable option (Brynjolfsson & McAfee, 2014: 92). In our example
of talent acquisition at JP Morgan Chase, the AI-based solution enabled the candidate
assessment’s automation, but HR managers are still needed for the subsequent candidate
selection (Riley, 2018), because no model can cover this task’s full complexity. Machines
cannot fully capture ambiguous predictors, such as cultural fit or interpersonal relations, for
which there simply are no codified data available. The same applies to the product development
process at Symrise, where the master perfumers ultimately choose one of the machine’s
suggested fragrance formulas (Bergstein, 2019).
Third, machines are limited to the specific task for which they have been trained. They
cannot take on other tasks, since they do not possess the general intelligence to learn from their
experience in one domain to conduct tasks in other domains (Davenport & Kirby, 2016: 35).9
Managers therefore need to ensure contextualization beyond an automated task. For example,
HR managers still need to spend hours coordinating meetings to ensure that their hiring
decisions are aligned with the business strategy, and product developers need to continue
interacting with marketing departments to align their products with the business models.
Fourth, machines do not possess human senses, perceptions, emotions, and social skills
(Braga & Logan, 2017).10 For example, HR managers can use their emotional and social
intelligence to provide a “human touch,” or the advanced communication required to build true
relationships, entice talent to work for their firm, and convince others to support the decisions
made (Davenport & Kirby, 2016: 74). In the Symrise case, machines can neither smell nor fully
predict how humans will perceive new fragrances, or the emotions and memories they trigger.
Master perfumers have these skills and can also use them to subsequently tell a compelling
story about a fragrance and its meaning, which is important for its commercialization.
To conclude, the augmentation of a managerial task may enable its subsequent
automation. Such automation can, in turn, trigger further augmentation in closely related
managerial tasks. While these dynamics are likely to promote increasing augmentation and
automation, technological and social limitations prevent progress toward the full automation
of managerial tasks in organizations. This is particularly true of managerial task contexts
characterized by high degrees of ambiguity, complexity, and rare events, which limit
deterministic approaches’ applicability (Davis & Marcus, 2015). In such contexts, automation
and augmentation provide different, partly conflicting, but also complementary logics and
functionalities that organizations require. While the optimal balance between automation and
augmentation depends on contingencies, such as organizations’ AI expertise and the nature of
the environmental contexts they face, organizations will experience a persistent tension
between these interrelated applications of AI in management.
Management Strategies
Recent technological progress has made the AI tension salient for organizations.
Organizations facing such a salient tension tend to apply management strategies to address it.
According to paradox theory, these organizational responses fuel reinforcing cycles that can
be either negative or positive (Smith & Lewis, 2011). If organizations are unaware of a
tension’s paradoxical nature, they risk applying partial strategies, which cause vicious cycles
that escalate the tension. Conversely, organizations that accept a tension as paradoxical and
pay attention to its competing demands could enable virtuous cycles (Schad et al., 2016).
Vicious cycles. Organizations are likely to prioritize automation due to its promise of
short-term cost efficiencies (Davenport & Kirby, 2016: 204). This strategy forces
organizations’ competitors to also pursue automation in order to remain cost competitive.
Consequently, the whole industry may be “entering (…) in a race toward the zero-margin
reality of commoditized work” (Davenport & Kirby, 2016: 204). Over time, these
organizations lose the human skills required to alter their processes (Endsley & Kiris, 1995).
Human experts are either made redundant through automation, or they lose their specific skills
regarding the tasks they no longer pursue. Prior research has shown that automation can deskill
humans, make them complacent, and diffuse their sense of responsibility (Parasuraman &
Manzey, 2010; Skitka, Mosier, & Burdick, 2000). Ultimately, organizations become
entrenched in their automated processes, because automation is limited to specific tasks in well-
understood domains and imposes formal rules that narrow organizations’ choices and penalize
deviation (Lindebaum et al., in press). To conclude, while automation can free up resources for
potential search activities, it is also associated with short-term thinking, the loss of human
expertise, and lock-in effects that, together, fuel a reinforcing cycle, which makes it
increasingly difficult for organizations to implement such search activities.11
In contrast, organizations could follow the three books’ combined advice and focus on
augmentation. This AI application requires extensive resources to work through iterative cycles
of human-machine learning. Contrary to automation, augmentation demands continued human
involvement and experimentation (Amershi et al., 2014). Since emotions and other subjective
factors affect humans, augmentation is difficult or even impossible to replicate, which means
every augmentation initiative is a new learning effort (Holzinger, 2016). Owing to their
inherent complexity and uncertainty, augmentation efforts often fail (Amershi et al., 2014).
Furthermore, the continued human involvement implies that human biases persist, which
means augmentation outcomes are never fully consistent, reliable, or persistent (Huang, Hsu,
& Ku, 2012). To legitimize their large augmentation investments, organizations experiencing
failure may be tempted to reinforce their augmentation efforts further, which could escalate
their commitment (Sabherwal & Jeyaraj, 2015; Staw, 1981), with failure leading to continued
augmentation, in turn leading to continued failure.
To conclude, one-sided orientations toward either automation or augmentation cause
vicious cycles, because they neglect the dynamic interdependencies between AI’s dual
applications in management. Managers limiting their perspective to either automation or
augmentation risk developing partial and incomplete managerial solutions. While these
solutions may be appropriate within the strict boundaries that time and space impose, the use
of AI in management causes an organizational tension that persists across time and space.
Virtuous cycles. Paradox theory offers a more constructive response to tensions by
envisioning a virtuous cycle, with organizations overcoming their defensiveness to embrace
these tensions and viewing them as an opportunity to find synergies that accommodate and
transcend the opposing poles (Schad et al., 2016).
A first step toward enabling such a virtuous cycle is the acceptance of the tensions as
paradoxical (Smith & Lewis, 2011). While managers initially perceive automation and
augmentation as a trade-off, they may eventually recognize that they cannot simply choose
between these dual AI applications, because either choice intensifies the need for its opposite.
However, transitioning to a more encompassing paradox perspective requires cognitive and
behavioral complexity (Miron-Spektor, Ingram, Keller, Smith, & Lewis, 2018). Stimulating an
exchange between organizational actors with different perspectives, such as data scientists and
business managers, could develop more complex understandings of the phenomenon. Once
actors accept that automation and augmentation can and should co-exist, they can explore the
dynamic relationship between them mindfully, which could be part of their organization’s
vision or guiding principles regarding the use of AI in management.
While acceptance lays the groundwork for virtuous cycles, it has to be complemented
with a subsequent resolution strategy (Smith & Lewis, 2011). Resolution involves seeking
responses to paradoxical tensions through a combination of differentiation and integration
practices (Poole & van de Ven, 1989).
Differentiation allows organizations to recognize and appreciate automation and
augmentation’s distinctive benefits and leverage them separately. Organizations can
purposefully iterate between distinct automation and augmentation tasks, allowing long-term
engagement with both forces. For example, Symrise’s master perfumers iterate between
automation (i.e., when generating alternative fragrance formulas) and augmentation (i.e., when
selecting and refining the most promising formula). The use of automation allows exploration
beyond humans’ abilities by searching through the whole landscape of possible options. Their
cognitive limitations mean that humans’ search field is restricted, while machines do not face
such information-processing limitations (Davenport & Kirby, 2016: 17). Excluding humans at
this stage may help break path dependencies and promote greater novelty. Switching to
augmentation allows machine limitations to be overcome by subsequently using humans’ more
holistic and intuitive information processing to choose between options and contextualize
beyond the specific task at hand (Brynjolfsson & McAfee, 2014: 92).
While such differentiation allows for engaging in both automation and augmentation,
integration enables the finding of linkages that transcend the two poles (Smith & Lewis, 2011).
By switching, the machine’s independent output can be used to challenge human intuition and
judgment, with human feedback enabling further rounds of machine analysis (Hoc, 2001). At
these transition points, automation and augmentation become mutually enabling. The two AI
approaches’ juxtaposition stimulates learning and fosters adaptability, allowing the
combination of (machine) rationality and (human) intuition, which enables more
comprehensive information processing and better decisions (Calabretta, Gemser, & Wijnberg,
2017). Through integration, automation and augmentation jointly generate outcomes that
neither application can enable individually.
It is no easy feat to ensure such integration. As described above, the risks of organizations
over-emphasizing either automation or augmentation are real. Integration therefore requires
humans to retain overall responsibility for a managerial process. Prior studies have shown that
maintaining overall human responsibility not only reduces human bias (Larrick, 2004), but also
prevents human-machine collaboration biases (Skitka et al., 2000). As these studies show,
assigning the overall responsibility for processes to humans leads to increased vigilance and
verification behavior, the consideration of a wider range of inputs prior to making decisions,
and the use of greater cognitive complexity when processing such information. Consequently,
retaining human responsibility for managerial processes promotes integration, which
transcends automation and augmentation.
Paradox theory suggests that managing tensions through the dynamic strategies of
acceptance and resolution fosters sustainability (Smith & Lewis, 2011). By managing paradox,
organizations enable learning and creativity, promote flexibility and reliance, and unleash
human potential. However, paradox scholars also acknowledge that narrow organizational
attention to just one of the tensions’ poles can trigger unintended organizational and societal
consequences (Schad & Bansal, 2018). We therefore conclude our analysis of the automation-
augmentation paradox by assessing its organizational and societal outcomes.
Organizational outcomes. The three books we reviewed argue that organizations
benefit greatly from using AI. In particular, the authors emphasize augmentation’s potential to
increase productivity, improve service quality, and foster innovation. Moreover, they assume
that the combination of complementary human and machine skills will increase the quality,
speed, and extent of learning in organizations (Brynjolfsson & McAfee, 2014: 182; Daugherty
& Wilson, 2018: 106; Davenport & Kirby, 2016: 206). In contrast, we have argued that
focusing on either automation or augmentation can lead to reinforcing cycles that harm long-
term performance. We suggest that organizations benefit if they differentiate between and
integrate across automation and augmentation.
Differentiation allows organizations to benefit from both AI applications’ unique
benefits. Automation enables organizations to drive cost efficiencies, establish faster processes,
and ensure greater information-processing rationality and consistency. As described before,
augmentation provides complementary benefits arising from the mutual enhancing of human
and machine skills. The integration of automation and augmentation leads to additional benefits
that accrue from the synergies between these interdependent activities. Automation could free
up scarce resources for augmentation, which, in turn, could help identify the rules and/or
models that enable automation. Balancing automation and augmentation helps prevent the
escalating cycles that focusing on just one of these AI applications could cause. Furthermore,
the combination of automation and augmentation could enable new business models. AI is, for
example, the major driver behind the current trend toward personalized medicine, with
treatments being tailored to each patient’s specific biological profile (Fleming, 2018; Lichfield,
2018). While augmentation allows for identifying patterns in large volumes of patient data,
automation makes the design and manufacturing of tailored drugs economically viable.
These varied benefits suggest that automation and augmentation’s combination creates
complementary returns that lead to superior firm performance. Together, AI’s dual applications
in management provide organizations with a range of benefits that neither automation nor
augmentation can provide alone. However, realizing these benefits is contingent upon
organizations’ active management of the automation-augmentation paradox.
Societal outcomes. Tensions’ systemic nature is a central tenet of paradox theorizing
(Jarzabkowski, Bednarek, Chalkias, & Cacciatori, 2019; Smith & Lewis, 2011). Paradoxes are
embedded in open systems and their implications extend beyond a single organization’s
boundaries. Consequently, it is important to adopt a more systemic perspective of paradox,
which not only takes the organizational outcomes into consideration, but also tensions’ and
their management’s larger, system-wide or societal implications (Schad & Bansal, 2018).
While firms may gain profits from their use of AI in management, the three books to a
varying extent also point out that the larger societal implications are less certain (e.g.,
Brynjolfsson & McAfee, 2014: 171). There is a risk that organizations could take a narrow
perspective of either automation or augmentation, triggering unintended consequences that
affect society negatively. However, if organizations adopt a more comprehensive perspective,
the outcomes could be positive for both business and society. We explore these issues further
by focusing our attention on two societal outcomes discussed extensively in the three books:
AI’s labor market impact (e.g., Brynjolfsson & McAfee, 2014: 147ff), and its effects regarding
social equality and justice (e.g., Daugherty & Wilson, 2018: 129ff).
First, a one-sided focus on automation could cause extensive job losses and result in the
deskilling of managers who relinquish tasks to machines, which could lead to the further risks
of rising unemployment and social inequality (Brynjolfsson & McAfee, 2014: 171; see also
Autor, 2015). Conversely, one-sided augmentation is likely to cause another “digital divide”
(Norris, 2001), with social tensions arising between the few who currently have the capabilities
and resources for augmentation and those who do not (Brynjolfsson & McAfee, 2014: 134f;
see also Brynjolfsson & Mitchell, 2017).
Balancing automation and augmentation could, however, enable a virtuous cycle of
selective deskilling (i.e., humans offload tasks where their abilities are inferior to those of
machines) and strategic requalification (i.e., humans stay ahead of machines in their core
abilities), thereby gradually enhancing both human and machine capabilities. This virtuous
cycle could help organizations reduce the digital transition’s negative effects on their
employees and the labor market at large. Employees and managers whose basic skills are made
redundant by automation could be given the opportunity to gradually build higher-level
augmentation skills that remain in demand. This skill-enhancement cycle could also help
“rehumanize work” by gradually shifting the focus from repetitive and monotonous tasks to
more creative and fulfilling tasks (Daugherty & Wilson, 2018: 214).
A recent initiative at UBS’s investment-banking division illustrates this virtuous cycle.
UBS used an AI-based solution to automate the task of reading and executing client demands
for fund transfers, which previously took an investment banker 45 minutes per demand.
Simultaneously, the bank implemented another AI-based solution to augment the development
of trading strategies. The investment bankers use the time freed up by automation to collaborate
closely with the AI-based tool to explore new strategies for adaptive trading. Data scientists
provide the investment bankers with organizational support to develop their augmentation
skills. Consequently, the combined use of automation and augmentation allowed UBS to
exploit cost efficiencies while exploring new client solutions (Arnold & Noonan, 2017).
Second, the use of AI in management could also have implications for social equality
and fairness. Automation takes humans “out of the loop,” reducing human biases and, in turn,
promising greater equality and fairness. For example, using automation for credit approval
could reduce bankers’ bias that might previously have kept people from qualifying for credit
due to their ethnicity, gender, or postal code (Daugherty & Wilson, 2018: 167). Similarly,
automated candidate assessment based on pre-determined criteria and consistent machine
processing could help eliminate humans’ implicit biases in their hiring decisions.
However, real-world applications show that machine biases caused by noisy data,
statistical errors, and/or socially vexed predictors often lead to new, even more systematic
discrimination. Daugherty and Wilson (2018: 179) cite the example of an automated AI system
used to predict defendants’ future criminal behavior, which has been shown to be biased against
black defendants. Another example involves Amazon, which discontinued the use of an
automated AI-hiring tool found to discriminate against female applicants for technical jobs
(Dastin, 2018). In contrast, augmentation is likely to reduce such machine biases through
human backtesting and feedback. However, the intense interaction between managers and
machines increases the risk of human biases being carried over to machines. This problem is
particularly pernicious, since machines then confirm humans’ biased intuition, which makes
humans even less likely to question their preconceived positions (Huang et al., 2012).
The solution could be once more to combine differentiation and integration practices to
address the paradox. Differentiation allows for independent analyses with (i.e., augmentation)
and without (i.e., automation) human involvement. Integration ensures that machine outputs
are used to challenge humans, and human inputs to challenge machines (Hoc, 2001), allowing
for mutual learning (Panait & Luke, 2005) and debiasing (Larrick, 2004). Furthermore, humans
“in the loop” could explain the system’s outputs, rendering algorithmic decisions more
accountable and transparent (Binns et al., 2018).
For example, much like JP Morgan Chase, Unilever differentiates clearly between
automation (used for the initial candidate assessment) and augmentation (used for the final
selection) in their talent acquisition process. Integration is ensured by retaining the overall
human responsibility for the entire process. Unilever reports that the combination of
automation and augmentation has led to a 16 percent increase in new hires’ ethnic and gender
diversity, making it the “most diverse class to date.” At the same time, the company managed
to save 70 000 person-days of interviewing and assessing candidates, which resulted in annual
cost savings of GBP 1 million and a reduction of the time-to-hire by 90 percent (from an
average of four months to two weeks) (Feloni, 2017).
Inspired by three recent business books, we have explored the emergent use of AI-based
applications in organizations to automate and augment managerial tasks. The central argument
of this review essay is that automation and augmentation are not only separable and conflicting,
but in fact fundamentally interdependent. We suggest that the prevailing tradeoff perspective
is overdrawn; viewed as a paradox, augmentation is both the driver and outcome of automation,
and the two applications of AI develop and fold into one another across space and time. The
automation-augmentation model’s popularity stems largely from its clear boundaries and
simplicity. However, adopting this intuitively appealing model uncritically obscures how
automation and augmentation intertwine in managerial practice. This review essay therefore
sheds light on automation and augmentation’s complementarities and identifies opportunities
to transcend the paradoxical relationship between them.
In the following, we discuss our reflections’ implications for organization and
management research. We will argue that the ways in which scientific research on AI is
currently conducted need to change to accurately capture and analyze its organizational and
societal implications for managerial practice. Our discussion will be loosely structured along
the famous “5w & 1h” questions (who, how, what, why, where, and when).12
Our analysis of the automation-augmentation paradox suggests that the jury is out on
whether the use of AI in management will turn out to be a blessing or a curse. Scholars can still
make a difference by exploring the topic and informing practice regarding the ways forward.
However, this will require a change in who conducts research on AI. Currently, computer
scientists, roboticists, and engineers are the scholars most commonly studying AI. Their
primary objective is to automate as far as possible, because they often regard humans as “a
mere disturbance in the system that can and should be designed out” (Cummings, 2014: 62).
Computer scientists may be expert technologists, but they are not generally trained social or
behavioral scientists (Rahwan et al., 2019). Instead, they tend to use laboratory settings for
their research, which reduce the inherent variability that characterizes human behavior. These
settings permit the use of methodologies aimed at maximizing algorithmic performance, but
disregard the role of humans and the wider organizational and societal implications.
The three business books’ augmentation perspective reveals the importance of inducing
social scientists to participate in the debate. With respect to managerial tasks, humans will
remain “in the loop” and interact closely with machines. Management scholars are particularly
well-equipped to study these human-machine interactions in real-world settings, as well as to
explore their organizational and societal implications. It is therefore no wonder that the three
business books devote the bulk of their attention to augmentation. However, our analysis
suggests that management research limited to augmentation may be as biased as computer
science’s traditional focus on automation. Research on AI in management would therefore
benefit greatly from more interdisciplinary efforts. The technological and the social worlds are
merging (Orlikowski, 2007), which means that computer scientists and management scholars
no longer study separate phenomena. By juxtaposing and integrating their different
perspectives, theories, and methodologies, computer scientists and management scholars can
jointly create a foundation for meaningful research on the use of AI in management.
Such efforts also require a change in how research on AI is conducted in the management
domain. The limited work to date provides separate accounts with clear-cut contrasts between
augmentation and automation. On the one side, the research and practice doomsayers warn us
that automation will enslave humans, supervise and control them, and drive out every iota of
humanity (e.g., Bostrom, 2014; Ford, 2015). For example, in their recent AMR review essay,
Lindebaum et al. (in press) maintain that automation may lead to a technology-enabled
totalitarian system with formal and oppressive rules representing the end of human choice. On
the other side, there are technology utopians (e.g., Kurzweil, 2014; More & Vita-More, 2013),
including the authors of two of the reviewed books (Daugherty & Wilson, 2018; Davenport &
Kirby, 2016), who argue that humans will remain in control and use augmentation to create
huge benefits for organizations and society.13
The automation-augmentation paradox suggests that both perspectives are equally
biased. Automation and augmentation are not good or evil per se. Derrida (1967) argued that
humans tend to construct binary oppositions in their narratives, with a hierarchy that privileges
one side of the dichotomy while repressing the other. He warns that this approach is overly
simplistic; it makes us forget that one side of a dichotomy cannot exist without the other.
Similarly, researchers need to accept that automation and augmentation are interdependent AI
applications in management that cannot be neatly separated and designated as either good or
evil. These applications provide complementary functionalities that are both potentially useful
for organizations. The complex interaction between these varied AI applications over time
could have both positive and negative organizational and societal implications.
Research on AI in management therefore needs to complexify its theorizing by moving
from simple either/or perspectives to more encompassing both/and ones. Complexifying
theories is essential for understanding complex phenomena (Tsoukas, 2017), because “it takes
richness to grasp richness” (Weick, 2007: 16). More encompassing perspectives, such as
paradox theory (Schad et al., 2016) or systems theory (Sterman, 2000), offer a vantage point
from which researchers can observe the dynamic interplay between automation and
augmentation. Accepting and embracing this complexity allows for studying AI and its
managerial applications “in the wild.” This leads to a more comprehensive and, ultimately,
more rigorous and relevant discussion of AI’s organizational and societal implications.
By working through this complexity, it becomes apparent what research needs to be
conducted. At the most basic level, management scholars need to acknowledge that humans
are no longer the sole agents in management, although most theories to date focus exclusively
on human agency. Scholars need to overcome this human bias and integrate intelligent
machines into their theories. The use of AI for managerial tasks implies that machines are no
longer simple artifacts, but a new class of agents in organizations (Floridi & Sanders, 2004).
While machines have fundamental limitations, their actions nevertheless enjoy far-reaching
autonomy, because humans delegate knowledge tasks to these machine agents and allow them
act on their behalf (Rai, Constantinides, & Sarker, 2019).
Such automation leads to machine behavior that deviates significantly from the human
behavior that management theories traditionally describe. The bulk of current management
research relies on the behavioral assumptions of boundedly rational human actors, who – due
to their information-processing limits and cognitive biasesengage in satisficing rather than
maximizing behavior (Argote & Greve, 2007; Cyert & March, 1963). However, intelligent
machines used for automation do not have these limitations; they have practically unlimited
information-processing capacity and exhibit perfectly consistent behavior. Nonetheless, they
can introduce statistical biases and have other limitations that humans do not have (Elsbach &
Stigliani, 2019). These differences lead to entirely new machine behaviors (Rahwan et al.,
2019). We could, for example, speculate that because machines do not have humans’
limitations when searching, organizations using automation and augmentation could suffer less
from learning myopia and path dependencies (Levinthal & March, 1993). Management
scholars thus need to broaden their perspective to include human and machine agents and
explore their distinct behaviors in organizational contexts.
If we increase the complexity further, we have to acknowledge that these human and
machine agents do not simply co-exist in separate worlds (working on separate tasks), but are
interdependent (interacting on the same or closely related tasks). Augmentation therefore
implies close collaboration between humans and machines. Since automation and
augmentation are interdependent, this interaction spreads across organizations. When
addressing this human-machine interaction, management scholars need to first explore how
machines shape managerial behavior. For example, Lindebaum et al. (in press) describe how
autonomous algorithms can direct and constrain human behavior by imposing formal
rationality. This perspective resonates with Foucault’s (1977) concept of panoptic surveillance,
characterizing IT as a form of omnipresent architecture of control that creates, maintains, and
cements central norms of expected behavior (see also Lyon, 2003; Zuboff, 1988, 2019).
However, our broader paradox perspective of AI in management reveals that humans
also shape machine behavior. Managers define the objectives, set constraints, generate and
choose the training data, and provide machines with feedback. In machine-learning systems,
humans shape and reshape algorithms through their daily actions and interactions (Deng, Bao,
Kong, Ren, & Dai, 2017). In a management context, machines may influence human behavior,
but without setting a static norm of conduct or an unsurpassable rule (Cheney-Lippold, 2011).
Managers participate in writing and rewriting the rules that shape behavior. Management
research should thus explore both machines’ influence on human behavior and humans’
influence on machine behavior in the context of AI use in management.
If we increase the complexity even further, we finally see that humans and machines
influence one another in an iterative process. While machine algorithms shape human actions,
humans shape these algorithms through their actions, which create a recursive relationship
(Beer, 2017). Through augmentation, humans and machines become so closely intertwined that
they collectively exhibit entirely new, emergent behaviors, which neither show individually
(Amershi et al., 2014). The use of AI in management leads to hybrid organizational systems
manifesting collective behaviors. It may therefore be difficult, or even impossible, to
distinguish between humans and machines or their respective learning and actions.
While it is convenient, and sometimes helpful, to separate research studies analyzing how
machines influence managers and vice versa, studies examining hybrid organizational systems
comprising both managers and machines should provide the greatest benefit. Only such studies
can examine the feedback loops between human influence on machine behavior and machine
influence on human behavior, which cause the emergent behaviors that are otherwise
impossible to predict (Rahwan et al., 2019). Management scholars must provide further
insights into how such hybrid organizational systems function by exploring their interactive
behaviors. This systemic perspective also enables them to predict and/or explore the
automation-augmentation paradox’s systemic dynamics, which, as we have described before,
can include unintended consequences and escalating cycles. Management research therefore
has a crucial mandate to explore managers’ continued interactions with machines, as well as
the emergent behaviors and systemic outcomes they cause.
This discussion leads us to the heart of the fourth question, namely why research on AI
in management is essential. As argued above, the emergent use of AI in management leads to
iterative interactions between humans and machines. The resulting hybrid organizational
systems exhibit behaviors and produce organizational, as well as societal, effects that are
impossible to predict precisely and are often entirely unanticipated (O’Neil, 2016). No single
actor in these systems has full control over these outcomes. Consequently, it is difficult to
apportion accountability for outcomes to specific actors, which creates an “accountability gap”
(Mittelstadt, Allo, Taddeo, Wachter, & Floridi, 2016). There is widespread fear of and the first
empirical evidence that this lack of accountability can have harmful societal consequences
regarding equality (Autor, 2015), privacy (Acquisti, Brandimarte, & Loewenstein, 2015),
security (Brundage et al., 2018), and transparency (Castelvecchi, 2016).
Traditional managerial and organizational solutions may be inadequate for addressing
such systemic problems sufficiently (Schad & Bansal, 2018). Management research should
therefore contribute to the development of new organizational solutions that allow AI’s benefits
to be realized, while mitigating the associated negative side effects. In order to fully understand
the automation-augmentation paradox’s societal implications, scholars could adopt a relational
ontology which accepts that, in the digital age, human and machine agents are so closely
intertwined in hybrid collectives that their relations determine their actions (Floridi & Taddeo,
2016). Rather than focusing on individual actors, the interactions between these actors should
be the unit of analysis. Such a perspective enables a discussion of “distributed morality”
(Floridi, 2013), which relies on shared ethical norms developed through collaborative practices
(Mittelstadt et al., 2016), and critically assesses whether, and to what extent, it replaces or
complements our current focus on individual responsibility in the digital age.
With regard to the fifth question, where refers to the locus of management scholars’
research attention. AI is a particularly broad research field with a great variety of organizational
and societal implications. Accordingly, researchers from disciplines such as astronomy
(Haiman, 2019), biology (Webb, 2018), law (Corrales, Fenwick, & Forgó, 2018), medicine
(Topol, 2019), politics (Helbing et al., 2019), psychology (Jaeger, 2016), and sociology
(McFarland, Lewis, & Goldberg, 2016) have all addressed and presented conceptual ideas. In
this regard, our focus was exclusively on the emergent use of AI for managerial tasks in
practice. While management scholars may (and should) become involved in the broader
discussion of AI and its societal implications, the use of automation and augmentation in
management relates to the core of management scholars’ research. We therefore suggest that
the management domain should be the specific focus of our attention.
In this review essay, our focus was predominantly on questions of organizational
functioning, such as those pertaining to the effect of collaboration between multiple humans
and machines on managerial tasks. While such meso-level topics will likely play a central role
in future research, organizations’ use of AI in management should also be explored on the
micro and macro levels of analysis. Micro-level research could investigate how the emergence
of AI-based solutions changes the role of managers in organizations. In the past, management
theories emphasized managers’ domain expertise, which granted them expert power and status
in their organizations (Finkelstein, 1992). Although domain expertise remains relevant for
managers regarding educating and challenging machines, automation and augmentation will
lead to institutionalized knowledge for example, in the form of algorithms – which is often
superior to individual managers’ expert knowledge. At the same time, general human skills
that complement machines, such as creativity, common sense, and advanced communication
(Davenport & Kirby, 2016: 30), as well as integration skills such as AI literacy, will gain further
importance in an era of automation and augmentation (Daugherty & Wilson, 2018: 191). These
developments could lead to important shifts in managers’ roles, competencies, and status.
Macro-level research could explore how the emergence of automation and augmentation
in management leads to institutional action and change. For example, AI is often applied in
open systems, blurring organizational boundaries (Panait & Luke, 2005); data are collected
widely, with diverse stakeholders updating them continuously and collectively through their
actions; and inputs from agents within and outside the organization impact the automation and
augmentation process, which, in turn, can have wide-reaching societal implications. A core
focus of management scholars’ future research attention should therefore be on studying how
broader networks of actors, comprising activists, companies, governments, public institutions,
and international organizations, collaborate to set standards, build institutions, and organize
collective action to address issues pertaining to the use of AI in management.
This leaves the last remaining question of when scholars should address the phenomenon
of emerging AI use in managerial practice. The answer is: Immediately!” Managers in key
organizational domains, including customer management, human resources, marketing,
product innovation, sales, and strategy have already started working closely with intelligent
machines on automated and augmented tasks. This introduction of AI in practice will
profoundly change the nature of management. These developments offer many fruitful areas
for scholarly research. Management scholars still have the opportunity to make a lasting impact
on how organizations perceive and cope with the complex challenges they face. Our review
essay shows that there is an urgent need for a better understanding, more reliable theories, and
sustainable managerial solutions. We therefore close with a call to action and encourage our
readers to embrace the topic of AI in management.
1 This step can also be done by using unsupervised machine learning (Russell & Norvig, 2009),
which allows the machine to induce rules directly from the data. If the task is deterministic,
and the rules are simple and clear, these rules can be readily used for automation.
2 A utility function represents the organization’s preference ordering over a choice set, allowing
it to assign a real value to each alternative. In the field of AI, utility functions are used to convey
various outcomes’ relative value to machines, which in turn allows them to propose alternatives
that optimize the utility function (Russell & Norvig, 2009).
3 Consistent with the three books, we adopt a broad definition of AI, comprising both rule-
based automation and machine learning. In rule-based automation, which is sometimes also
called “robotic process automation” (RPA), the machine is static in the sense that it adheres to
the explicit rules it has been given (Daugherty & Wilson, 2018: 50; Davenport & Kirby, 2016:
48). In contrast, machine learning gives the machine the ability to learn from experience
without being explicitly programmed to do so (Mitchell, 1997).
4 Machine-learning solutions already employ measures against overfitting, such as cross-
validation and regularization. However, these measures can only complement, not replace
human responsibility and intervention in managerial tasks (Greenwald & Oertel, 2017).
5 The computer-science literature distinguishes between tractable or polynomial (P) problems
and intractable or non-polynomial (NP) problems (Dean, 2016; Hartmanis & Stearns, 1965).
Less complex (P) problems are amenable to optimization and rule-based automation. In
contrast, machines working on more complex (NP) problems encounter optimization problems.
While the optimal solution may be out of reach, machine-learning solutions can find models
that approximate such a solution with certain accuracy. These solutions are therefore
suboptimal (given that they inevitably relax certain real-life constraints), but may be close
enough to the optimal solution to be suitable for practical application (Fortnow, 2013).
6 The talent acquisition process in our JP Morgan Chase example functions similarly: HR
managers now engage in augmentation to initially set the machine’s objectives and constraints
(objective setting) and to subsequently select from the candidates that the machine suggested
(candidate selection) (Riley, 2018).
7 A recent example is the media hype around AlphaGo Zero, an AI-based system representing
state-of-the-art Chess and Go play (Silver et al., 2017). AlphaGo Zero learned these games
through trial-and-error (or reinforcement) without human guidance, only playing games against
itself. However, people often overlook that programmers still needed to feed AlphaGo Zero an
important piece of human knowledge: the rules of the game. Chess and Go have explicit, finite,
and stable goals, rules, and reward signals, which allow machine learning to be optimized.
Most real-world managerial problems are far more complex than games like Chess or Go. For
example, the rules of managerial problems might not be known, might be ambiguous, and/or
might change over time. While AlphaGo Zero is impressive, it represents little if any progress
toward artificial general intelligence.
8 Intractable (NP) problems imply discrete optimization problems. While the optimal solution
is out of reach, relaxation of the constraints allows these problems to be addressed (Fortnow,
2013). Humans use their experience to reduce the search space of exponential possibilities by
means of heuristic selection. Machines can subsequently use approximation algorithms to
provide a range of possible solutions that all relax certain real-life constraints.
9 The phenomenon of “catastrophic forgettingexplains this machine limitation; that is, having
learned one task and subsequently transferred to another, a machine-learning system simply
“forgets” how to perform the previously learned task (Taylor & Stone, 2009). Humans, on the
other hand, possess the capacity to transfer learning, allowing them to generalize from one task
context to another (Parisi, Kemker, Part, Kanan, & Wermter, 2019).
10 Several current projects are aimed at deploying AI-based agents capable of perceiving and
responding to emotional cues. However, these agents are very limited in their capabilities,
because fundamental technical and ethical challenges limit their potential for human-level
emotional sentience in the foreseeable future (McDuff & Czerwinski, 2018).
11 Prior studies have shown that slack resources are a necessary, but insufficient prerequisite
for organizational search. Slack resources can also induce complacency and inertia, especially
if organizational factors work against leveraging slack resources for search (Desai, in press).
12 The “5w & 1h” questions are a method used in areas as varied as journalism, research, and
police investigations to describe and evaluate a subject comprehensively. The method’s origins
have been traced to Aristotle’s Nicomachean Ethics (Sloan, 2010).
13 We acknowledge that Brynjolfsson and McAfee (2014) provide a far more balanced
discussion of AI’s organizational and societal implications than the two more recent business
books (Daugherty & Wilson, 2018; Davenport & Kirby, 2016).
Acquisti, A., Brandimarte, L., & Loewenstein, G. 2015. Privacy and human behavior in the
age of information. Science, 347: 509–514.
Amershi, S., Cakmak, M., Knox, W. B., & Kulesza, T. 2014. Power to the people: The role
of humans in interactive machine learning. AI Magazine, 35(4): 105–120.
Andriopoulos, C., & Lewis, M. W. 2009. Exploitation-exploration tensions and
organizational ambidexterity: Managing paradoxes of innovation. Organization
Science, 20: 696–717.
Argote, L., & Greve, H. R. 2007. A behavioral theory of the firm - 40 years and counting:
Introduction and impact. Organization Science, 18: 337–349.
Arnold, M., & Noonan, L. 2017. Robots enter investment banks’ trading floors. Financial
Times. July 7.
Autor, D. H. 2015. Paradox of abundance: Automation anxiety returns. In S. Rangan (Ed.),
Performance and progress: Essays on capitalism, business, and society: 237–260.
Oxford: Oxford University Press.
Bartunek, J. M., & Ragins, B. R. 2015. Extending a provocative tradition: Book reviews and
beyond at AMR. Academy of Management Review, 40: 474–479.
Beer, D. 2017. The social power of algorithms. Information, Communication & Society, 20:
Benlian, A., Kettinger, W. J., Sunyaev, A., & Winkler, T. J. 2018. The transformative value
of cloud computing: A decoupling, platformization, and recombination theoretical
framework. Journal of Management Information Systems, 35: 719–739.
Bergstein, B. 2019. Can AI pass the smell test? MIT Technology Review, 122(2): 82–86.
Binns, R., Van Kleek, M., Veale, M., Lyngs, U., Zhao, J., & Shadbolt, N. 2018. “It’s
reducing a human being to a percentage”: Perceptions of justice in algorithmic
decisions. Proceedings of the 2018 CHI Conference on Human Factors in Computing
Systems, 377–391. New York: ACM Press.
Bostrom, N. 2014. Superintelligence: Paths, dangers, strategies. Oxford: Oxford University
Braga, A., & Logan, R. 2017. The emperor of strong AI has no clothes: Limits to artificial
intelligence. Information, 8: 156–177.
Brundage, M., Avin, S., Clark, J., Toner, H., Eckersley, P., Garfinkel, B., Dafoe, A., Scharre,
P., Zeitzoff, T., Filar, B., Anderson, H., Roff, H., Allen, G. C., Steinhardt, J., Flynn, C.,
Ó hÉigeartaigh, S., Beard, S., Belfield, H., Farquhar, S., Lyle, C., Crootof, R., Evans,
O., Page, M., Bryson, J., Yampolskiy, R., & Amodei, D. 2018. The malicious use of
artificial intelligence: Forecasting, prevention, and mitigation. February 20.
Brynjolfsson, E., & McAfee, A. 2014. The second machine age: Work, progress, and
prosperity in a time of brilliant technologies. New York: W.W. Norton.
Brynjolfsson, E., & McAfee, A. 2017. The business of artificial intelligence. Harvard
Business Review, July issue.
Brynjolfsson, E., & Mitchell, T. 2017. What can machine learning do? Workforce
implications. Science, 358: 1530–1534.
Calabretta, G., Gemser, G., & Wijnberg, N. M. 2017. The interplay between intuition and
rationality in strategic decision making: A paradox perspective. Organization Studies,
38: 365–401.
Cariani, P. 2010. On the importance of being emergent. Constructivist Foundations, 5: 86–
Castelvecchi, D. 2016. Can we open the black box of AI? Nature, 538: 20–23.
Cheney-Lippold, J. 2011. A new algorithmic identity. Theory, Culture & Society, 28(6):
Corrales, M., Fenwick, M., & Forgó, N. (Eds.). 2018. Robotics, AI and the future of law.
Singapore: Springer.
Cummings, M. M. 2014. Man versus machine or man + machine? IEEE Intelligent Systems,
29(5): 62–69.
Cyert, R. M., & March, J. G. 1963. A behavioral theory of the firm. Englewood Cliffs, NJ:
Dastin, J. 2018. Amazon scraps secret AI recruiting tool that showed bias against women.
idUSKCN1MK08G. October 10.
Daugherty, P., & Wilson, H. J. 2018. Human + machine: Reimagining work in the age of
AI. Boston, MA: Harvard Business Review Press.
Davenport, T. H., & Kirby, J. 2016. Only humans need apply: Winners and losers in the
age of smart machines. New York: HarperCollins.
Davis, E., & Marcus, G. 2015. Commonsense reasoning and commonsense knowledge in
artificial intelligence. Communications of the ACM, 58(9): 92–103.
Dean, W. 2016. Computational complexity theory. In E. N. Zalta (Ed.), The Stanford
encyclopedia of philosophy. Stanford, CA: Metaphysics Research Lab, Stanford
Deng, Y., Bao, F., Kong, Y., Ren, Z., & Dai, Q. 2017. Deep direct reinforcement learning for
financial signal representation and trading. IEEE Transactions on Neural Networks
and Learning Systems, 28: 653–664.
Derrida, J. 1967. De la grammatologie. Paris: Les Éditions de Minuit.
Desai, V. M. In press. Can busy organizations learn to get better? Distinguishing between the
competing effects of constrained capacity on the organizational learning process.
Organization Science, doi: 10.1287/orsc.2019.1292.
Deutsche Telekom. 2018. Deutsche Telekom’s guidelines for artificial intelligence. Deutsche
l-180710-ki-leitlinien-en-data.pdf. April 24.
Elsbach, K. D., & Stigliani, I. 2019. New information technology and implicit bias. Academy
of Management Perspectives, 33: 185–206.
Endsley, M. R., & Kiris, E. O. 1995. The out-of-the-loop performance problem and level of
control in automation. Human Factors, 37: 381–394.
Fails, J. A., & Olsen, D. R. 2003. Interactive machine learning. Proceedings of the 8th
International Conference on Intelligent User Interfaces, 39–45. New York: ACM
Fan, J., Han, F., & Liu, H. 2014. Challenges of big data analysis. National Science Review,
1: 293–314.
Feloni, R. 2017. Consumer-goods giant Unilever has been hiring employees using brain
games and artificial intelligence - and it’s a huge success. Business Insider.
June 28.
Finkelstein, S. 1992. Power in top management teams: Dimensions, measurement, and
validation. Academy of Management Journal, 35: 505–538.
Fleming, N. 2018. How artificial intelligence is changing drug discovery. Nature, 557: S55–
Floridi, L. 2008. Information ethics: A reappraisal. Ethics and Information Technology, 10:
Floridi, L. 2013. The philosophy of information. Oxford, UK: Oxford University Press.
Floridi, L., & Sanders, J. W. 2004. On the morality of artificial agents. Minds and Machines,
14: 349–379.
Floridi, L., & Taddeo, M. 2016. What is data ethics? Philosophical Transactions of the
Royal Society A: Mathematical, Physical and Engineering Sciences, 374: 1–4.
Ford, M. 2015. The rise of the robots: Technology and the threat of mass unemployment.
New York: Basic Books.
Fortnow, L. 2013. The golden ticket: P, NP, and the search for the impossible. Princeton,
NJ: Princeton University Press.
Foucault, M. 1977. Discipline and punish: The birth of the prison. New York: Pantheon
Gillespie, T. 2014. The relevance of algorithms. In T. Gillespie, P. J. Boczkowski, & K. A.
Foot (Eds.), Media technologies: Essays on communication, materiality, and society:
167–194. Cambridge, MA: The MIT Press.
Goodwin, R., Maria, J., Das, P., Horesh, R., Segal, R., Fu, J., & Harris, C. 2017. AI for
fragrance design. Proceedings of the machine learning for creativity and design
workshop at NIPS. San Diego, CA: Neural Information Processing Systems
Greenwald, H. S., & Oertel, C. K. 2017. Future directions in machine learning. Frontiers in
Robotics and AI, 3: 79.
Haiman, Z. 2019. Learning from the machine. Nature Astronomy, 3: 18–19.
Hartmanis, J., & Stearns, R. E. 1965. On the computational complexity of algorithms.
Transactions of the American Mathematical Society, 117: 285–306.
Helbing, D., Frey, B. S., Gigerenzer, G., Hafen, E., Hagner, M., Hofstetter, Y., van den
Hoven, J., Zicari, R. V, & Zwitter, A. 2019. Will democracy survive big data and
artificial intelligence? In D. Helbing (Ed.), Towards digital enlightenment: 73–98.
Cham, Switzerland: Springer.
Hoc, J.-M. 2001. Towards a cognitive approach to human-machine cooperation in dynamic
situations. International Journal of Human-Computer Studies, 54: 509–540.
Holzinger, A. 2016. Interactive machine learning for health informatics: When do we need
the human-in-the-loop? Brain Informatics, 3: 119–131.
Huang, H.-H., Hsu, J. S.-C., & Ku, C.-Y. 2012. Understanding the role of computer-mediated
counter-argument in countering confirmation bias. Decision Support Systems, 53: 438–
IBM Think Blog. 2017. Transparency and trust in the cognitive era. IBM Think Blog. January 17.
Jaeger, H. 2016. Deep neural neasoning. Nature, 538: 467–468.
Jarzabkowski, P., Bednarek, R., Chalkias, K., & Cacciatori, E. 2019. Exploring inter-
organizational paradoxes: Methodological lessons from a study of a grand challenge.
Strategic Organization, 17: 120–132.
Jordan, M. I., & Mitchell, T. M. 2015. Machine learning: Trends, perspectives, and prospects.
Science, 349: 255–260.
Kellogg, K., Valentine, M., & Christin, A. In press. Algorithms at work: The new contested
terrain of control. Academy of Management Annals, doi: 10.5465/annals.2018.0174.
Kurzweil, R. 2014. The singularity is near. In R. L. Sandler (Ed.), Ethics and emerging
technologies: 393–406. London: Palgrave Macmillan.
La Roche, J. 2017. IBM’s Rometty: The skills gap for tech jobs is “the essence of divide.”
Yahoo! Finance.
essence-divide-175847484.html. November 16.
Langley, P., & Simon, H. A. 1995. Applications of machine learning and rule induction.
Communications of the ACM, 38(11): 54–64.
Larrick, R. P. 2004. Debiasing. In D. J. Koehler & N. Harvey (Eds.), Blackwell handbook of
judgment and decision making: 316–338. Malden, MA: Blackwell.
Levinthal, D. A., & March, J. G. 1993. The myopia of learning. Strategic Management
Journal, 14: 95–112.
Lichfield, G. 2018. The precision medicine issue. MIT Technology Review.
issue. October 23.
Lindebaum, D., Vesa, M., & den Hond, F. In press. Insights from The Machine Stops to
better understand rational assumptions in algorithmic decision-making and its
implications for organizations. Academy of Management Review, doi:
Lüscher, L. S., & Lewis, M. W. 2008. Organizational change and managerial sensemaking:
Working through paradox. Academy of Management Journal, 51: 221–240.
Lyon, D. 2003. Surveillance as social sorting: Privacy, risk, and digital discrimination.
London: Routledge.
Marr, B. 2018. The amazing ways how Unilever uses artificial intelligence to recruit & train
thousands of employees. Forbes.
uses-artificial-intelligence-to-recruit-train-thousands-of-employees. December 14.
McDuff, D., & Czerwinski, M. 2018. Designing emotionally sentient agents.
Communications of the ACM, 61(12): 74–83.
McFarland, D. A., Lewis, K., & Goldberg, A. 2016. Sociology in the era of big data: The
ascent of forensic social science. The American Sociologist, 47: 12–35.
Miron-Spektor, E., Ingram, A., Keller, J., Smith, W. K., & Lewis, M. W. 2018.
Microfoundations of organizational paradox: The problem is how we think about the
problem. Academy of Management Journal, 61: 26–45.
Mitchell, T. M. 1997. Machine learning. Boston, MA: McGraw-Hill.
Mittelstadt, B. D., Allo, P., Taddeo, M., Wachter, S., & Floridi, L. 2016. The ethics of
algorithms: Mapping the debate. Big Data & Society, 3(2): 1–21.
More, M., & Vita-More, N. 2013. The transhumanist reader: Classical and contemporary
essays on the science, technology, and philosophy of the human future. Malden, MA:
Nadella, S. 2016. The partnership of the future. Slate.
work-together-to-solve-societys-challenges.html. June 28.
Newell, A., Shaw, J. C., & Simon, H. A. 1959. Report on a general problem solving program.
International Conference on Information Processing, 256–264. Santa Monica, CA:
Rand Corporation.
Newell, A., & Simon, H. 1956. The logic theory machine - A complex information
processing system. IRE Transactions on Information Theory, 2: 61–79.
Nilsson, N. J. 1971. Problem-solving methods in artificial intelligence. New York:
Norris, P. 2001. Digital divide: Civic engagement, information poverty, and the internet
worldwide. Cambridge, UK: Cambridge University Press.
O’Neil, C. 2016. Weapons of math destruction: How big data increases inequality and
threatens democracy. New York: Crown.
Orlikowski, W. J. 2007. Sociomaterial practices: Exploring technology at work.
Organization Studies, 28: 1435–1448.
Panait, L., & Luke, S. 2005. Cooperative multi-agent learning: The state of the art.
Autonomous Agents and Multi-Agent Systems, 11: 387–434.
Parasuraman, R., & Manzey, D. H. 2010. Complacency and bias in human use of automation:
An attentional integration. Human Factors, 52: 381–410.
Parisi, G. I., Kemker, R., Part, J. L., Kanan, C., & Wermter, S. 2019. Continual lifelong
learning with neural networks: A review. Neural Networks, 113: 54–71.
Poole, M. S., & van de Ven, A. H. 1989. Using paradox to build management and
organization theories. Academy of Management Review, 14: 562–578.
Press, G. 2016. Only Humans Need Apply is a must-read on AI for Facebook executives.
a-must-read-on-ai-for-facebook-executives. September 21.
Putnam, L. L., Fairhurst, G. T., & Banghart, S. 2016. Contradictions, dialectics, and
paradoxes in organizations: A constitutive approach. Academy of Management Annals,
10: 65–171.
Rahwan, I., Cebrian, M., Obradovich, N., Bongard, J., Bonnefon, J.-F., Breazeal, C.,
Crandall, J. W., Christakis, N. A., Couzin, I. D., Jackson, M. O., Jennings, N. R.,
Kamar, E., Kloumann, I. M., Larochelle, H., Lazer, D., McElreath, R., Mislove, A.,
Parkes, D. C., Pentland, A. “Sandy,” Roberts, M. E., Shariff, A., Tenenbaum, J. B., &
Wellman, M. 2019. Machine behaviour. Nature, 568: 477–486.
Rai, A., Constantinides, P., & Sarker, S. 2019. Next-generation digital platforms: Toward
human-AI hybrids. MIS Quarterly, 43: iii–ix.
Raisch, S., Hargrave, T. J., & van de Ven, A. H. 2018. The learning spiral: A process
perspective on paradox. Journal of Management Studies, 55: 1507–1526.
Riley, T. 2018. Get ready, this year your next job interview may be with an A.I. robot.
mya-other-start-ups.html. March 13.
Russell, S., & Norvig, P. 2009. Artificial intelligence: A modern approach (3rd ed.).
Englewood Cliffs, NJ: Prentice-Hall.
Sabherwal, R., & Jeyaraj, A. 2015. Information technology impacts on firm performance: An
extension of Kohli and Devaraj (2003). MIS Quarterly, 39: 809–836.
Schad, J., & Bansal, P. 2018. Seeing the forest and the trees: How a systems perspective
informs paradox research. Journal of Management Studies, 55: 1490–1506.
Schad, J., Lewis, M. W., Raisch, S., & Smith, W. K. 2016. Paradox research in management
science: Looking back to move forward. Academy of Management Annals, 10: 5–64.
Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T.,
Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L., van den
Driessche, G., Graepel, T., & Hassabis, D. 2017. Mastering the game of Go without
human knowledge. Nature, 550: 354–359.
Simon, H. A. 1987. Two heads are better than one: The collaboration between AI and OR.
Interfaces, 17: 8–15.
Skitka, L. J., Mosier, K., & Burdick, M. D. 2000. Accountability and automation bias.
International Journal of Human-Computer Studies, 52: 701–717.
Sloan, M. C. 2010. Aristotle’s Nicomachean Ethics as the original Locus for the Septem
Circumstantiae. Classical Philology, 105: 236–251.
Smith, W. K., & Lewis, M. W. 2011. Toward a theory of paradox: A dynamic equilibrium
model of organizing. Academy of Management Review, 36: 381–403.
Smith, W. K., & Tracey, P. 2016. Institutional complexity and paradox theory:
Complementarities of competing demands. Strategic Organization, 14: 455–466.
Staw, B. M. 1981. The escalation of commitment to a course of action. Academy of
Management Review, 6: 577–587.
Stephan, M., Brown, D., & Erickson, R. 2017. Talent acquisition: Enter the cognitive
recruiter. Deloitte Insights.
capital-trends/2017/predictive-hiring-talent-acquisition.html. February 28.
Sterman, J. 2000. Business dynamics: Systems thinking and modeling for a complex world.
Boston, MA: Irwin/McGraw-Hill.
Taylor, M. E., & Stone, P. 2009. Transfer learning for reinforcement learning domains: A
survey. Journal of Machine Learning Research, 10(1): 1633–1685.
The Economist. 2017. Harvard Business School risks going from great to good. The
risks-going-from-great-to-good. May 4.
Topol, E. J. 2019. High-performance medicine: The convergence of human and artificial
intelligence. Nature Medicine, 25: 44–56.
Tsoukas, H. 2017. Don’t simplify, complexify: From disjunctive to conjunctive theorizing in
organization and management studies. Journal of Management Studies, 54: 132–153.
Walsh, T. 2017. The singularity may never be near. AI Magazine, 38(3): 58–62.
Webb, S. 2018. Deep learning for biology. Nature, 554: 555–557.
Weick, K. E. 2007. The generative properties of richness. Academy of Management
Journal, 50: 14–19.
Westcott Grant, K. 2018. Netflix’s data-driven strategy strengthens claim for “best original
content” in 2018. Forbes.
strategy-strengthens-lead-for-best-original-content-in-2018. May 28.
Wladawsky-Berger, I. 2018. Human + Machine: The impact of AI on business
transformation. Wall Street Journal.
machine-the-impact-of-ai-on-business-transformation. April 13.
Zuboff, S. 1988. In the age of the smart machine: The future of work and power. New
York: Basic Books.
Zuboff, S. 2019. The age of surveillance capitalism: The fight for a human future at the
new frontier of power. New York: PublicAffairs.
Sebastian Raisch ( is Professor of Strategy at GSEM, University
of Geneva, Switzerland. He received his Ph.D. in Management from the University of Geneva
and his habilitation from the University of St. Gallen. His research interests include artificial
intelligence, organizational ambidexterity, and organizational paradox.
Sebastian Krakowski ( is a Ph.D. candidate at GSEM,
University of Geneva, and a research associate at the House of Innovation, Stockholm School
of Economics, Sweden. He was a visiting researcher at Warwick Business School. His research
explores artificial intelligence’s behavioral implications in organizations and society.
... Yet, we are currently lacking empirical research that analyzes how organizations decide about the level of augmentation/automation (Coombs et al., 2020), what kind of value organizations try to generate with AI (Lyytinen et al., 2020), and what organizational strategies they employ to pursue value from AI (Berente et al., 2021;Günther et al., 2017). We address the calls from both information systems (IS) (Coombs et al., 2020;Galliers et al., 2017;Markus, 2017;Rai et al., 2019) and management scholars (Raisch & Krakowski, 2020;von Krogh, 2018) to study how organizations pursue their value targets through AI applications. ...
... While the types and subtypes are not new per se, they do provide evidence that organizations employ them sequentially on a project level but simultaneously on a portfolio level to pursue different value targets from ML applications. Scholars have largely discussed the use of AI for achieving value targets through augmentation and/or automation mechanisms (Grønsund & Aanestad, 2020;Raisch & Krakowski, 2020). The ML value creation mechanism of knowledge creation has so far remained primarily detached from the augmentation and automation mechanisms (e.g., Berente et al., 2019;Tremblay et al., 2021). ...
... Our process model contributes to the current literature on designing and implementing strategies in the context of ML systems in three ways. First, as mentioned in the background section, most studies on ML or AI value creation in organizations are either conceptual (e.g., Coombs et al., 2020;Davenport & Ronanki, 2018;Raisch & Krakowski, 2020) or pursue a variance perspective (Jöhnk et al., 2021;Mikalef & Gupta, 2021;Pumplun et al., 2019). This suggests lists of AI readiness factors (e.g., AI resources and capabilities that organizations need) without, however, necessarily taking into account that these capabilities or environmental factors might change over time. ...
Full-text available
Advancements in artificial intelligence (AI) technologies are rapidly changing the competitive landscape. In the search for an appropriate strategic response, firms are currently engaging in a large variety of AI projects. However, recent studies suggest that many companies are falling short in creating tangible business value through AI. As the current scientific body of knowledge lacks empirically-grounded research studies for explaining this phenomenon, we conducted an exploratory interview study focusing on 56 applications of machine learning (ML) in 29 different companies. Through an inductive qualitative analysis, we uncover three broad types and five subtypes of ML value creation mechanisms, identify necessary but not sufficient conditions for successfully leveraging them, and observe that organizations, in their efforts to create value, dynamically shift from one ML value creation mechanism to another by reconfiguring their ML applications (i.e., the shifting practice). We synthesize these findings into a process model of ML value creation, which illustrates how organizations engage in (resource) orchestration by shifting between ML value creation mechanisms as their capabilities evolve and business conditions change. Our model provides an alternative explanation for the current high failure rate of ML projects.
... These approaches are promising, not only for applications where it is necessary to make algorithmic judgement interpretable to humans (e.g., for legal or ethical decisions), but also for applications where AI is employed to provide more insights to humans-uncover patterns in data, not only making predictions-enabling more informed human 2 decisions. Therefore, XAI supports the use of AI to augment work (Grønsund and Aanestad, 2020;Raisch and Krakowski, 2020) instead of replacing humans (Frey and Osborne, 2017). Current XAI approaches, however, are criticized for (i) focusing too heavily on technical aspects or data perspectives of developers and (ii) including few aspects of social sciences and human-computer interaction (Abdul et al., 2018;Miller, 2019). ...
... Yet, the energy consumption context is also an interesting study site from an information systems perspective, because extensive time series data is available, which contains complex patterns that may not be easy to recognize by humans. Given 13 the recent calls for empirical research to get a better understanding on how to integrate AI in human workplaces (Lyytinen et al., 2020;Rai et al., 2019), and the importance of AI technology to support humans by augmenting reality rather than replacing humans by AI (Raisch and Krakowski, 2020), our study demonstrated the power of XAI methods in human-AI interface design and highlights areas of further research and development. ...
Full-text available
Machine learning (ML) methods can effectively analyse data, recognize patterns in them, and make high-quality predictions. Good predictions usually come along with "black-box" models that are unable to present the detected patterns in a human-readable way. Technical developments recently led to eXplainable Artificial Intelligence (XAI) techniques that aim to open such black-boxes and enable humans to gain new insights from detected patterns. We investigated the application of XAI in an area where specific insights can have a significant effect on consumer behaviour, namely electricity use. Knowing that specific feedback on individuals' electricity consumption triggers resource conservation, we created five visualizations with ML and XAI methods from electricity consumption time series for highly personalized feedback, considering existing domain-specific design knowledge. Our experimental evaluation with 152 participants showed that humans can assimilate the pattern displayed by XAI visualizations, but such visualizations should follow known visualization patterns to be well-understood by users.
... Indeed, a considerable amount of research has been conducted in the recent past on the influence of algorithms on organizations (e.g. Fang et al., 2019;Kellogg et al., 2020), and as to how AI may support human management (Raisch & Krakowski, 2020;Simpson & Berti, 2019). A few authors pointed out the dangers and risks of human leadership (e.g. ...
... The AI-based clustering approach applied in this study is best described as augmented and far from automatized (Raisch and Krakowski, 2021). This study has chosen a more conservative approach, including human interference, to identify shared needs carefully. ...
Full-text available
Seeking inspiration from other perspectives is a prominent mechanism to support ideation. AI-based language models can help overcome information processing limits and efficiently structure large solution spaces spanned by prior ideas. However, it remains unclear how the search through a solution space affects the subsequent idea generation. This study explores the influence of different sets of prior idea stimuli pre-structured by an AI-supported clustering on ideation outcomes. The sets varied in quantity and semantic diversity. In a survey experiment, 181 participants generated 447 ideas evaluated according to major idea performance characteristics. Results indicate that seeing an extensive set of ideas from various clusters improves idea novelty and positively and semantic diversity. In a survey experiment, 181 participants generated 447 ideas evaluated according to major idea performance characteristics. Results indicate that seeing an extensive set of ideas from various clusters improves idea novelty and positively interacts with domain-specific knowledge. However, it negatively affects idea feasibility and specificity. These findings encourage innovators seeking particularly novel ideas to complement their current processes with AI-supported clustering tools while taking steps to avoid vagueness.
... In addition to rapid technological advancements affecting where work is done, technology is also affecting how work is performed. Already, the use of artificial intelligence (AI) is altering how the jobs of managerial decision-makers and medical personnel are completed (Bartleby, 2019;Malone, 2018;Raisch & Krakowski, 2020). Whereas in the past automation tended to replace blue-collar jobs, the increased use of AI is predicted to replace the jobs of many white-collar workers (Schwab, 2017). ...
Increased globalization and rapid technological advances have greatly affected how careers are enacted. Many of today's workers are transitioning across the boundaries of occupations, organizations, and countries more frequently than previous generations of workers. Despite these great changes in the contemporary workplace, relatively little recent academic consideration has been devoted to the career transition (CT) construct or the commonalities in the decision-making process used by individuals to determine whether to make a CT. Given the many significant changes that have affected careers over the last several decades, the purpose of this paper is to offer a contemporary typology of CTs, a theory-driven model of the voluntary inter-role CT decision-making process, and propositions for future research. The proposed typology and model should also help people make better decisions about CTs.
... Whereas automation involves handing over human tasks to machine with little or no further human involvement, augmentation implies continued close interaction between humans and machines. Augmentation is conceived as a co-evolutionary process in which humans learn from machine and machine from humans (Raisch and Krakowski, 2020). ...
Full-text available
This systematic review brings together the collection of recent scholarly outputs on the disruptive impact of digital transformation on the work. This paper draws from a sample of 68 outputs from 2011 to 2022. We identify three key theoretical perspectives: socio-technical systems theory, skill-biased technological change, and political economy of digital transformation. The articles provide complementary insights on cross-cutting themes of technological unemployment, wage inequality and job polarization. They also highlight often conflicting views about technology ownership, work-less utopia, education reforms and the imperative of human-centricity in appropriation of technology. Drawing on the findings across the whole spectrum of theoretical and analytical perspectives, we offer critical reflections about the factors that will define the work of the future, in terms of skills, creativity and opportunities for autonomous workers. We also discuss the political and institutional processes that will shape the future of work. Finally, we offer recommendations for future research and policy interventions.
In developing the highly useful technologies, knowledge from human factors and ergonomics (HF/E) can be of great use, especially to designers charged with the difficult task of dovetailing humans and machines in complex systems built to navigate sometimes chaotic environments. The role of HF/E in A ³ design remains centered around the goal that A ³ self-action is ideal to provide maximum benefit to humans while increasing the likelihood of task success. This chapter is written between “artificial intelligence (AI) winters,” times of decreased funding of AI technologies, indeed at a time of great optimism and investments in A ³ technologies. It argues that design is key to all A ³ technologies, and that the advent of autonomy which has no need for humans is not only unlikely, but likely undesirable. The chapter endeavors to provide tools from the HF/E literature with which to shape the development of A ³ with respect to our knowledge of human factors.
Full-text available
The procurement function and the sales and marketing function are a supply chain organization’s main boundary-spanning functions. Each has a specific focus and different views and objectives - so together, the two often fight for relative competitive advantage. The digitization of processes and the rise of enterprise resources planning systems led to a data abundance, yet the outcomes are not yet satisfactory in practical application. This presentation will share results from an explanatory study that considered key takeaways related to procurement in contrast sales and marketing.
In industry, the Fourth Industrial Revolution is transforming the roles of people, technology and work on the shop floor. Despite ongoing strides towards automation, people are anticipated to remain integral contributors in future manufacturing. Where full automation is ineffective or in-feasible, Operator Assistance Systems (OAS) can augment workers' cognitive or physical capabilities. We frame OAS as a subset of Human-Computer Interaction (HCI) systems designed for the purpose of workforce augmentation in production systems. However, while OAS are anticipated to address key needs in industry, a challenge for both OAS researchers and industrial practitioners is to identify the most promising applications of OAS and justify them from a value-added perspective. This paper addresses this challenge by presenting a systematic literature review of 2,928 papers, revealing (a) 11 application areas for OAS; and (b) 12 approaches for assessing the value-added of OAS. Moreover, we discuss implications for OAS, with a particular focus on integrating OAS in industry.
Full-text available
Advancing AI capabilities make the technology increasingly relevant for enabling better and faster decisions. AI plays different roles in different types of decisions, with the most common AI-enabled decisions involving repetitive, tactical, and structured situations. These are also the types of decisions that are most likely to be fully or partially automated. For unstructured and semi-structured decisions, AI plays a more supporting role in informing human decision-makers. Although the primary effect of AI is to augment human intelligence rather than replace it, AI is capable of transforming managerial decision processes, allowing managers to make earlier, simulated, and complementary decisions. It also becomes incumbent upon managers to understand the ways that AI-enabled decision tools operate, and when the models on which they rely no longer reflect current reality and need to be retrained. Organizations should begin to redesign key decision processes with these new capabilities and responsibilities in mind.
Artificial intelligence and related technologies are changing both the law and the legal profession. In particular, technological advances in fields ranging from machine learning to more advanced robots, including sensors, virtual realities, algorithms, bots, drones, self-driving cars, and more sophisticated “human-like” robots are creating new and previously unimagined challenges for regulators. These advances also give rise to new opportunities for legal professionals to make efficiency gains in the delivery of legal services. With the exponential growth of such technologies, radical disruption seems likely to accelerate in the near future. This collection brings together a series of contributions by leading scholars in the newly emerging field of artificial intelligence, robotics, and the law. The aim of the book is to enrich legal debates on the social meaning and impact of this type of technology. The distinctive feature of the contributions presented in this edition is that they address the impact of these technological developments in a number of different fields of law and from the perspective of diverse jurisdictions. Moreover, the authors utilize insights from multiple related disciplines, in particular social theory and philosophy, in order to better understand and address the legal challenges created by AI. Therefore, the book will contribute to interdisciplinary debates on disruptive new AI technologies and the law.
Insomnia affects up to 15% of the US population. There are effective pharmacologic and behavioral treatments for insomnia; however, there is often no one-size-fits-all intervention. This article discusses the leading behavioral treatment of insomnia, cognitive behavioral therapy for insomnia, and its ability to be tailored to an individual's specific symptoms. It then discusses pharmacologic options for treating insomnia, and offers some guidance on medication selection to enhance personalized care. In addition, it discusses how the current evidence base can help providers make choices between pharmacologic and behavioral treatments.
The widespread implementation of algorithmic technologies in organizations prompts questions about how algorithms may reshape organizational control. We use Edwards’ (1979) perspective of “contested terrain,” wherein managers implement production technologies to maximize the value of labor and workers resist, to synthesize the interdisciplinary research on algorithms at work. We find that algorithmic control in the workplace operates through six main mechanisms, which we call the “6 Rs”—employers can use algorithms to direct workers by restricting and recommending, evaluate workers by recording and rating, and discipline workers by replacing and rewarding. We also discuss several key insights regarding algorithmic control. First, labor process theory helps to highlight potential problems with the largely positive view of algorithms at work. Second, the technical capabilities of algorithmic systems facilitate a form of rational control that is distinct from the technical and bureaucratic control used by employers for the past century. Third, employers’ use of algorithms is sparking the development of new algorithmic occupations. Finally, workers are individually and collectively resisting algorithmic control through a set of emerging tactics we call algoactivism. These insights sketch the contested terrain of algorithmic control and map critical areas for future research.
Organizations are getting busier, but can they still learn to get better? This question has urgent practical importance, since competitive pressures in a wide variety of industries have resulted in organizations that increasingly strain their operating limits. This question is deeply connected with organizational learning theory, since organizations operating with constrained capacity may gain experience but lose the ability to digest it—challenging the overall organization’s ability to learn and improve. Some research, though, suggests a seemingly contradictory perspective, with constrained capacity perhaps motivating organizations to adopt more flexible approaches and learn out of necessity. This study integrates the perspectives to examine how constrained capacity impacts organizational learning. To explore this question, the study develops separate theory regarding the amount and timing of capacity crises, suggesting that increasingly constrained capacity tends to detract from learning, but, uniquely, that consistently constrained capacity, rather than periodic spikes, may instead lead to better learning. Hypothesis tests provide support for several of the study’s arguments.
The use of artificial intelligence, and the deep-learning subtype in particular, has been enabled by the use of labeled big data, along with markedly enhanced computing power and cloud storage, across all sectors. In medicine, this is beginning to have an impact at three levels: for clinicians, predominantly via rapid, accurate image interpretation; for health systems, by improving workflow and the potential for reducing medical errors; and for patients, by enabling them to process their own data to promote health. The current limitations, including bias, privacy and security, and lack of transparency, along with the future directions of these applications will be discussed in this article. Over time, marked improvements in accuracy, productivity, and workflow will likely be actualized, but whether that will be used to improve the patient–doctor relationship or facilitate its erosion remains to be seen.
TODAY, PEOPLE INCREASINGLY rely on computer agents in their lives, from searching for information, to chatting with a bot, to performing everyday tasks. These agent-based systems are our first forays into a world in which machines will assist, teach, counsel, care for, and entertain us. While one could imagine purely rational agents in these roles, this prospect is not attractive for several reasons, which we will outline in this article. The field of affective computing concerns the design and development of computer systems that sense, interpret, adapt, and potentially respond appropriately to human emotions. Here, we specifically focus on the design of affective agents and assistants. Emotions play a significant role in our decisions, memory, and well-being. Furthermore, they are critical for facilitating effective communication and social interactions. So, it makes sense that the emotional component surrounding the design of computer agents should be at the forefront of this design discussion. © 2018 Association for Computing Machinery. All Rights Reserved.
Large cosmological datasets have been probing the properties of our universe and constraining the parameters of dark matter and dark energy with increasing precision. Deep learning techniques have shown potential to be smarter, and to greatly outperform human-designed statistics.