ArticlePDF Available

Abstract and Figures

The paper presents an approach for implementing inscrutable (i.e., nonexplainable) artificial intelligence (AI) such as neural networks in an accountable and safe manner in organizational settings. Drawing on an exploratory case study and the recently proposed concept of envelopment, it describes a case of an organization successfully "enveloping" its AI solutions to balance the performance benefits of flexible AI models with the risks that inscrutable models can entail. The authors present several envelopment methods-establishing clear boundaries within which the AI is to interact with its surroundings, choosing and curating the training data well, and appropriately managing input and output sources-alongside their influence on the choice of AI models within the organization. This work makes two key contributions: It introduces the concept of sociotechnical envelopment by demonstrating the ways in which an organization's successful AI envelopment depends on the interaction of social and technical factors, thus extending the literature's focus beyond mere technical issues. Secondly, the empirical examples illustrate how operationalizing a sociotechnical envelopment enables an organization to manage the trade-off between low explainability and high performance presented by inscrutable models. These contributions pave the way for more responsible, accountable AI implementations in organizations, whereby humans can gain better control of even inscrutable machine-learning models.
Content may be subject to copyright.
ISSN 1536-9323
Journal of the Association for Information Systems (2021) 22(2), 325-352
doi: 10.17705/1jais.00664
RESEARCH ARTICLE
325
Sociotechnical Envelopment of Artificial Intelligence:
An Approach to Organizational Deployment
of Inscrutable Artificial Intelligence Systems
Aleksandre Asatiani1, Pekka Malo2, Per Rådberg Nagbøl3,
Esko Penttinen4, Tapani Rinta-Kahila5, Antti Salovaara6
1University of Gothenburg, Sweden, aleksandre.asatiani@ait.gu.se
2Aalto University School of Business, Finland, pekka.malo@aalto.fi
3IT University of Copenhagen, Denmark, pena@itu.dk
4Aalto University School of Business, Finland, esko.penttinen@aalto.fi
5The University of Queensland, Australia, t.rintakahila@uq.edu.au
6Aalto University School of Arts, Design and Architecture, Finland, antti.salovaara@aalto.fi
Abstract
The paper presents an approach for implementing inscrutable (i.e., nonexplainable) artificial
intelligence (AI) such as neural networks in an accountable and safe manner in organizational
settings. Drawing on an exploratory case study and the recently proposed concept of envelopment,
it describes a case of an organization successfully “enveloping” its AI solutions to balance the
performance benefits of flexible AI models with the risks that inscrutable models can entail. The
authors present several envelopment methodsestablishing clear boundaries within which the AI is
to interact with its surroundings, choosing and curating the training data well, and appropriately
managing input and output sourcesalongside their influence on the choice of AI models within the
organization. This work makes two key contributions: It introduces the concept of sociotechnical
envelopment by demonstrating the ways in which an organization’s successful AI envelopment
depends on the interaction of social and technical factors, thus extending the literature’s focus beyond
mere technical issues. Secondly, the empirical examples illustrate how operationalizing a
sociotechnical envelopment enables an organization to manage the trade-off between low
explainability and high performance presented by inscrutable models. These contributions pave the
way for more responsible, accountable AI implementations in organizations, whereby humans can
gain better control of even inscrutable machine-learning models.
Keywords: Artificial Intelligence, Explainable AI, XAI, Envelopment, Sociotechnical Systems,
Machine Learning, Public Sector
Hind Benbya was the accepting senior editor. This research article was submitted on February 29, 2020 and underwent
three revisions.
1 Introduction
Advances in big data and machine-learning (ML)
technology have given rise to systems using artificial
intelligence (AI) that bring significant efficiency gains
and novel information-processing capabilities to the
organizations involved. While ML models may be able
to surpass human experts’ performance in demanding
analysis and decision-making situations (McKinney et
al., 2020), their operation logic differs dramatically
from humans’ ways of approaching similar problems.
Rapid growth in the volumes of data and computing
power available has made AI systems increasingly
complex, rendering their behavior inscrutable and,
therefore, hard for humans to interpret and explain
Copyright owned by Association for Information Systems. Use for profit is not allowed.
Link to the Journal of AIS: https://aisel.aisnet.org/jais/
Sociotechnical Envelopment of Artificial Intelligence
326
(Faraj et al., 2018; Stone et al., 2016). While the
economic value of such systems is rarely in doubt,
broader organizational and societal implications,
including negative side-effects such as undetected
biases, have started to cause concerns (Benbya et al.,
2020; Brynjolfsson & McAfee, 2014; Newell &
Marabelli, 2015). Thus, humans’ ability to explain how
AI systems produce their outputs, referred to as
“explainability” (e.g., Rosenfeld & Richardson, 2016),
has become a prominent issue in various fields.
The inscrutability of AI systems leads to a host of
ethics-related, legal, and practical issues. ML models,
by necessity, operate mindlessly, meaning that they
approach the work from a single perspective, with no
conscious understanding of the broader context
(Burrell, 2016; Salovaara et al., 2019). For example,
ML models cannot reflect on the ethics or legality of
their actions. Accordingly, an AI system may exhibit
unintended biases and discrimination after learning to
consider inappropriate factors in its decision-making
(Martin, 2019). Through such problems during the
training stage and beyond, an organization may
(wittingly or not) end up operating in a manner that
conflicts with its values (Firth, 2019), with models
being susceptible to biases and errors connected with
vexing ethics issues, such as discrimination against
specific groups of people. Designing models with solid
ethics in mind could provide means to identify, judge,
and correct such biases and errors (Martin, 2019), but
all of this is impossible if the model’s actions are
inscrutable. Alongside ethics matters, there are
legislative factors that impose concrete and
inescapable requirements for explainability (Desai &
Kroll, 2017). Public authorities often must honor
requirements for transparency in their actions, and
private companies may also be compelled to explain
and justify, for instance, how they use customer data.
The European Union’s General Data Protection
Regulation (GDPR) serves as a prominent example of
recent legislative action that promotes the rights of data
subjects to obtain an explanation of any decision based
on data gathered on them (European Union, 2016).
Yet producing an explainable AI system may not
always be feasible. Inscrutability takes many forms,
linked to such elements as intentional corporate or state
secrecy, technical illiteracy, and innate characteristics
of ML models (Burrell, 2016). This multifaceted
nature, combined with limitations on human logic,
means there are no simple solutions to explainability
problems (Edwards, 2018; Robbins, 2020). For
example, some legal scholars maintain that the
GDPR’s provision for a right to explanation is
insufficient and could result in meaningless
“transparency” that does not actually match user needs
(Edwards & Veale, 2017): while there may technically
be an explanation for a given decision, this might not
be understandable for the person(s) affected. Though
approaches such as legal auditing (O’Neil, 2016;
Pasquale, 2015), robust system design (Rosenfeld &
Richardson, 2019), and user education may improve
explainability in some cases, they are unidimensional
and inadequate for tackling the fundamental challenges
presented by the mindless operation of AI (Burrell,
2016). In an organizational setting, information-
technology (IT) systems affect a broad spectrum of
stakeholders who display differing, often sharply
contrasting, demands and expectations (Koutsikouri et
al., 2018). Explanation of AI agents’ behavior is
further complicated by the environment wherein AI
development takes place, with various incumbent work
processes, structures, hierarchies, and legacy
technologies. These challenges have prompted calls
for human-centered and pragmatic approaches to
explainability (Mittelstadt et al., 2019; Ribera &
Lapedriza, 2019). This invites us to approach
explainability from a sociotechnical perspective to
account for the interconnected nature of technology,
humans, processes, and organizational arrangements,
and thereby give balanced attention to instrumental
and humanistic outcomes of technology alike (Sarker
et al., 2019).
It is against this backdrop that we set out to address the
following research question (RQ): How can an
organization exploit inscrutable AI systems in a safe
and socially responsible manner? Our inquiry was
inspired by a desire to understand how organizations
cope with AI models’ inscrutability when facing
explainability demands. The sociotechnical nature of
the problem became apparent during the early phases
of a research project at the case organization. We
observed a need to integrate the organization’s social
side (people, processes, and organizational structures)
with its technical elements (information technology
and AI systems) synergistically if the organization
wished to take advantage of a wider array of AI
models, including some of the inscrutable models
available. This pursuit involved two types of goals,
explainability- and performance-oriented goals,
which, in the case of AI implementation, present
conflicting demands. Here, we draw on Sarker et al.’s
(2019) concepts of instrumental and humanistic
outcomes of information-system implementation to
analyze the well-known tradeoff between
explainability and accuracy. In its development of
powerful AI models, the organization sought
instrumentally oriented outcomes (better performance
and greater efficiency) but also needed to cater to
humanistic outcomes by making sure that the use of
such models would not diminish human agency or
harm people affected by the models’ use. As we drilled
down to precisely how the organization addressed both
sets of desired outcomes, envelopment emerged as an
illuminating lens for conceptualizing the various
approaches.
Journal of the Association for Information Systems
327
This conceptenvelopment of AIhas recently
emerged as a potentially useful approach to cope with
the explainability challenges described above
(Robbins, 2020). It suggests that, by controlling the
training data carefully, appropriately choosing both
input and output data, and specifying other boundary
conditions mindfully, one may permit even inscrutable
AI to make decisions, because these specific
precautions erect a predictable envelope around the
agent’s virtual maneuvering space. Thus far, however,
envelopment has been illustrated in only a handful of
contexts (e.g., autonomous driving, playing Go, and
recommending apparel) and on a conceptual level
only; thus, relatively limited insights have been
presented for tackling explainability challenges in
complex real-world organizations. To address this gap,
we describe how envelopment is practiced in one
pioneering organization that has embarked on utilizing
AI in its operations, and we show that envelopment is
fundamental to enabling an organization to use
inscrutable systems safely even in settings that
necessitate explainability. Further, we deepen the
concept of envelopment by showing how it emerges via
sociotechnical interactions in a complex organizational
setting. With the empirical findings presented here, we
argue that the sociotechnical envelopment concept has
widespread relevance and offers tools to mitigate many
challenges that stand in the way of making the most of
advanced AI systems.
2 Review of the Literature and
Theory Development
This section offers a review of lessons already learned
from organizational AI implementations and their
sociotechnical underpinnings. Also, we address the
properties of good explanations and provide a more
detailed picture of the envelopment concept.
2.1 A Sociotechnical Approach to
Organizational AI
The recent emergence and proliferation of new
generations of ML tools have reawakened interest in
organizational AI research (Faraj et al. 2018; Keding
2021; Sousa et al. 2019). Like human intelligence, AI
is notoriously difficult to define as a concept. For the
purposes of our study, we follow Kaplan and Haenlein
(2019) in defining AI as a “system’s ability to interpret
external data correctly, to learn from such data, and to
use those learnings to achieve specific goals” (p. 17).
Complementing conceptual works, empirical studies
on the topic have started to appear (e.g.,
Ghasemaghaei, Ebrahimi, & Hassanein, 2018;
Salovaara et al., 2019; Schneider & Leyer, 2019). The
papers have increasingly shifted the position of AI
research from a largely technical one to a perspective
encompassing the social component (Ågerfalk, 2020).
Whereas the technical facet involves the information
systems (IS) angle, IT infrastructure, and platforms,
the social aspect brings in people, work processes,
organizational arrangements, and cultural and societal
factors (Sarker et al., 2019). Although scholars have
discussed issues such as replacing humans with
machines versus augmenting humans’ capabilities
(e.g., Davenport, 2016; Jarrahi, 2018; Raisch &
Krakowski, in press), there is still little critical
empirical work investigating the human aspects
involved with deploying and managing AI in
organizations (Keding, 2021).
Research on organizations’ implementation and use of
AI and other forms of automated decision-making has
highlighted some recurrent patterns. First, AI’s
mindless and, thereby, error-prone nature necessitates
careful control of the AI’s agency and autonomy in the
implementation. Humans can serve as important
counterweights in this equation (Butler & Gray, 2006;
Pääkkönen et al., 2020; Salovaara et al., 2019). The
division of labor and knowledge between humans and
AI can be arranged in various ways whereby
organizations can balance rigidity and predictability
against flexibility and creative problem-solving
(Asatiani et al., 2019; Lyytinen et al., in press). Second,
organizations’ AI agents interact with many types of
human stakeholders, each with a particular dependence
on AI and distinct abilities to understand its operation
(Gregor & Benbasat, 1999; Preece, 2018; Weller,
2019). Studies indicate that AI is rarely considered a
“plug-and-play” technology and that an organization
deploying it requires a clear implementation strategy
that takes into account the wide spectrum of stakeholders
(Keding, 2021). For instance, since the impact of AI’s
implementation varies greatly between stakeholders,
decisions to decouple stakeholders from the process of
designing, implementing, and using it increase the
likelihood of unethical conduct and breach of social
contracts, often leading to the systems’ ultimate failure
(Wright & Schultz, 2018).
Collectively, the literature on organizational AI shows
how important it is for organizations to balance the
risks associated with AI against the efficiency gains
that may be reaped. These considerations also show
that organizational AI deployment entails a significant
amount of coordination and mutual adaptation between
humans and AI and is thus inescapably a matter of
sociotechnical organization design (Pääkkönen et al.,
2020). Those advocating a sociotechnical approach
maintain that attention must be given both to the
technical artifacts and to the individuals/collectives
that develop and utilize the artifacts in social (e.g.,
psychological, cultural, and economic) contexts
(Bostrom et al., 2009; Briggs et al., 2010). In a
corollary to this, taking a sociotechnical stance is
aimed at meeting instrumental objectives (e.g.,
effectiveness and accuracy of the model or other
Sociotechnical Envelopment of Artificial Intelligence
328
artifact developed) and humanistic objectives (e.g.,
engaging users and retaining employee skills) alike
(Mumford, 2006).
Sarker et al. (2019) have reviewed the intricate ways in
which the social and the technical may become
interwoven such that neither the social nor technical
aspects come to dominate. They show that this
relationship is quite varied, and they demonstrate this
by presenting examples of reciprocal as well as
moderating influence, inscription of the social in the
technical, entanglement, and imbrication. For instance,
from the perspective of reciprocal influence,
technology and organizational arrangements may be
seen to coevolve throughout an IS implementation as
they mutually appropriate each other (Benbya &
McKelvey, 2006). From the sociomaterial perspective
of imbrication, in turn, humans and technologies are
viewed as agencies whose abilities interlock to
produce routines and other stable emergent processes.
2.2 Challenges of Inscrutable AI
As noted in the introduction, complex AI models often
promise better performance than simple ones, but such
models also tend to lack transparency, and their
outputs can be hard or even impossible to explain.
Writings on AI explainability often employ the
interrelated concepts of transparency, interpretability,
and explainability in efforts to disentangle the threads
of this problem. Transparency refers to the possibility
of monitoring AI-internal operationse.g., tracing the
paths via which the AI reaches its conclusions
(Rosenfeld & Richardson, 2019; Sørmo et al., 2005).
Its opposite is opacity, a property of “black-box
systems, which hide the decision process from users
and sometimes even from the system’s developers
(Lipton, 2018). The two other concepts
interpretability and explainabilityrefer to the AI
outputs’ understandability for a human (e.g., Doshi-
Velez & Kim, 2017; Miller 2019). On occasion, the
terms are used interchangeably (e.g., Došilović et al.,
2018; Liu et al., 2020) while sometimes authors
employ separate definitions. Often, interpretability has
strong technical connotations while explainability is
more human centered in nature and hence a more
sociotechnically oriented concept.
Many of the more traditional AI models, such as linear
regression, with its handling of only a limited number
of known input variables, and decision trees, which can
display the if-then sequence followed, are considered
explainable. However, more and more of today’s AI
models are so complex that explainability is rendered
virtually impossible. For instance, when a traditional
decision-tree model is “boosted” via a machine-
learning technique called gradient boosting, its
performance improves but its behavior becomes far
more difficult to explain. Other examples of highly
accurate models that lack explainability are deep and
recurrent neural networks, complexly layered
computing systems whose structure resembles that of
the biological networks of a brain’s neurons. Then, one
deems them inscrutable (Dourish, 2016; Martin,
2019), referring to situations wherein the system’s
complexity outstrips practical means of analyzing it
comprehensively. A recent open-domain chatbot
developed at Google, which has 2.6 billion free
parameters in its deep neural network (Adiwardana et
al., 2020), is an extreme example of an AI system
whose inner workings are inscrutable for humans even
if they are transparent.
Unrestrained use of inscrutable systems can be
problematic. Humans interacting with such systems
are unable to validate whether the decisions made by
the system correspond to real-world requirements and
adhere to legal or ethics norms (Rosenfeld &
Richardson, 2019). The issue is far from academic;
after all, reliance on inscrutable systems could lead to
systematic biases in decision-making, completely
invisible to humans interacting with or affected by the
system (Došilović et al., 2018).
In consequence, organizations intending to deploy AI
systems face an explainability-accuracy tradeoff
(Došilović et al., 2018; Linden et al., 2019; London,
2019; Martens et al., 2011; Rosenfeld & Richardson,
2019). On the one hand, complex models with greater
flexibility, such as deep neural networks, often yield
more accurate predictions than do simple ones such as
linear regression or decision trees. On the other hand,
simple models are usually easier for humans to
interpret and explain. The tradeoff that seems to exist
between explainability and accuracy forces the design to
prioritize one over the other: an organization wishing to
reduce the risks associated with inscrutable AI must
settle for AI models with a high degree of explainability.
Figure 1 illustrates this tradeoff, following depictions
by Linden et al. (2019) and Rosenfeld and Richardson
(2019).
One approach recently introduced to address the risks
brought by black-boxed systems is envelopment. In
recognition of its potential for managing the
explainability-accuracy tradeoff, the following section
delves into the suggestions that researchers have
presented in relation to this approach.
2.3 Envelopment
As noted above, we identified envelopment (Floridi,
2011; Robbins, 2020) as a suitable sensemaking
concept when examining the domain of organizational
AI development. In its original context in robotics, a
work envelope is “the set of points representing the
maximum extent or reach of the robot hand or working
tool in all directions” (RIA Robotics Glossary, 73;
cited by Scheel, 1993, p. 30).
Journal of the Association for Information Systems
329
Figure 1. The Explainability-Accuracy Tradeoff
Robots’ work envelopes, often presented as shaded
regions on factories’ floor maps and as striped areas on
factory floors, are a practical solution for fulfilling
what is known as the “principle of requisite variety
(Ashby, 1958)i.e., meeting the requirement that the
number of states of a robot’s logic be larger than the
number of environmental states in which it operates. If
a robot acts in an environment whose complexity
exceeds its comprehension, it will pose a risk to the
surroundings. Work envelopesareas that no other
actors will entercan guarantee that the physical
environment of the robot is simplified sufficiently (i.e.,
that the number of possible states of the environment is
reduced enough). Through this modification, the robot
can handle those states that still need to be controlled,
thereby fulfilling the principle of requisite variety. In
addition to physical parameters, a robot’s envelope
may be specified by means of time thresholds, required
capabilities/responsibilities, and accepted tasks
(McBride & Hoffman, 2016, p. 79). These parameters
are dynamic: when a robot faces new problems, the
envelope parameters are adjusted to accommodate
what the requisite variety now entails (p. 81).
Our research is a continuation of work wherein this
concept has been applied to cases that involve humans
and nonphysical work performed by AI agents. In this
context, the envelope is not physically specified but
relates to the realm of information processing. This
domain change notwithstanding, there remains a need
for collaboration with a human partner who maintains
the envelope and thus guarantees the safety and
correctness of the AI’s operation (Floridi, 2011). Also,
the underlying principle of requisite variety continues
to persist, meaning that the AI should not be used for
tasks it cannot master and that it should not be trained
with data irrelevant to the tasks. Such undesired
effects—“excessive risks” in Figure 1—can manifest
themselves in several forms, among them erroneous
input-action mappings, ethics dilemmas that an AI
agent should not be allowed to tackle by itself, and
behaviors that demonstrate bias (e.g., Robbins, (2020).
Even if the realization of such risks does not impair the
financial bottom line or operations’ efficiency, it can
result in problematic humanistic outcomes. For
example, an AI system that processes job applications
to identify the most promising candidates may increase
the efficiency of an HR department, and consistently
identify candidates that meet requirements for the
position. At the same time, the system could
consistently discriminate against certain groups of
applicants who would otherwise qualify because of a
bias in an underlying model. In such scenarios, AI
actions may not impact the bottom line of the
company, at least in the short term, but may be
nevertheless problematic.
Envelopment can be advanced via several methods.
Figure 2 presents our interpretation of the five methods
that Robbins (2020) articulated. We summarize them
below, then build on them in relation to our study.
Boundary envelopes represent the most general of the
Sociotechnical Envelopment of Artificial Intelligence
330
envelopment methods. The envelope delineates where
the AI operatesfor example, only analyzing images of
human faces photographed in good lighting conditions.
An AI model enveloped in this way will not encounter
any tasks other than those carefully designated for it
(condition A in Figure 2). Robbins (2020) takes the
design of a robot vacuum cleaner as an example. Its
boundary envelopment mechanism means that the robot
does not need to be able to avoid threats that never exist
in indoor domestic spaces (e.g., puddles of water). The
benefit of boundary envelopment is that the AI does not
need to incorporate methods to recognize whether the
agent is being made to operate in scenarios that extend
beyond its ability to comprehend the surroundings (i.e.,
requisite variety).
Figure 2. Illustration of AI Envelopment Methods Suggested by Robbins (2020)
Journal of the Association for Information Systems
331
Among the other envelopment methods are three that
refer to the notion of what content the AI will
manipulate (Robbins, 2020). The first of them is the
training-data envelope, related to the curation of the
correct input-output mappings with which the AI
model is trained. Robbins cites biases and other
representativeness problems (“B” in Figure 2) as
particularly likely to propagate or uphold societal
stereotypes if the envelope is not handled properly.
Input envelopes, in turn, address the technical details
of inputs to the AI. For example, in Robbins’s
example, a recommendation AI uses various pieces of
weather and user data (e.g., temperature, real-time
weather status, and the user’s calendar) to produce
clothing recommendations (e.g., the suggestion to
wear a raincoat). For good results, the data should
arrive from sources that are high quality, noise free,
and of appropriate granularity. Input envelopment
limits input channels to those that meet appropriate
criteria in this regard and prevents poorly understood
sources from affecting the model’s behavior. The
third envelopment method in the “what” category is
the use of output envelopes. These define the set of
actions that may be performed within the realm of the
AI’s operation. In the case of an autonomously
driving car, the outputs might be specified as
speeding up, turning the wheels, and braking. Even if
speeding would be technically possible and
sometimes useful, it presents risks to passengers and
other traffic. Therefore, that output is enveloped out
of an autonomous car’s actions. In Figure 2, “C” and
“D” illustrate the input- and output-envelopment
methods described above.
The fifth and final method, use of a function envelope,
addresses the question of why the AI exists and what
goals and ethics it has been designed to advance. This
category of envelopment is applied to limit the AI’s
use for malicious or otherwise problematic purposes,
even in cases wherein it operates correctly. For
example, the functions of conversational home
assistants such as Echo or Alexa are limited to only a
small set of domestic activities to avoid privacy
infringements (Robbins, 2020). Such filtering out of
functions is denoted as “E” in Figure 2.
Robbins suggests that with such variety of
envelopment methods available, one can either
overcome some problems connected with black-box
AI or neutralize their effects. Our work is thus
informed by the envelopment concept, and we
consider its applicability in complex and emergent
sociotechnical settings. In particular, we maintain that
humans play an important role in an AI agent’s
envelopment and in how it is organized by striving to
guarantee that the AI does not face tasks it is unable
to process or interpret correctly—where the problems
exceed its requisite variety (e.g., Salovaara et al.,
2019). Next, we report on our case study.
3 The Case Study: Machine
Learning in a Governmental
Setting
To examine how an organization may tackle
explainability challenges, we conducted an exploratory
case study at a government agency that actively pursues
the deployment of AI via several ML projects. We
selected a case organization with both extensive capabil-
ities to develop AI/ML tools and a commitment to
accountability and explainability.
3.1 The Study Setting
The Danish Business Authority (DBA) is a government
entity operating under the Ministry of Industry,
Business, and Financial Affairs of Denmark. It has
approximately 700 employees and is based in
Copenhagen, with satellite departments in Silkeborg and
Nykøbing Falster. The authority is charged with a wide
array of core tasks related to business, clustered around
enhancing the potential for business growth throughout
Denmark. The DBA maintains the digital platform
VIRK, through which Danish companies can submit
business documents and that allows the DBA to
maintain an online business register (containing
approximately 809,000 companies, with roughly
812,000 registrations in all and together filing about
292,000 annual statements per year). The DBA has
maintenance and enforcement remits related to laws
such as Denmark’s Companies Act, Financial
Statements Act, Bookkeeping Act, and Act on
Commercial Foundations. In the past, the DBA also
collaborated with Early Warning Europe (EWE)a
network established to help companies and
entrepreneurs across Europeto produce support
mechanisms for companies in distress. The ML projects
analyzed in our study are related to the DBA’s core
tasksfor example, understanding VIRK users’
behavior and checking business registrations and annual
statements for mistakes and evidence of fraud.
The idea of using ML at the DBA originated in 2016.
The agency embarked on AI-related market research,
which culminated in several data-science projects and
the establishment of the Machine Learning Lab (“the
ML Lab” from here on) in 2017. One factor creating the
impetus for establishing the ML Lab was tremendous
growth in the quantities of various types of documents
processed by the DBA. Rather than engage and rely on
external consultants, the DBA opted to hire its own data
engineers and data scientists. The main reasons for this
in-house approach were cost-management concerns and
a desire to retain relevant knowledge within the agency.
Creating ML solutions internally by combining
technologies such as Neo4j graph database
management, Docker containers, and Python offers a
better fit for the organization than commercial
off-the-shelf solutions. Also, the ML Lab’s role is
Sociotechnical Envelopment of Artificial Intelligence
332
restricted largely to experimentation and development
surrounding proof-of-concept models. If a solution is
deemed useful and meets the quality criteria set, its
deployment is offloaded to external consulting firms,
which then put the model into production use. This
decision was primarily based on DBA culture, in which
vendors take responsibility for the support and
maintenance functions related to their code: the ML
models follow the same governance as other IT projects
within the DBA.
Hence, DBA operations related to ML are divided
between two main entities: a development unit (the ML
Lab) and an implementation unit (external consultants).
The ML Lab’s role is to collaborate closely with domain
experts (hereafter “case workers”) to develop functional
prototypes as part of a proof of concept. The lab’s main
objective is to prove that the problems identified by the
case workers can be solved by means of ML. In
combination, the proof of concept and documentation
such as the evaluation plan form the foundation for the
DBA steering committee’s decision-making on whether
to forward the model to the implementation unit.
Different stakeholders are accountable for different
parts of the process. The ML Lab is responsible for
developing the prototype, and the case workers provide
domain knowledge to the lab’s staff as that prototype is
developed. The case workers also answer for the ML
models’ operational correctness, being charged with
evaluating each model and with its retraining as needed.
The steering committee then decides which models will
enter production use and when. Finally, the implemen-
tation unit is accountable for implementing the model
and overseeing its technical maintenance.
3.2 Data Collection
Interviews and observations at the DBA served as our
main data sources. We used purposive sampling
(Bernard, 2017) and selected the case organization by
applying the following criteria. The organization needed
to have advanced AI and ML capabilities, in terms of
both resources and know-how. It also had to be
committed to developing explainable systems. Finally,
the researchers needed access to the AI/ML projects,
associated processes, and relevant stakeholders. The last
criterion was especially important for giving us a
broader perspective on the projects and for enabling the
verification of explainability claims made by the
informants. The DBA met all of these criteria.
To gain access to the DBA, we used the known-sponsor
approach (Patton, 2001): we had access to a senior
manager at the DBA working with ML initiatives within
the organization, who helped us arrange interviews at
the early stages of data collection. Piggybacking on that
manager’s legitimacy and credibility helped us establish
our legitimacy and credibility within the DBA from the
start (Patton, 2001). In addition, one of the authors had
a working relationship with the organization at the
operations level, allowing us to arrange interviews
further along in the data-collection work. This helped us
to establish mutual trust with the informants and
prevented us from being seen as agents of the upper
management.
We collected and analyzed data in a four-stage iterative
process (presented in Table 1), in which the phases
overlapped and earlier stages informed subsequent
stages. To prevent elite bias, we sought to interview a
wide range of DBA employees with varying tenure at
several levels in the hierarchy (Miles et al., 2014; Myers
& Newman, 2007). Phase 1 was explorative in nature.
Its purpose was to establish research collaboration and
create a picture of the DBA’s current and future ML
projects and visions from a data-science and case-work
perspective. The second phase was aimed at gaining in-
depth understanding of the DBA’s various ML projects
and the actors involved. In this phase, we focused on the
ML Lab and its roles and responsibilities in the projects,
along with explainability in relation to ML. Then, in
Phase 3, we interviewed all ML Lab employees as well
as two case workers who acted in close collaboration
with the lab. The final phase involved validating the
interpretations from our analysis and obtaining further
insight into the technical infrastructure supporting the
lab.
We conducted semi-structured interviews in all phases,
taking place from August 2018 to October 2020. Initial
impressions are important for establishing trust between
researchers and informants (Myers & Newman, 2007);
hence, we always presented ourselves as a team of
impartial researchers conducting an academic study. At
the start of each interview, we explained the overall
purpose of the study and our reasons for selecting the
informant(s) in question to participate. We promised
anonymity and confidentiality to all the informants and
asked for explicit consent to record the interviews. Also,
we explained the right to withdraw consent at any time
during the interview or after it, up to the time of the final
publication of a research article. We made sure to
address any concerns the informants expressed about the
procedure and answered all questions.
The interviews were conducted in English, with one of
the authors, a native Danish speaker, being present for
all of them and clarifying terminology as necessary. In
addition, the informants had the opportunity to speak
Danish if they so preferred. The choice of English as the
primary language was made in consideration of the fact
that most members of the research team did not speak
Danish, whereas all informants were highly proficient in
English. Though we recognize potential downsides to
conducting interviews in a language that is not native to
the interviewees, we accepted the remaining risk for the
sake of enabling the whole research team to be involved
in the data-collection process and data analysis. All
interviews were audio-recorded and transcribed,
yielding 167,006 words of text.
Journal of the Association for Information Systems
333
Table 1. The Four Phases of Gathering the Data
Phase number, theme,
and date range
Method and duration
Informants pseudonym and role
Focus of outcomes
1. ML projects overall,
August-September
2018
Group interview (105 minutes)
James (ML Lab team leader / chief
data scientist); Mary (chief
consultant)
Responsibilities of
the DBA;
organization
structure
2. ML Lab functions,
October 2018 to
January 2019
Personal interview (90 minutes)
James
The role of
explainability in ML
projects; allocation
of tasks among
stakeholders (the
ML Lab,
implementation unit,
and case workers)
Group interview (83 minutes)
David; John (both Early Warning
Europe external case workers)
Personal interview (70 minutes)
Daniel (an internal case worker)
Personal interview (59 minutes)
Steven (a data scientist at the ML
Lab)
Personal interview (51 minutes)
Mary
Personal interview (116 minutes)
James
3. Explainability in
ML projects,
September 2019
Personal interview (51 minutes)
Steven
Practical means to
address
explainability issues;
the sociotechnical
environment of
model development
Personal interview (54 minutes)
Thomas (a data scientist at the ML
Lab)
Personal interview (50 minutes)
Linda (a data scientist at the ML
Lab)
Personal interview (48 minutes)
Michael (a data scientist at the ML
Lab)
Personal interview (52 minutes)
Mark (a data scientist at the ML
Lab)
Personal interview (53 minutes)
Joseph (a data scientist at the ML
Lab)
Personal interview (54 minutes)
Jason (a team leader at the ML Lab)
Personal interview (48 minutes)
Susan (a data scientist at the ML
Lab)
Personal interview (62 minutes)
William (an internal case worker)
Personal interview (54 minutes)
Daniel
4. Verification of
interpretations from
analysis, December
2019 to October 2020
Personal interview (55 minutes)
Jason
Validation of
interpretations via
interview feedback
and an assessment
exercise involving
mapping via project
templates
Assessment exercise (time N/A)
Steven; Mary; Thomas; Linda;
Michael; Mark; Joseph; Jason;
Susan
Personal interview (27 minutes)
Jason
Personal interview (32 minutes)
Steven
Personal interview (49 minutes)
Daniel
In addition to interviews, we employed participant
observation and document analysis. Hand-written field
diaries kept by the Danish-speaking author provided
background information. These go back to September
2017, when he became involved with ML at the DBA.
Covering work as an external consultant and then a
collaborative PhD student funded equally by the IT
University of Copenhagen and the DBA, the diary
material comprises observations, task descriptions, and
notes taken at meetings. The diaries extended over the
full duration of our research period, including the time
when most ML projects were either very early in their
development or had not even begun. Accounting for
approximately every other workday at the DBA, the
doctoral student’s observations give a realistic view of
day-to-day work life at the case organization. We used
the field diaries for memory support, to fill gaps in the
interview data, and as a reference for basic information
about key informants, organization structure, and
organizational processes and work practices. In
addition, the diaries helped to corroborate some claims
made by informants. Similarly, the document analysis
addressed the entire time span of interest. This work
included analyzing documentation and user stories
extracted from the DBA’s Jira system, a project
management tool. The document analysis also
extended to accessing the DBA’s Git repository (used
in version control) and verifying which model was
Sociotechnical Envelopment of Artificial Intelligence
334
applied in each project. In addition, the collaborative
doctoral researcher had access to a personal email
account at the organization and could search old
conversations and start new ones if decisions made
during ML projects needed further explanation.
Finally, to verify the interpretations arising in the
course of the authors’ analysis, we asked the ML Lab
data scientists to fill in an outline document for each of
the ML projects alongside the authors in an assessment
exercise. This exercise produced an inputML-model
output framework that allowed us to verify the ML
projects’ fundamentals and establish uniform project
descriptions characterizing, for example, the data fed
into the model, the type of ML model employed, and
the nature of the output produced. Appendix A
provides a summary of this framework.
3.3 Data Analysis
Overall, our analysis approach can be considered
abductive: it began as inductive but was later informed
by a theoretical lens that emerged as a suitable
sensitizing device (Sarker et al., 2018; Tavory &
Timmermans, 2014). We coded all interview data in
three stages, utilizing coding and analysis techniques
adopted from less procedure-oriented versions of
grounded theory (Belgrave & Seide, 2019; Charmaz,
2006). In practice, this entailed relying on constant
comparative analysis to identify initial concepts. The
processes of data collection and analysis were
mutually integrated (Charmaz, 2006), constantly
taking us between the specific interview and the larger
context of the case organization (Klein & Myers,
1999). Later, we linked the emerging concepts to
higher-level categories. Similarities can be seen
between our approach to using elements of grounded
theory for qualitative data analysis and methods
established in earlier IS studies (e.g., Asatiani &
Penttinen, 2019; Sarker & Sarker, 2009).
The three stages of coding produced concepts (first-
order constructs), themes (second-order constructs),
and aggregate dimensions (see Appendix C), paralleling
the structure proposed by Gioia, Corley, and Hamilton
(2013). In the first stage, we performed open coding
with codes entirely grounded in our data. This involved
paragraph-by-paragraph coding, using in vivo codes
taken directly from the informants’ discourse
(Charmaz, 2006) with minimal interpretation by the
coders. For example, the extract: There would be a
guidance threshold. Actually, no. For this model, there
would be some guidance set by us, yeah. And then case
workers will be free to move it up and down” was
assigned two codes: “case workers’ control thresholds
and “guidance threshold. Two of the authors performed
open coding independently, after which the two sets of
codes were revisited, compared, and refined.
Conceptually similar codes were merged into the set of
concepts.
In the second stage, we analyzed the results from the
open coding and started to look for emerging themes.
We iterated between the open codes and interview
transcripts, coding data for broader themes connecting
several concepts (axial coding). While these themes
were at a higher level than the in vivo codes from the
first stage, they still were firmly grounded in the data.
All the authors participated in this stage, which
culminated in the codes identified being compared and
consolidated to yield the second-order constructsthe
themes.
In the third stage, we applied theoretical coding to our
data. That term notwithstanding, the goal for this stage
was not to validate a specific theory. Rather, we
wanted to systematize the DBA’s approaches to
tackling explainable AI challenges where building a
transparent system was not an option. For this, the
envelopment framework of Robbins (2020) served as
a sensitizing lens to help us organize the themes that
emerged in the second stage of analysis. The decision
was data-drivenwe had not anticipated finding such
strong focus on envelopment at the case organization,
but the first two stages of analysis inductively revealed
that the DBA’s strategy resembled an envelopment
rather than a method whereby the DBA would attempt
to guarantee explainability in all of its AI model
implementations. All authors participated in this stage
of the work, performing coding independently. Then,
the codes were compiled, compared, and synthesized
into a single code set.
4 Findings
Our findings draw from the DBA ML Lab’s work in
eight AI projects, denoted here as Auditor’s Statement,
Bankruptcy, Company Registration, Land and
Buildings, ID Verification, Recommendation, Sector
Code, and Signature (see Appendix A for project
details). While every project had a distinct purpose, each
was aimed at supporting the DBAs role in society as a
government business authority. At the time of writing
this paper, many of these projects had been deployed
and entered continuous use. The DBA had faced
pressure to be highly efficient while remaining a
transparent and trustworthy actor in the eyes of the
public, and AI-based tools represented an efficient
alternative to the extremely resource-intensive fully
human-based processing of data. At the same time, the
use of such tools presented a risk of coming into conflict
with the DBA’s responsibility to be transparent. To
situate the set of envelopment methods employed by the
DBA in this context, we begin by analyzing the DBA’s
viewpoint on requirements for the AI systems to be used
in the agency’s operations. This sets the stage for
discussing the envelopment methods that the DBA
developed to address the challenges of the
explainability-accuracy tradeoff (see Figure 1)
introduced by its development of ML solutions.
Journal of the Association for Information Systems
335
4.1 Requirements for AI at the DBA
Our interviews showed that, given the drive to improve
its operations by using AI models, the DBA must
devote significant attention to making sure
instrumental outcomes do not come bundled with
ignoring humanistic ones. Two factors have shaped the
organization’s quest to find balance in terms of the
explainability-accuracy tradeoff: its positions as a
public agency and diverse stakeholder requirements.
First, as a public agency, the DBA has significant
responsibility for making sure that its decisions are as
fair and bias-free as possible. Recent discussion
surrounding regulations such as the GDPR has brought
further attention to the handling of personal data and to
citizens’ rights to explanation. These reasons have
impelled the DBA to be sure that the organization’s
ML solutions respond to explanability requirements
sufficiently. This comment from a chief consultant on
the DBA annual statements team, Mary, addresses
transparency’s importance:
I think in Denmark, generally, we have a lot
of trust towards systems …. I’m very fond of
transparency. I think it’s the way to go that
it’s fully disclosed why a system reacts [the
way] it does. Otherwise, you will feel unsafe
about why the system makes the decisions it
does For me, it’s very important that it’s
not a black box.
Still, the DBA has ample opportunities to benefit from
deploying AI in its operations, in that it has access to
vast volumes of data and boasts proactive case workers
who are able to identify relevant tasks for the AI.
Sometimes inscrutable models clearly outperform
explainable ones, so the agency has a strong incentive
to seek ways of expanding the range of AI models that
are feasible for its operations, in pursuit of higher
accuracy and better performance. However, it needs to
do so without incurring excessive risks associated with
inscrutable models:
If the output of the algorithm is very bad
when using the [explainable] models and
we see a performance boost in more
advanced or black-box algorithms, we will
use [the more advanced ones]. Then, we
will afterwards check like “okay, how to
make this transparent, how to make this
explainable…” (Steven, ML Lab)
Secondly, the quest for explainable AI is made even
more complex by the diversity of explanation-related
requirements among various DBA stakeholders. The
internal stakeholders comprise several distinct
employee categories, including managers, data
scientists, system developers, and case workers.
Externally, the DBA interacts with citizens and the
companies registered in Denmark, as well as with the
IT consulting firms that maintain the agency’s AI
models deployed in the production environment.
Each of these stakeholders requires a specific kind of
explanation of a given model’s internal logic and
outputs. While an expert may consider it helpful to
have a particular sort of explanation for the logic
behind the model’s behavior, that explanation may be
useless to someone who is not an expert user. For a
nonexpert user, a concise, directed, and even partially
nontransparent explanation may have more value than
a precise technical account. David, a case worker with
Early Warning Europe, offered an example: “When [a
data scientist] explained this to us, of course it was like
the teacher explaining brain surgery to a group of
five-year-olds.”
These two factors together explain why expanding the
scope of candidate models can pose problems even if
more accurate models are available and technically
able to be brought into use. Because of the different
stakeholders’ various needs, a suitable level of
explainability is hard to reach. Therefore, approaches
that could broaden the range of modelsvisualized as
a circle with a dashed outline in Figure 1are sorely
needed.
Our findings indicate that envelopment offers a
potential solution to the explainability-accuracy
tradeoff. With a variety of envelopment methods, the
risks of inscrutable AI may be controlled in a manner
that is acceptable to the different stakeholders, even
when technical explanations are not available. As
Steven stated:
Often, we [are] able to unpack the black box
if necessary and unpack it in a way that
would be more than good enough for our
case workers to understand and to use it
and also for us to explain how the model
came to the decision it did.
Next, we discuss how the DBA has succeeded in this
by enveloping its AI systems’ boundaries, training
data, and input and output data. We then consider our
findings with regard to the connection between the
choice of AI model and envelopment.
4.2 Boundary Envelopment
The notion of boundary envelopment suggests that an
AI agent’s limits can be bounded by well-defined
principles that demarcate the environment within
which it is allowed to process data and make decisions.
One example of boundary envelopment at the DBA is
the document filter implemented in the Signature
project. It filters out images that are not photographs of
a paper document. The need for such a filter was
identified when an external evaluator tested the model
with a picture of a wooden toy animal and the model
judged the image to be a signed document because it
Sociotechnical Envelopment of Artificial Intelligence
336
was operating beyond its intended environment.
Having not been trained to analyze images other than
scans and photographs of black-and-white documents,
the model returned unpredictable answers. By limiting
the types of input images to ones that the model had
been trained to recognize, the filter created in response
acts as a boundary envelope guaranteeing the requisite
variety for the AI model that constitutes the next
element in the information-processing pipeline. Thus,
the AI model was enveloped in two ways: technically,
via the development of a filter for its input data, and
socially, via a change in workflow, whereby
documents now undergo screening before they are
assessed for completeness.
Both social and technical dimensions of envelopment
were evident also in other instances at the case
organization. The following quotes exemplify how the
DBA orchestrates its AI agents’ boundary-creation
work and makes sure that its AI solutions speak to very
different stakeholders’ concerns. To ensure that AI
systems’ abilities and limitations are controlled and
therefore enveloped, the DBA decided to divide its AI
development into a process of incremental stages by
introducing multiple small-scale solutions, each
dedicated to a certain set of relatively simple and well-
defined actions. The following comment summarizes
this method:
Well, I’m working at an organization
where, luckily, the management wants us to
develop results fast or fail fast, so they are
happy with having small solutions put into
production [use] rather than having large
projects fail …. We decided to use an event-
driven architecture, because when dealing
with complex systems, it’s better to allow an
ordered chaos than try to have a chaotic
order. By having an event-driven
architecture, you can rely on loosely coupled
systems, and by having sound metadata it
will help you create order in the chaos of
different systems interacting with the same
data. (Jason, ML Lab)
Thus, from a purely technical angle, the event-driven
architecture and loosely coupled systems constitute a
technique in which the various components of a larger
architecture operate autonomously and malfunctions
are limited to local impacts only. For instance,
erroneous decisions are less likely to be passed onward
to other systems, and if this somehow does occur, the
loose coupling allows the DBA to rapidly curb the
failure’s escalation. Each component is therefore
operating in its own envelope, and larger envelopes are
created to control AI components’ operation as a
network.
However, as highlighted by the reference above to
envelopes that meet various stakeholders’ needs,
boundary envelopes do not serve a technical purpose
alone. The following extract from the data shows how
important the understanding of these boundaries is for
those human stakeholders that are tasked with judging
the correctness of the model’s operation when, for
example, the complexity of the environment exceeds
the model’s comprehension capability:
We have around 160 rules. We have
technical rules that look into whether the
right taxonomy is being used, whether it is
the XBRL format, and whether it is
compliant. We also have business rules. For
example, do assets and liabilities match?
Some rules only look at technical issues in
the instance report. Other rules are what we
called full-stop rules filers are not allowed
to file the report until they have corrected the
error. We also have more guidance[-type]
rules, where we say, “It looks like you’re
about to make a mistake. Most people do it
this way. Are you sure you want to continue
filing the report?” And then [users] can
choose whether to ignore the rule [or not].
(Mary)
In addition to the technical issues connected with
accounting for multiple kinds of failure, the comment
attests to boundary envelopes’ social dimension. The
boundaries are clearly explained to internal users at the
DBA, who can overrule the models if necessary.
Moreover, customer-facing models operate within an
environment that has clearly defined rules constraining
their operation. Wherever nonexpert employees
interact directly with a model, these rules are explained
to them, and the human always has the power to ignore
the models’ recommendations if they seem
questionable.
Thus, importantly, for every customer-facing AI model
at the DBA, the final boundary envelope is a human. A
decision suggested by an AI model is always verified
by a case worker. In simple terms, human rationality
creates a boundary that envelops the model’s
operation. This serves a dual purpose: it denies any
model the power to make unsupervised decisions while
it also makes certain that every DBA decision is
compliant with legal requirements. According to
Jason:
The agency can be taken into court when we
dissolve a company, when we end a
company [forceably] by means of the law.
And we, in that situation, in court, will have
to provide … full documentation of why that
decision has been made. Now, legally
speaking, as soon as there’s a human
involved, as there always is, we always keep
a human in [the] loop, [so we are on the
safe side]. In that context, it’s only legally
Journal of the Association for Information Systems
337
necessary to present that human’s decision.
But we want to be able to explain also
decision support, so that’s why we need
explainability in our model and information
chain. Explainability, on the microscale, is
beneficial to understanding [the]
organization on a sort of macroscale.
In other instances, expert case workers are allowed to
set thresholds for the model in question, to make
certain it produces the most useful and precise
recommendations. This has a knock-on effect in
facilitating DBA workers acceptance of the relevant
model:
For some [of our] models, there would be
some guidance threshold set by us. And then
case workers are free to move it up and
down. (Susan, ML Lab)
The ability to “mute” a model or change the
threshold has been a major cultural factor
in [the] business adaptation of this
technology. (Jason)
In summary, envelopment of boundaries involves both
resolving technical issues (understanding the limits of
the model’s abilities, etc.) and addressing social factors
(providing the various stakeholders with sufficient
explainability and, thereby, affording trust in the
model’s accuracy, etc.).
4.3 Training-Data Envelopment
The crucial importance of the data used in AI systems’
training is widely acknowledged in the AI/ML
community. If trained on different data sets, two
models with otherwise identical structure produce
vastly different outputs (Alpaydin, 2020; Robbins,
2020). Accordingly, close control of the training data
and the training process form an important aspect of
envelopment: if the spectrum of phenomena that the
training data represent is considered with care, one can
better understand what the model willand will not
be able to interpret.
Since the DBA wants to avoid any undesired outcomes
from an uncontrolled model roaming freely on a sea of
potentially biased training data, the organization has
decided to maintain full control over the learning
process; thus, it abstains from using online-learning
models, which continue learning autonomously from
incoming data. This aids the DBA in protecting its
systems from the unintended overfitting and bias that
less tightly controlled training data could more easily
introduce. The training may be implemented in a
controlled, stepwise manner:
We have taken a conscious decision not to
use [online-learning] technologies,
meaning that we train a model to a certain
level and then we accept that it will not
become smart until we retrain it. (Jason,
ML Lab)
Avoidance of models that learn “on the fly” has a
downside in that models’ training at the DBA is a
highly involved periodic process that requires human
expertise. Successful training-data envelopment
therefore entails numerous stakeholders at the agency
cooperating periodically to assess the needs for
retraining and to perform that retraining. Paying
attention to training data stimulates internal discussion
of the data’s suitability and of possible improvements
in detecting problematic cases that are flagged for
manual processing.
To plan retraining appropriately, data scientists at the
ML Lab communicate with case workers regularly
with regard to analyzing the models’ performance and
new kinds of incoming data. Though time-consuming,
this process supports employees’ mutual
understanding of how the models arrive at specific
results. A case worker described the effect as follows:
I’m not that technically [grounded a] person,
but doing thattraining the model and
seeing what output actually came out from
me training the model…made my
understanding of it a lot better. (William,
Company Registration)
Through interaction during the retraining steps, the
stakeholders gain greater appreciation of each other’s
needs:
In the company team, we would very much
like [a model that] tells us, “Look at these
areas,” areas we didn’t even think about:
“Look at these because we can see there is
something rotten going on here,” basically.
Other control departments would rather
say, “We have seen one case that look[s]
like this; there were these eight things
wrong. Dear machine, find me cases that
are exactly the same.And we have tried
many times to tell them that that’s fine. We
had a case years ago where there were a lot
of bakeries that did a lot of fraud, but now
it doesn’t make sense to look for bakeries
anymore, because now these bakeries …
are selling flowers or making computers or
something different. (Daniel, Company
Registration)
In summary, training-data envelopment involves social
effort in tandem with the purely technical endeavor of
preparing suitable input-output mappings in machine-
readable form that the AI can then be tasked with
learning. For the training-data envelopment to succeed,
the screening and ongoing monitoring of a model’s
performance requires the cooperation of many different
stakeholders. Only this can guarantee that biases and
Sociotechnical Envelopment of Artificial Intelligence
338
other deficiencies in the data are reducedand that the
model remains up to date. Otherwise, as the environment
changes around the model, its boundary envelope
becomes outdated. Training-data envelopment helps
address this alongside issues of bias.
4.4 Input and Output Envelopment
Input and output determine, respectively, what data
sources are used to create predictions and what types of
decisions, classifications, or actions are created as the
model’s output. Any potential inputs and outputs that
exhibit considerable noise, risk of bias, data omissions,
or other problems are enveloped out of the AI’s
operation through these decisions. The selection of input
sources is thus closely tied to conceptions of data
quality. In the concrete case of the ID-recognition model
PassportEye, the benefits of input control in conditions
of poor and variable end-user-generated content became
clear to the lab’s staff:
I think our main problem was that, yeah, we
had to go a little bit back and forth because
the input data was [of] very varied quality.
Mostly low quality. Out of the box,
PassportEye actually returned very bad
results, and that reflects the low quality of the
input data, because people just take pictures
in whatever lighting, [against] whatever
background, and so on. So we actually
figured out a way to rotate the images back
and forth to get a more reliable result.
Because, it turned out, PassportEye was
quite sensitive to angle of an image. We
didn’t write it [the image analysis software],
so this is maybe one of the risky parts when
you just import a library instead of writing it
yourself. (Thomas, ML Lab)
As for output envelopment, the interplay between social
and technical is more prominent here. Instead of simply
preventing production of outputs that may be
untrustworthy, the DBA takes a more nuanced
approach. Output of appropriate confidence ratings and
intervals from the models is a subject of active
deliberation at the DBA. Estimates such as probabilities
that a financial document is signed are important for the
agency’s case workers, who need them for identifying
problematic cases. When an AI model yields a clearly
specified and understandable confidence value, the case
worker’s attention can be rapidly drawn to the model’s
output as necessary:
If there’s no signature, [the case workers]
will simply reject it. Because the law says this
document has to be signed, so the human will
look at the papers and say, “It’s not here.
You will not get your VAT number, or your
business number, because you didn’t sign the
document.(James, ML Lab)
When able to verify judgments on the basis of
confidence ratings, the case worker can act in an
accountable manner in the interactions with DBA clients
(e.g., companies that have submitted documents) and
respond convincingly to their inquiries. As Steven
explained:
If a person calls and asks, “Why was my
document rejected?” then a case worker will
say, “That’s because you have not signed it.”
“How do you know that?” “I have looked at
the document. It is not signed.” So they don’t
have to answer, “Well, the neural network
said it because of a variable 644 in the
corner.” That’s why you can get away [with]
using a neural network in this case,
regardless of explainability.
However, sometimes it is trickier to verify the model’s
output unequivocally, in which case the organization
strives to understand the AI model’s behavior by
consulting domain experts who understand the social
context of the model’s output. As Steven put it, “When
[it is] harder to determine if the model is right or wrong,
we can push the cases to the case workers and say,
‘Please look at this.’”
These examples of input and output envelopment
demonstrate clear interplay between the social and the
technical. While an opaque model is able to process a
large quantity of unstructured data efficiently and
produce recommendations on whether to accept or reject
particular documents, this process is closely guided by
case workers who rely on organizational objectives and
legislative limitations to be sure the AI-produced
decisions are in line with their needs. Thus, final
decisions are produced at the intersection of actions by
humans and AI.
4.5 The Implications of Envelopment for
Model Choice
Having demonstrated the use of several envelopment
methods in concert at the DBA, we now turn to their
implications for the choice of a suitable AI model.
Overall, the adoption of envelopment practices has
enabled the DBA to use models that could otherwise
pose risks. Different AI models are based on different
architectures, which has ramifications for what the
models can and cannot do. Models differ in, for
example, their maturity, robustness to noise, ability to
unlearn and be retrained quickly, and scalability. These
qualities are dependent on the choice of the model
type. For instance, robustness against noise is often
easier to achieve with neural networks, while abilities
of quick unlearning and retraining may be more rapidly
exploited with decision trees. Depending on the needs
for accuracy and/or explainability associated with a
given model type, alongside the use case, suitably
chosen envelopment methods can be implemented as
Journal of the Association for Information Systems
339
layers that together guarantee safe and predictable
operation.
Boundary envelopment has given the DBA more
degrees of freedom in choosing its models by limiting
the AI agent’s sphere of influence. This has allowed
the staff to take advantage of complex models that,
were it not for envelopment, could be rendered
problematic by their lack of explainability. Jason
characterized this as follows: “You can sort of say
we’re feeding the dragon, organization-wise, with one
little biscuit at a time, so we can produce models that
can be brought into production and are indeed put into
production.” In this way, human agents adjust the
organization’s processes and structures in order to
contain the technological agent’s operations safely.
Similarly, understanding and controlling data through
training-data and input-data envelopment combined
guarantee that the model’s behavior is within safe
limits and that the DBA possesses sufficient
understanding of how the outputs are generated, even
in the absence of full technical traceability. As James
at the ML Lab mused:
Here’s a new data set. What can we say
about it? What should we be aware of?
That’s becoming increasingly important
also as we are using more data connected
to people’s individual income, which is
secret in Denmark …. Our experience with
the initial use of the model has
emphasized that this model and the data it
[encompasses] needs some additional
governance to safeguard that we’re not
going outside our initial intentions … We’ve
revisited some of the metadata handling
that’s built into the platform to ensure that
we get the necessary data about how the
model behaves in relation to this case
handling so we can survey model output.
With regard to output, provided that a human is able to
judge its validity, one can easily opt for black-boxed
models that yield superior performance. The following
comment by James demonstrates how exercising
output control has enabled the use of an inscrutable
model: “I don’t have to be able to explain how I got to
the result in cases such as identifying a signature on a
paper. You can just do deep learning because it’s easy
to verify by a human afterward.”
The interviews illustrate how a need for new models
may arise in response to new legislative initiatives, a
new organizational strategy, or changes in taxpayer
behavior. An incumbent model may have to be
retrained or even entirely overhauled if metrics for
accuracy or explainability indicate that it is no longer
performing satisfactorily (e.g., its classifications are no
longer accurate or they start leading to nonsensical
estimates that cannot be explained). James gave an
example illustrating the use of a boundary envelope to
“mute” a model in such a case while it was directed to
retraining or replacement: “The caseworkers found
that the output of the model was not of quality that they
could use to anything, so they muted the model. That
comes back to us. We take the model down. Retrain
it….” Through this process, humans decreased the
AI’s agency in the work process by muting it and
renegotiating its agency via retraining or replacement.
4.6 Summary
The concept of envelopment has helped us flesh out
our view of the conceptual and practical mechanisms
of countering challenges posed by inscrutable AI. The
subsections above provide empirical evidence for
several distinct envelopment methods in an
organizational setting. It is worth noting that, while we
found evidence of the DBA actively applying
boundary, training-data, and input- and output-data
envelopment, we did not observe discussions about the
last of the five envelopment methods listed by Robbins
(2019): function envelopment, which the reader may
recall refers to deciding that an AI agent will not be
used for certain purposes even though it could do so
accurately. Behind this decision may be ethics
considerations, for instance. We believe that the lack
of discussion of topics related to function envelopment
at the DBA can be explained by the goals for each
system having already been narrowly specified based
on government regulations for every process.
We summarize the findings as follows. Considering,
first, that the DBA has been able to implement several
AI-based solutions successfully in its operations and,
second, the evidence of envelopment in the DBA’s
practices (both in general and pertaining to the various
methods), the concept of envelopment appears to
effectively capture some of the ways in which the
explainability-accuracy tradeoff presented in Figure 1
can be managed in AI implementation. Specifically,
our findings indicate that, although envelopment does
not change the relationship between accuracy and
explainability, it allows organizations to choose from a
wider range of AI models without facing an
insurmountable risk of harmful consequences (e.g.,
wildly unpredictable outcomes). Envelopment can
permit an organization to compromise some
explainability for the sake of greater accuracy without
needing to worry, as long as this takes place within
some limits of predictable behavior. The principal
benefit of envelopment is depicted in Figure 3 below.
Sociotechnical Envelopment of Artificial Intelligence
340
Figure 3. How Envelopment Expands the Set of Models an Organization May Adopt Without Excessive Risks
Second, in terms of the sociotechnical perspective,
regardless of which envelopment method they were
discussing, the interviewees never spoke of a purely
technical solution for limiting AI agents’ capabilities.
Analysis revealed that, rather than in isolation, such
actions were always carried out via iterative
negotiations that took into account several stakeholder
views, responsibility to society, and particular
implications for the personnel’s work processes.
5 Discussion
In this research, we asked: How can an organization
exploit inscrutable AI systems in a safe and socially
responsible manner? We sought answers to this
question by conducting a case study of a publicly funded
organization that regularly deploys AI to improve its
operations, which are of importance for society. As
described above, the study and analysis of the results
built on the concept of envelopment as a possible
approach to balancing accuracy with explainability and
finding good harmony between efficiency and safety.
The analysis presented above clearly identified three
significant findings. First, the case study showed that
AI’s envelopment, as a concept, holds empirical validity
in an organizational knowledge-work setting. This
complements prior envelopment literature (see Floridi,
2011; Robbins, 2020), which is of a purely conceptual
nature. Second, we demonstrated that envelopment is far
more than a technical matterto be effective, it has to
be situated at the intersection of the technical and the
social. Our study showed how social factors pervade all
aspects of envelopment and that human agents are an
integral part of envelopment, responsible for defining
suitable envelopes as well as maintaining and
renegotiating them. Finally, the analysis articulated
connections between envelopment methods and the
choice of ML model. Together, these findings
demonstrate the utility of envelopmentsociotechnical
envelopment in particularas an approach to
understanding the ways in which AI’s role in an
organization can be conceptualized and the ways in
which its responsibilities can be defined and managed.
We discuss specific implications for theory and practice
next.
5.1 Implications for Theory
Attending to the considerations described above allows
for deeper sociotechnical discussion of enveloping AI,
anchored in the DBA case as an example. This is
possible via synthesis of prior literature and our
empirical results. Sarker et al.’s (2019) review of
sociotechnical approaches in IS research, discussed
near the beginning of this paper, warns that today’s IS
work is in danger of too often being focused on
technologies’ instrumental outcomes, since they are
easier to measure and evaluate. Sarker and colleagues
suggest that sociotechnically oriented IS scholars
would do well to address both the instrumental and
humanistic outcomes of systems.
Journal of the Association for Information Systems
341
In the case of the DBA, any given AI deployment’s
possible instrumental outcomes would indeed be easier
to analyze and declare than its humanistic outcomes,
since they tie in with typical reasons for automating
processes, such as aims of increased efficiency and
higher precision. However, we saw that such
instrumental outcomes are not the only consideration at
the DBA: it was deemed crucial that AI projects not lead
to misuses of government power or unnecessary
profiling/surveillance of either citizens or private
enterprises. Such outcomes would be problematic from
a humanistic perspective and would compromise the
organization’s integrity as a public authority, potentially
introducing erosion of public trust. Moreover, AI
projects have humanistic outcomes even internally to
the DBA. They expand case workers’ opportunities to
redesign their work processesin fact, most of the
agency’s projects are undertaken in light of their
proposalsand case workers are also directly involved
in AI development processes. This serves to increase
workplace democracy, empowerment, and occupational
well-being. The DBA’s AI envelopment is clearly a
sociotechnical process: the technical specification of
limits for AI’s operations takes place via a social process
wherein the case workers and other stakeholders are
central actors.
The fact that the DBA’s AI development is typically
triggered by case workers suggests that the organization
has adopted an emergent mode of operation. Case
workers identify practical domain problems for the ML
Lab to work on and they also participate in the AI
models’ development. In the search for a suitable model,
ML experts and case workers analyze the capabilities
and constraints entailed by various ML models, then
match them interactively with the properties of the
problems to be solved. When suitable models are not
found for the problem at hand, the problem is broken
into an alternative structure. Another approach, in such
cases, is to adapt the case workers’ role in resolution to
mesh with the AI system’s capabilities.
We propose theoretical implications for (1) describing
organizational AI implementation as a balancing act
between human and AI agency, and (2) conceptualizing
sociotechnical envelopment as the primary tool for this
crucial balancing act. Addressing the first implication
builds on considering how AI development processes
consist of action sequences in which case workers and
AI systems, as partnered agents, carry out tasks together.
The desired level of agency (that is, a suitable balance
between humans and AI systems) is determined in the
course of developing models and governed by the
capabilities and constraints of the possible AI solutions.
AI technologies’ powerful information-processing
1
For more detailed managerial recommendations based on
the case of the DBA please refer to Asatiani et al. (2020).
capabilities offer an abundance of opportunities for
numerous kinds of implementation (Kaplan & Haenlein,
2019). At the same time, thanks to ready availability of
scalable computing resources, AI places few constraints
on data-processing capacity (Lindebaum et al., 2020).
Therefore, there are multitudes of possibilities for using
such technology. However, because of the complexity
of many AI models, the technology presents constraints
with regard to its ability to provide technical
explanations for its workings. Therefore, AI’s potential
still must be curbed appropriately: for example, it is
necessary to find an acceptable explainability-accuracy
tradeoff and, to this end, one must also establish the
required level of meaningful explainability for a given
context (Ribera & Lapedriza, 2019; Robbins, 2019),
which takes place via negotiations across the agency
among social actors. Hence, AI implementations tend to
involve a balancing act between human and AI agency
to arrive at a suitable level of agency for the AI. In this
context, the power balance between the two parties is
more equal than in many other human-technology
relationships (e.g., implementing enterprise resource
planning systems) in which the technology’s workings
are known and its capabilities seem less likely to
represent unexpected negative consequences for
stakeholders.
This discussion leads us to the second implication:
conceptualization of sociotechnical envelopment.
Two-pronged envelopment of this nature emphasizes
the social dimension that is missing from existing
envelopment literature (Floridi, 2011; Robbins, 2020)
by focusing on the interaction of human and AI
agencies, instead of on merely limiting or adjusting an
AI system’s capabilities. In doing so, we have been able
to extend discussion on envelopment by revealing how
envelopes can be constructed and maintained in a
sociotechnical setting. We posit that this sociotechnical
view of envelopment may offer a powerful tool for
success in the balancing act between human and AI
agency by offering a rich mechanism through which AI
capabilities can be curbed in settings where ethics,
safety, and accountability are vital to operations. This
should help to offset the impact of uncertainty
introduced by the inscrutability of AI and thus allow
organizations to obtain efficiency gains from AI systems
that offer powerful capabilities but lack explainability.
5.2 Practical Implications
For managers, whose expertise often lies in managing
humans rather than AI agents, the envelopment methods
presented and illustrated in this paper offer a suitable
vocabulary and toolbox for handling AI development.
1
Through a process of analyzing the risks a given AI
Sociotechnical Envelopment of Artificial Intelligence
342
solution creates for business, ethics, consumer rights
(e.g., the right to explanation), and environmental
safety, a manager may be able to apprehend the
organization’s needs for envelopment. On this basis,
sociotechnical approaches may be implemented and
aligned with operations management and AI solution
development, all in a manner that renders the models
more understandable to stakeholders and addresses AI
interpretability needs specific to data scientists.
A word of caution is crucial, however. Even in the
presence of envelopment, one should not accept black-
box models without having devoted significant effort
to finding interpretable models. While a black-box
model may initially appear to be the only alternative,
there are good reasons to believe that accurate yet
interpretable models may exist in many more domains
than now recognized. Identifying such models offers
greater benefit than does the sociotechnical
envelopment of a black-box model. For every decision
problem involving uncertainty and a limited training
data set, numerous nearly optimal, reasonably accurate
predictive models usually can be identified. This
assertion stems from the so-called Rashomon set
argument (Rudin, 2019), under which there is a good
chance that at least one of the acceptable models is
interpretable yet still accurate. Another recommended
approach that simplifies envelopment is to strive for
“gray-box models,” as exemplified by the creation of
“digital twins” that can simulate real, physical
processes (see El Saddik, 2018; Kritzinger et al.,
2018). Gray-box ML solutions are modeled in line
with laws, theories, and principles known to hold in the
given domain. For example, such an approach can
establish a structure for a neural network, whereupon
the free parameters can be trained more quickly to
achieve high performance, without any reduction in
explainability.
Another practical benefit of adopting envelopment as
a tool for AI implementation is its relationship to
technical debt. In an AI context, at least two kinds of
debt can be identified. The first is related to selecting
models that do not offer the best accuracy for the
problems at hand (Cunningham, 1992; Kruchten et al,.
2012), as occurs if an organization needs to ensure
explainability in its implementation. The other source,
connected with documentation, applies to software
development in general: organizations may decide to
expedite their implementation efforts if they decide to
relax the requirements for documenting their decisions
and code (see Allman, 2012; Rolland et al., 2018). This
may backfire if employee turnover rears its head and
no one remains who can explain the underlying logic
of the AI system. After all, answers only exist in
individuals’ heads or buried in code.
Envelopment may offer a means to address both types
of debt: debt resulting from risk-averse choices in AI
implementation that lag behind the problem’s
development, and debt occurring because of decisions
to relax documentation requirements. Since
envelopment involves carefully making and
documenting decisions, it may serve as a practice
whereby design decisions are rendered explicit; for
example, implicit assumptions about the problem and
model may be recorded. Envelopment, therefore, not
only supports documentation but, by enabling the use of
more accurate models, it can also decrease the
accumulation of technical debt rooted in a conservative
model-choice strategy.
5.3 Limitations and Further Research
Our research has some limitations. First, we used
purposive sampling and studied a government unit as
our empirical case since we presumed it would provide
an empirically rich setting for gathering data on the use
of AI. This choice, while supplying ample evidence of
the envelopment strategies employed, did restrict us to
studying such strategies in the specific setting of a
public organization. Further research could examine
envelopment of AI in a larger variety of contexts. For
example, private firms driven by differently weighted
objectives might use other types of envelopment
strategies or employ the ones we studied in different
ways. Moreover, our study did not find evidence
pertaining to function envelopmentlikely because
the purposes of AI’s use at the DBA are already strictly
mandated by laws and regulations. Indeed, there was
seldom reason to discuss whether the DBA’s AI
solutions should be applied to purposes for which they
were never designed. Second, while our access to the
case organization permitted in-depth analysis of the
envelopment strategies applied, we could not examine
their long-term implications. Further research is
needed to probe the impacts of these envelopment
strategies over time. Finally, while we were granted
generous access for conducting interviews and
analyzing secondary material, our corpus of interview
data is naturally limited to what the informants
expressed. To mitigate the risks associated with
informant bias, we strove to obtain multiple views on
all critical pieces of evidence associated with
envelopment strategies. For example, we interviewed
every employee working at the DBA’s ML Lab, with
the aim of harnessing several perspectives on each
project.
With regard to both the utility of this paper and
outgrowths of the efforts presented here, we wish to
emphasize the value of developing a fuller
understanding of the various methods by which AI and
ML solutions can be controlled in order to harness the
strengths they bring to the table. Envelopment
strategies and their deeper examination can offer a
practical means toward this end. Although the
application of envelopment at the DBA was not
grounded in the literature conceptualizing these
Journal of the Association for Information Systems
343
practices (e.g., Floridi, 2011; Robbins, 2020), given
DBA developers’ awareness of this prior work, more
informed harvesting of the methods’ potential could
follow. Alongside such opportunities, future research
could investigate whether the dynamics between
humans and AI agents discussed here carry over to
contexts other than AI implementation. We believe
that similar logic might be identifiable, albeit in
different forms, in other contexts where safe, ethical,
and accountable IS implementation is crucial.
6 Conclusion
We find considerable promise in our definition and
operationalization of sociotechnical envelopment in an
organizational context. The findings shed light on
specific instances of envelopment and they aid in
identifying particular socially and technically oriented
approaches to envelopment. We have been able to
offer, as a starting point, a tantalizing glimpse of the
capabilities and limitations of various sociotechnical
envelopment approaches for addressing issues related
to the safer use of AI for human good.
Acknowledgments
We are grateful to the Danish Business Authority and
Early Warning Europe for the opportunity to conduct
this study. We wish to thank the special issue editors
and three anonymous reviewers whose insightful
comments and constructive criticism helped us to
greatly improve the quality of our paper. We also thank
the roundtable participants at the ICIS 2019
JAIS/MISQE Special Issue Session for their feedback
on our project proposal. Naturally, all remaining errors
are ours.
Sociotechnical Envelopment of Artificial Intelligence
344
References
Adiwardana, D., Luong, M.-T., So, D. R., Hall, J.,
Fiedel, N., Thoppilan, R., Yang, Z.,
Kulshreshtha, A., Nemade, G., Lu, Y., et al.
(2020). Towards a human-like open-domain
chatbot. https://arxiv.org/pdf/2001.09977v1.pdf.
Ågerfalk, P. J. (2020). Artificial intelligence as digital
agency. European Journal of Information
Systems, 29(1), 1-8.
Allman, E. (2012). Managing technical debt.
Communications of the ACM, 55(5), 50-55.
Alpaydin, E. (2020. Introduction to Machine Learning,
(4th ed.). MIT Press.
Asatiani, A., Malo, P., Nagbøl, P. R., Penttinen, E.,
Rinta-Kahila, T., & Salovaara, A. (2020).
Challenges of explaining the behavior of black-
box AI systems. MIS Quarterly Executive,
19(4), 259-278.
Asatiani, A., & Penttinen, E. (2019). Constructing
continuities in virtual work environments: A
multiple case study of two firms with differing
degrees of virtuality. Information Systems
Journal, 29(2), 484-513.
Asatiani, A., Penttinen, E., Rinta-Kahila, T., &
Salovaara, A. (2019). Implementation of
automation as distributed cognition in
knowledge work organizations: Six
recommendations for managers. Proceedings of
the 40th International Conference on
Information Systems.
Ashby, W. R. (1958). Requisite variety and its
implications for the control of complex
systems. Cybernetica, 1(2), 83-99.
Belgrave, L. L., & Seide, K. (2019). Coding for
grounded theory. In A. Bryant and K. Charmaz
(eds.), The SAGE Handbook of Current
Developments in Grounded Theory, (pp. 167-
185). SAGE.
Benbya, H., Davenport, T. H., & Pachidi, S. (2020).
Special issue editorial: Artificial intelligence in
organizations: Current state and future
opportunities. MIS Quarterly Executive, 19(4),
ix-xxi.
Benbya, H., & McKelvey, B. (2006). Using
coevolutionary and complexity theories to
improve IS alignment: A multi-level approach.
Journal of Information Technology, 21(4),
Springer, 284-298.
Bernard, H. R. (2017). Research methods in
anthropology: Qualitative and quantitative
approaches. Rowman & Littlefield.
Bostrom, R., Gupta, S., & Thomas, D. (2009). A meta-
theory for understanding information systems
within sociotechnical systems. Journal of
Management Information Systems, 26(1) 17-
48.
Briggs, R. O., Nunamaker, J. F., & Sprague, R. H.
(2010). Special section: Social aspects of
sociotechnical systems. Journal of
Management Information Systems, 27(1), 13-
16.
Brynjolfsson, E., & McAfee, A. (2014). The second
machine age: Work, progress, and prosperity in
a time of brilliant technologies. Norton.
Burrell, J. (2016). How the machine thinks:
Understanding opacity in machine learning
algorithms. Big Data and Society, 3(1), 1-12.
Butler, B. S., & Gray, P. H. (2006). Reliability,
mindfulness and information systems. MIS
Quarterly, 30(2), 211-224.
Charmaz, K. (2006). Constructing grounded theory: A
practical guide through qualitative analysis.
SAGE.
Cunningham, W. (1992). The WyCash portfolio
management system. In Addendum to the
Proceedings on Object-Oriented Programming
Systems, Languages, and Applications, 29-30.
Davenport, T. (2016). Rise of the strategy machines.
MIT Sloan Management Review, 58(1), 29-30
Desai, D. R., & Kroll, J. A. (2017). Trust but verify: A
guide to algorithms and the law. Harvard
Journal of Law & Technology, 31(1), 1-63.
Doshi-Velez, F., & Kim, B. (2017). Towards a rigorous
science of interpretable machine learning.
https://arxiv.org/pdf/1702.08608v2.pdf
Došilović, F. K., Brčić, M., & Hlupić, N. (2018).
Explainable artificial intelligence:
A survey. Proceedings of the 41st International
Convention on Information and
Communication Technology, Electronics and
Microelectronics (MIPRO).
Dourish, P. (2016). Algorithms and their others:
Algorithmic culture in context. Big Data &
Society, 3(2), 1-11.
Edwards, L., & Veale, M. (2017). Slave to the
algorithm: Why a right to an explanation is
probably not the remedy you are looking for.
Duke Law and Technology Review, 16(1), 18-
84.
Edwards, P. N. (2018). We have been assimilated:
Some principles for thinking about algorithmic
systems. Proceedings of the IFIP WG 8.2
Journal of the Association for Information Systems
345
Working Conference on the Interaction of
Information Systems and the Organization.
European Union. (2016). Regulation (EU) 2016/679 of
the European Parliament and the Council.
Official Journal of the European Union, L
119(1), 1-88.
Faraj, S., Pachidi, S., & Sayegh, K. (2018). Working
and organizing in the age of the learning
algorithm. Information and Organization,
28(1), 62-70.
Firth, N. (2019). Apple card is being investigated over
claims it gives women lower credit limits. MIT
Technology Review. https://www.technology
review.com/2019/11/11/131983/apple-card-is-
being-investigated-over-claims-it-gives-
women-lower-credit-limits/
Floridi, L. (2011). Children of the fourth revolution.
Philosophy and Technology, 24(3), 227-232.
Ghasemaghaei, M., Ebrahimi, S., & Hassanein, K.
(2018). Data analytics competency for
improving firm decision making performance.
The Journal of Strategic Information Systems,
27(1), 101-113.
Gioia, D. A., Corley, K. G., & Hamilton, A. L. (2012).
Seeking Qualitative Rigor in inductive
research: Notes on the Gioia methodology.
Organizational Research Methods, 16(1), 15-
31.
Gregor, S., & Benbasat, I. (1999). Explanations from
intelligent systems: Theoretical foundations
and implications for practice. MIS Quarterly,
23(4), 497-530.
Jarrahi, M. H. (2018). Artificial intelligence and the
future of work: Human-AI symbiosis in
organizational decision making. Business
Horizons, 61(4), 577-586.
Kaplan, A., & Haenlein, M. (2019). Siri, Siri, in my
hand: Who’s the fairest in the land? On the
interpretations, illustrations, and implications
of artificial intelligence. Business Horizons,
62(1), 15-25.
Keding, C. (2021). Understanding the interplay of
artificial intelligence and strategic
management: Four decades of research in
review. Management Review Quarterly, 71(1),
91-134.
Klein, H., & Myers, M. M. D. (1999). A set of
principles for conducting and evaluating
interpretive field studies in information
systems. MIS Quarterly, 23(1), 67-93.
Koutsikouri, D., Lindgren, R., Henfridsson, O., &
Rudmark, D. (2018). Extending digital
infrastructures: A typology of growth tactics.
Journal of the Association for Information
Systems, 19(10), 1001-1019.
Kritzinger, W., Karner, M., Traar, G., Henjes, J., &
Sihn, W. (2018). Digital twin in manufacturing:
A categorical literature review and
classification. IFAC-PapersOnLine, 51(11),
Elsevier, 1016-1022.
Kruchten, P., Nord, R. L., & Ozkaya, I. (2012).
Technical debt: From metaphor to theory and
practice. IEEE Software, 29(6), 18-21.
Lindebaum, D., Vesa, M., & Den Hond, F. (2020).
Insights from the Machine Stops to better
understand rational assumptions in algorithmic
decision making and its implications for
organizations. Academy of Management
Review, 45(1), 247-263.
Linden, A., Reynolds, M., & Alaybeyi, S. (2019). 5
Myths about explainable AI. Gartner Research.
Lipton, Z. C. (2018). The mythos of model
interpretability. ACM Queue, 16(3), 1-27.
Liu, N., Du, M., & Hu, X. (2020). Adversarial machine
learning: An interpretation perspective.
https://arxiv.org/pdf/2004.11488.pdf.
London, A. J. (2019). Artificial intelligence and black-
box medical decisions: Accuracy versus
explainability. Hastings Center Report, 49(1),
15-21.
Lyytinen, K., Nickerson, J. V, & King, J. L. (in press).
Metahuman systems = humans + machines that
learn. Journal of Information Technology.
https://doi.org/10.1177/0268396220915917.
Martens, D., Vanthienen, J., Verbeke, W., & Baesens,
B. (2011). Performance of classification models
from a user perspective. Decision Support
Systems, 51(4), 782-793.
Martin, K. (2019). Designing ethical algorithms. MIS
Quarterly Executive, 18(2), 129-142.
McBride, N., & Hoffman, R. R. (2016). Bridging the
ethical gap: From human principles to robot
instructions. IEEE Intelligent Systems, 31(5),
76-82.
McKinney, S. M., Sieniek, M., Godbole, V., Godwin,
J., Antropova, N., Ashrafian, H., Back, T.,
Chesus, M., Corrado, G. C., Darzi, A., & others.
(2020). International evaluation of an AI
system for breast cancer screening. Nature,
577(7788), 89-94.
Miles, M. B., Huberman, M. A., & Saldana, J. (2014).
Drawing and verifying conclusions. In
Qualitative data analysis: A methods
sourcebook (pp. 275-322). SAGE.
Sociotechnical Envelopment of Artificial Intelligence
346
Miller, T. (2019). Explanation in artificial intelligence:
Insights from the social sciences. Artificial
Intelligence, 267, 1-38.
Mittelstadt, B., Russell, C., & Wachter, S. (2019).
Explaining explanations in AI. Proceedings of
the Conference on Fairness, Accountability,
and Transparency.
Mumford, E. (2006). The story of socio-technical
design: Reflections on its successes, failures
and potential. Information Systems Journal,
16(4), 317-342.
Myers, M. D., & Newman, M. (2007). The qualitative
interview in IS Research: Examining the craft.
Information and Organization, 17(1), 2-26.
Newell, S., & Marabelli, M. (2015). Strategic
opportunities (and challenges) of algorithmic
decision-making: A call for action on the long-
term societal effects of datification. Journal
of Strategic Information Systems, 24(1), 3-14.
O’Neil, C. (2016). Weapons of math destruction: How
big data increases inequality and threatens
democracy. Broadway Books.
Pääkkönen, J., Nelimarkka, M., Haapoja, J., &
Lampinen, A. (2020). Bureaucracy as a lens for
analyzing and designing algorithmic systems.
Proceedings of the CHI Conference on Human
Factors in Computing Systems.
Pasquale, F. (2015). The black box society. Harvard
University Press.
Patton, M. Q. (2001). Qualitative Evaluation and
Research Methods (3rd ed.). SAGE.
Preece, A. (2018). Asking “why in AI: Explainability
of intelligent systemsperspectives and
challenges. Intelligent Systems in Accounting,
Finance and Management, 25(2), 63-72.
Raisch, S., & Krakowski, S. (in press). Artificial
intelligence and management: The automation-
augmentation paradox. Academy of
Management Review. https://journals.aom.org/
doi/10.5465/2018.0072
Ribera, M., & Lapedriza, A. (2019). Can we do better
explanations? A proposal of user-centered
explainable AI. In . In Joint Proceedings of the
ACM IUI 2019 Workshops.
Robbins, S. (2020). AI and the path to envelopment:
Knowledge as a first step towards the
responsible regulation and use of AI-powered
machines. AI & Society, 25, 391-400.
Robbins, S. (2019). A misdirected principle with a
catch: Explicability for AI. Minds and
Machines, 29(4), 495-514.
Rolland, K. H., Mathiassen, L., & Rai, A. (2018).
Managing digital platforms in user
organizations: The interactions between digital
options and digital debt. Information Systems
Research, 29(2), 419-443.
Rosenfeld, A., & Richardson, A. (2019).
Explainability in human-agent systems.
Autonomous Agents and Multi-Agent Systems,
33, 673-705.
Rudin, C. (2019). Stop explaining black box machine
learning models for high stakes decisions and
use interpretable models instead. Nature
Machine Intelligence, 1(5), 206-215.
El Saddik, A. (2018). Digital twins: The convergence
of multimedia technologies. IEEE MultiMedia,
25(2), 87-92.
Salovaara, A., Lyytinen, K., & Penttinen, E. (2019).
High reliability in digital organizing:
Mindlessness, the frame problem, and digital
operations. MIS Quarterly, 43(2), 555-578.
Sarker, S., Chatterjee, S., Xiao, X., & Elbanna, A.
(2019). The sociotechnical axis of cohesion for
the IS discipline: Its historical legacy and its
continued relevance. MIS Quarterly, 43(3),
695-719.
Sarker, S., Xiao, X., Beaulieu, T., & Lee, A. S. (2018).
Learning from first-generation qualitative
approaches in the IS discipline: An
evolutionary view and some implications for
authors and evaluators (Part 1/2). Journal of the
Association for Information Systems, 19(8),
752-774.
Sarker, Saonee, & Sarker, Suprateek. (2009).
Exploring agility in distributed information
systems development teams: An interpretive
study in an offshoring context. Information
Systems Research, 20(3), 440-461.
Scheel, P. D. (1993). Robotics in industry: A safety
and health perspective. Professional Safety,
38(3), 28-32.
Schneider, S., & Leyer, M. (2019). Me or information
technology? Adoption of artificial intelligence
in the delegation of personal strategic decisions.
Managerial and Decision Economics, 40(3),
223-231.
Sørmo, F., Cassens, J., & Aamodt, A. (2005).
Explanation in case-based reasoning-
perspectives and goals. Artificial Intelligence
Review, 24, 109-143.
Sousa, W. G. de, Melo, E. R. P. de, Bermejo, P. H. D.
S., Farias, R. A. S., & Gomes, A. O. (2019).
How and where is artificial intelligence in the
public sector going? a literature review and
Journal of the Association for Information Systems
347
research agenda. Government Information
Quarterly, 36(4), 101392.
Stone, P., Brooks, R., Brynjolfsson, E., Calo, R.,
Etzioni, O., Hager, G., Julia, H.,
Kalayanakrishnan, S., Kamar, E., Kraus, S.,
Leyton-Brown, K., Parkes, D., Press, W.,
Saxenian, A., Shah, J., Tambe, M., & Teller, A.
(2016). Artificial intelligence and life in 2030:
One Hundred Year Study on Artificial
Intelligence: Report of the 2015-2016 Study
Panel. Stanford University. https://ai100.
stanford.edu/sites/g/files/sbiybj9861/f/ai_100_
report_0831fnl.pdf
Tavory, I., & Timmermans, S. (2014). Abductive
analysis: Theorizing qualitative research.
University of Chicago Press.
Weller, A. (2019). Transparency: Motivations and
Challenges. Proceedings of the ICML
Workshop on Human Interpretability in
Machine Learning.
Wright, S. A., & Schultz, A. E. (2018). The rising tide
of artificial intelligence and business
automation: developing an ethical framework.
Business Horizons, 61(6), 823-832
Sociotechnical Envelopment of Artificial Intelligence
348
Appendix A: The DBA’s ML Projects
Project name
Project description (use case within
the DBA and end users)
Purpose
Input
Output
Model and
tool
Auditor’s
Statement
The Auditor’s Statement model speeds
up verification that the valuations of
company assets given in an auditor's
statement are correct and that the
statement does not feature violations.
The algorithm is used by internal DBA
case workers.
Prevent
misreporting
of company
assets
Text from
auditors
statements
that present
asset
valuations
Probability of
violations in asset
valuations
Random
forest, bag of
words
Bankruptcy
The Bankruptcy model predicts company
distress and insolvency and ties in with
the Early Warning Europe (EWE)
initiative. The algorithm is used not at
the DBA but by external consultants in
the EWE community in Denmark and
elsewhere in the European Union. The
DBA is not responsible for actions and
consequences related to the tool.
Identify
companies in
distress, to
enable timely
intervention
Data from the
business
registry and
annual
statements
Probability of
bankruptcy
Scikit-learn,
gradient
boosting
Company
Registration
The Company Registration model is
aimed at detecting fraud-indicating
behavior among newly registered Danish
companies. The algorithm is used by
internal DBA case workers.
Prevent
abusing
incorporation
to commit
fraud
Data from the
business
registry,
annual
statements,
and VAT
reports
Probability of
fraudulent
actions
XGBoost
Land and
Buildings
The Land and Buildings model predicts
violations of accounting policies related to
property holdings and long-term
investments. The algorithm is used by
internal DBA domain experts.
Prevent
violations of
accounting
policy
Text about
accounting
policies, from
the auditor's
statement
Probability of
violations of
accounting
policies
Random
forest, bag of
words
ID Verification
The ID Verification model expedites
processing of the documents submitted, by
supplying a text string from the machine-
readable portion of an ID document and
comparing it against input data from the
user. The algorithm is used by internal
DBA case workers.
Facilitate
processing of
documents
Pictures of
IDs submitted
to the DBA
JSON string with
text from the
machine-readable
portion of the ID
PassportEye
Recommendati
on
The Recommendation model improves
the user experience of the DBA’s virk.dk
online portal by focusing on personalized
content and optimized interfaces. The
algorithm improves the portals usability
for external customers (end users).
Improve
usability of
the online
portal
Telemetry
data from
virk.dk
Recommendation
of relevant
content
[Not decided
by the time of
this study]
Sector Code
The Sector Code model speeds up
verifying a company’s industry-sector
code. At present, 25% of the company
codes are incorrect. The algorithm is
used by internal DBA case workers.
Prevent
misreporting
of industry-
sector codes
Activity-
description text
from a
company’s
annual
statements
Probability
distribution over
the set of sector
codes
Neural
network
Signature
The Signature model, in combination with
the associated document filter, speeds up
verification of whether a company-
establishment document is signed or not.
The algorithm, used by internal DBA
case workers, returns three probabilities:
of whether the document is physically
signed, whether it is digitally signed, and
whether the signature is missing.
Facilitate the
process of
establishing a
company
An image of a
company-
establishment
document
Probability of
whether a
document is
signed or not
Neural
network
(ResNet16)
Journal of the Association for Information Systems
349
Appendix B: The Interview Protocol
Personal background
Could you tell us about your academic and professional background?
How long have you been part of the DBA, and how long have you held your current position?
Could you tell us about projects you are involved in at the DBA?
ML and AI projects at the DBA
Could you list machine-learning and AI projects currently being carried out by the ML Lab?
Could you describe ML/AI projects that you are involved with?
What types of algorithms and models are used in these projects?
What is the rationale behind using these models?
In your own words, could you please explain
Which data go into the system and what type of output the algorithm provides?
How well you understand how the algorithm works?
How you interpret the output?
Use of black-box models and explainability
How explainable are the decisions of the AI used in the projects you are involved in?
Who is able to understand how the AI produces its outputs (data scientists, developers, case workers, …)?
Have you encountered a case in which you needed to explain a particular AI decision? Could you describe the case in
detail?
Has this explanation been documented? Could you provide documents?
Could you give a concrete example of a typical decision your AI makes?
How would you explain the resulting decision if requested to do so…
By qualified auditors?
By an affected organization?
By the general public?
What would be the procedure for requesting the explanation, and for delivering it?
Is explanation embedded in the algorithm (or predefined protocol)’s design, or is it ad hoc / emergent?
Explainability requirements
How does the requirement for explainability manifest itself in algorithm development?
Do you use different machine-learning platforms for projects that require explainable AI?
Have you had any issues or problems with explainability (in development, in relations with external stakeholders,
DBA-internally, or with regard to managers)?
Have explanations been requested? By whom?
Have you been able to provide satisfactory explanations upon request?
Have you experienced inability to provide explanations to a stakeholder or to obtain explanations from one?
How should explainability be taken into account in system development?
What design principles were applied in development of PROJECTX (cost, time, etc.)?
How was the design of PROJECTX organized (following a waterfall model, in sprints, etc.)?
Was explainability a system requirement in the AI design?
What did this mean for the design process?
Sociotechnical Envelopment of Artificial Intelligence
350
If explainability was initially specified as a system requirement, did it materialize in the final design as was
intended? That is, did the final design’s explainability correspond to what was envisioned?
Describe the process of crafting an explanation:
Who creates it?
How often, and for whom?
What are the steps?
Were any of the design principles in conflict with explainability during the design phase?
If so, how did you navigate through the issue?
Have you noticed conflicts related to differing understandings of the work done by the algorithm?
Could you give examples?
Is such conflict acceptable, or do contradictions need to be reconciled?
How are they reconciled?
What do you consider the best way to resolve conflicts?
Reasons for developing explainable AI and its implications
What are the main reasons for the requirement to explain AI?
Why do you need explainability?
For internal purposes: for finding out how to improve your AI, or to double-check its outputs?
For external purposes: to be accountable as a governmental authority with defendable unbiased processes?
External pressure for explainability:
Do you have to be able to explain AI decisions to clients (taxpayers)? How, and at what level of detail?
Which regulations, internal policies, outside pressure, etc. force you to explain the AI’s decisions?
Who are the main actors for whom you craft explanations? Could you name them and provide examples of
what those explanations are like?
How do explainability requirements constrain the process of AI development? Could you describe these constraints?
Do you have to limit your use of AI approaches because of a need for explainability?
How does needing to produce explainable systems affect the systems’ performance?
Overall, how does explainability influence your ability to achieve organizational objectives?
Journal of the Association for Information Systems
351
Appendix C: The Coding
Concepts
(first-order)
Themes
(second-order)
Aggregate
dimensions
Example quotations
Case workers’ control of thresholds
Guidance on threshold-setting
The thresholds’ dependence on the
code
Thresholds
Boundary
envelopes
“But we’re involved more or less the whole way
because if suddenly there is a problem or
suddenly there is ‘Okay, we can deploy this, but
do you want the machine to do this or this? Do
you want it to have a marker saying this case
cannot go further, or do you just want it to go
through and [we] have a special marker where
we can look it up later?’... So we are involved
the whole way, but at some points we are more
[in the goals or in practice] helping or [asking]
‘Can we do...?’”
Conversion of probabilities into flags
The AI flagging only basic flaws in
documents
Flags
Designing AI that is easier to hand
over
Basic AI tools with wide
applicability
Division of a
task into smaller
parts
Simple algorithms’ ease of
explanation
An explainability/performance
tradeoff not always existingsimple
models work just fine
Choosing of
interpretable
algorithms
Close communication links for
reducing misunderstandings during
development
Communication with developers
Social dialogue
Understanding of input data as
important
Quality of inputs
Input control
Input and
output
envelopes
“An example could be that our model [for
whether a document is] signed or not, as it is
now, if the model forecasts that the document is
signed, then it gets a special code, ‘document
signed, everything is okay,’ and if it’s not
signed, then it gets another marking, for
‘document not signed.’ These cases we go
through, and then you can see that was correct
and that was not correct. In that case, there isn’t
really any- we don’t need to know- I don’t need
to know as [a case worker] why the model said
‘signed’ or ‘not signed,’ because I can see
instantly if it’s right or not right.”
Compensation for explainability-
induced lower performance, via
control over the output’s use
Acceptability of having a black box
if checking the outputs is simple
Output control
Verification as an aid to establishing
trust in ML
a human holding ultimate
responsibility
Simple algorithms that a human
expert can follow and reproduce
Human
verification
External stakeholdersinvolvement
in early stages of development
Establishment of feedback channels
between technical and business
teams
Human
feedback
Model-choice
envelopes
“We have around 160 rules. We have technical
rules that look into whether the right taxonomy
is being used, whether it is the XBRL format,
and whether it is compliant. We also have
business rules. For example, do assets and
liabilities match? Some rules only look at
technical issues in the instance report. Some
rules are what we called full-stop rules: … filers
are not allowed to file the report until they have
corrected the error. We also have more
guidance[-type] rules, where we say, ‘It looks
like you’re about to make a mistake. Most
people do it this way. Are you sure you want to
continue filing the report?’ And then [users] can
choose to ignore the rule.”
Governance of AI development
In-house development, to improve
understanding
Continuous-
improvement
procedure
Internal accumulation of training
data
Data “red herrings”
Training on in-house data
Knowledge of
data
Training-data
envelopes
“I think it’s important with these models to look
at them often to see if something is changing.
And, maybe, train them again. Because I think
there might be some issues, with the robustness.
We haven’t gotten this system into production
yet, but I think it’s on its way.”
Challenges of creating models
The dangers of training a model on
the open internet
Training of models in stages
Phased training
of a model
Sociotechnical Envelopment of Artificial Intelligence
352
About the Authors
Aleksandre Asatiani is an assistant professor in information systems at the Department of Applied Information
Technology, at the University of Gothenburg. He is also an affiliated researcher with the Swedish Center for Digital
Innovation (SCDI). His research focuses on artificial intelligence, robotic process automation, virtual organizations,
and IS sourcing. His work has previously appeared in leading IS journals such as Information Systems Journal, Journal
of Information Technology, and MIS Quarterly Executive.
Pekka Malo is a tenured associate professor of statistics at Aalto University School of Business. His research has been
published in leading journals in operations research, information science, and artificial intelligence. Pekka is
considered as one of the pioneers in the development of evolutionary optimization algorithms for solving challenging
bilevel programming problems. His research interests include business analytics, computational statistics, machine
learning, optimization and evolutionary computation, and their applications to marketing, finance, and healthcare.
Per Rådberg Nagbøl is a PhD fellow at the IT University of Copenhagen doing a collaborative PhD with the Danish
Business Authority within the field of information systems. He uses action design research to design systems
and procedures for quality assurance and evaluation of machine learning, focusing on accurate, transparent, and
responsible use in the public sector from a risk management perspective.
Esko Penttinen is a professor of practice in information systems at Aalto University School of Business in Helsinki.
He holds a PhD in information systems science and an MSc in Economics from Helsinki School of Economics. Esko
leads the Real-Time Economy Competence Center and is the co-founder and chairman of XBRL Finland. He studies
the interplay between humans and machines, organizational implementation of artificial intelligence, and governance
issues related to outsourcing and virtual organizing. His main practical expertise lies in the assimilation and economic
implications of interorganizational information systems, focusing on application areas such as electronic financial
systems, government reporting, and electronic invoicing. Esko’s research has appeared in leading IS outlets such as
MIS Quarterly, Information Systems Journal, Journal of Information Technology, International Journal of Electronic
Commerce, and Electronic Markets.
Tapani Rinta-Kahila is a postdoctoral research fellow at the UQ Business School and Australian Institute for Business
and Economics, at the University of Queensland in Australia. He holds a doctoral degree in information systems
science from the Aalto University School of Business. His research addresses issues related to IT discontinuance,
organizational implementation of artificial intelligence and automation, and the dark side of IS.
Antti Salovaara is a senior university lecturer at Aalto University, Department of Design and an adjunct professor in
the Department of Computer Science at the University of Helsinki. He studies human-AI collaboration and online
trolling and the methodology of user studies. His research has been published both in human-computer interaction and
information systems journals and conferences, including CHI, Human Computer Interaction and International Journal
of Human-Computer Studies, as well as MIS Quarterly and European Journal of Information Systems.
Copyright © 2021 by the Association for Information Systems. Permission to make digital or hard copies of all or part
of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for
profit or commercial advantage and that copies bear this notice and full citation on the first page. Copyright for
components of this work owned by others than the Association for Information Systems must be honored. Abstracting
with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists requires prior
specific permission and/or fee. Request permission to publish from: AIS Administrative Office, P.O. Box 2712 Atlanta,
GA, 30301-2712 Attn: Reprints, or via email from publications@aisnet.org.
... Learning refers to the AI system's ability to improve through data and experience (Ågerfalk, 2020;Janiesch et al., 2021). Inscrutability refers to the unintelligibility of AI systems to some audiences, given their complex inner workings and probabilistic outputs (Asatiani et al., 2021;Jöhnk et al., 2021). These characteristics are expected to even exacerbate as the field of AI moves forward and new techniques and approaches emerge. ...
... We propose that the identified capabilities help organizations cope with two characteristics in AI: inscrutability and data dependency. Inscrutability refers to the unintelligibility of AI systems to some audiences due to their complex inner workings and probabilistic nature (Asatiani et al., 2021;Berente et al., 2021;Jöhnk et al., 2021). Data dependency refers to the high dependence of AI systems on the underlying data, as these systems are typically built by learning and generalizing from data (Ågerfalk, 2020;Berente et al., 2021;Janiesch et al., 2021). ...
... On the other hand, business functions require transparency and explanations from the IT department regarding the inner functioning, performance, and limitations of the developed AI system (Watson, 2017). Business functions need to understand how an AI system works to deal with its potential limitations (Asatiani et al., 2021). Explaining the workings of AI systems can also help address potential user resistance caused by misconceptions about the developed AI system (Reis et al., 2020). ...
Article
Full-text available
Artificial Intelligence (AI) implementation incorporates challenges that are unique to the context of AI, such as dealing with probabilistic outputs. To address these challenges, recent research suggests that organizations should develop specific capabilities for AI implementation. Currently, we lack a thorough understanding of how certain capabilities facilitate AI implementation. It remains unclear how they help organizations to cope with AI’s unique characteristics. To address this research gap, we employ a qualitative research approach and conduct 25 explorative interviews with experts on AI implementation. We derive four organizational capabilities for AI implementation: AI Project Planning and Co-Development help to cope with the inscrutability in AI, which complicates the planning of AI projects and communication between different stakeholders. Data Management and AI Model Lifecycle Management help to cope with the data dependency in AI, which challenges organizations to provide the proper data foundation and continuously adjust AI systems as the data evolves. We contribute to our understanding of the sociotechnical implications of AI’s characteristics and further develop the concept of organizational capabilities as an important success factor for AI implementation. For practice, we provide actionable recommendations to develop organizational capabilities for AI implementation.
... End-users often cannot comprehend how AI systems reach their decisions (Waardenburg et al., 2020). However, explainability is crucial for using AI in joint decision-making (Asatiani et al., 2021). ...
... Comprehensibility of the algorithmic mechanisms is the key barrier that prevents farmers from working with AI systems (Asatiani et al., 2021). Statistical capabilities are lacking on the farmers and manufacturing sides (Klerkx et al., 2019). ...
... Sociotechnical environments refer to system environments defined by regular interactions between people and technology such as human-machine interactions and tactile internet [14]- [16]. They are enabled by sociotechnical systems in which designers attempt to jointly optimize both the social and technical elements so that social criteria such as human well-being and productivity have the same weight as technical criteria such as device lifetimes or efficiencies. ...
... Decentralization is an important approach for designing complex social environments with distributed actors, giving several benefits such as privacy preservation, self-adaption, independence, and social welfare [17]. The concept of sociotechnical systems and environments has evolved from simple one-to-one interactions between humans and machines to today's idea that sociotechnical environments comprise a collection of massive number of IoT devices, where IoT refers to a network of ''intelligent'' physical objects-i.e., objects embedded with sensors, software, and technologies-that connect and exchange data with other devices and systems over the Internet [14]- [16]. ...
Article
Full-text available
Distributed intelligence is a well-known approach for optimizing interactions among numerous smart devices that interconnect and operate together as Internet of Things (IoT) systems. A modern form of human-machine collective intelligence emerges when humans interact with IoT systems in sociotechnical environments such as smart homes. Fifth-generation (5G) communication networks are designed for high-speed reliable wireless connectivity and expected to boost IoT and (distributed) collective intelligence by revolutionizing human–device–human interactions. In this paper, we contribute a comprehensive review of state-of-the-art sociotechnical environments that exhibit collective intelligence, supported by 5G-enabled IoT. We discuss the latest developments in 5G and their implications for collective intelligence. Further, we explain the key challenges for using 5G to support collective intelligence, e.g., data processing, security, and radio resource management. Finally, we describe four practical applications of collective intelligence to sociotechnical environments—road traffic control, unmanned aerial vehicles, electrical load demand response, and augmented democracy.
... This transparency issue has also drawn attention from the professions, where for instance audit researchers working on AI for audit data analytics, recognize the concerns of not being able to explain their decisions (Zhang et al., 2021). Organizational forces may envelop the AI techniques to control the unknown (Asatiani et al., 2021), but the unknown invariably limits experts' reliance. ...
... The hesitancy from the professions comes largely from not knowing what those technologies are doing (Sutton et al., 2018;Asatiani et al., 2021;Zhang et al., 2021). ...
Preprint
Full-text available
The Theory of Technology Dominance (TTD) provides a theoretical foundation for understanding how intelligent systems impact human decision-making. The theory has three phases with propositions related to (1) the foundations of reliance, (2) short-term effects on novice versus expert decision-making, and (3) long-term epistemological effects related to individual deskilling and profession-wide stagnation. In this theory paper, we propose an extension of TTD, that we refer to as TTD2, primarily to increase our theoretical understanding of how, why, and when the short-term and long-term effects on decision-making occur and why advances in technology design have exacerbated some weaknesses and eroded some benefits. Recently, researchers have called for reconsideration of how we design intelligent systems to mitigate the detrimental effects of technology; in TTD2 we provide a theory-based understanding for reimagining how such systems are designed.
... We need more studies to uncover what sorts of organization change approaches work, for whom and under what circumstances. As an example, Asatiani et al. (2021) proposed ways to implement AI systems when these systems are non-explainable and inscrutable; a process the authors referred to as 'sociotechnical envelopment'. ...
... Moreover, they require a strategic frame that combines a technical, managerial, and temporal perspective (Raisch and Krakowski, 2021). Incumbent banks thus need means to structurally evaluate investments in AI-related capabilities to seize opportunities from their existing resources (Oberländer et al., 2021) while addressing potentially unintended adverse outcomes (Benbya et al., 2020) that may lie in AI's inscrutability (Asatiani et al., 2021;Teodorescu et al., 2021) or flawed ground truth assumptions in training and evaluating AI models (Lebovitz et al., 2021). Against this backdrop, we ask the following research question: How can banks successfully manage their investments in AI-related IT capabilities? ...
Conference Paper
Full-text available
Technology-driven challenges, both existing and emerging, require banks to invest in IT capabilities, especially in artificial intelligence (AI). Digital options theory presents a valuable guide rail for these investments. However, the nature of AI as a moving frontier of computing requires certain extensions to established digital option thinking. Based on interviews with 23 experts in the retail banking industry, we highlight the importance of thinking broadly when laying the foundation for AI options and being mindful of the dynamic effects of contextual factors. Drawing from digital options theory and the Technology-Organization-Environment framework as dual lens, our study adds a structured approach to consciously balance resources and AI-related capability investments with a broader consideration of the banking industry’s complex environment. In this way, our study complements recent research on the interplay between incumbents’ resources and digital opportunities.
... An empirical quantification of end user explainability is necessary to provide first-hand knowledge to the engineers of intelligent systems for the development of intelligent systems (Jauernig et al., 2022;. Despite this apparent deficiency, they are commonly referenced as a motivation for a user-or organization-centered XAI research or intelligent system deployment (e. g., Asatiani et al., 2021;Guo et al., 2019;Rudin, 2019). ...
Preprint
Full-text available
Machine learning algorithms enable advanced decision making in contemporary intelligent systems. Research indicates that there is a tradeoff between their model performance and explainability. Machine learning models with higher performance are often based on more complex algorithms and therefore lack explainability and vice versa. However, there is little to no empirical evidence of this tradeoff from an end user perspective. We aim to provide empirical evidence by conducting two user experiments. Using two distinct datasets, we first measure the tradeoff for five common classes of machine learning algorithms. Second, we address the problem of end user perceptions of explainable artificial intelligence augmentations aimed at increasing the understanding of the decision logic of high-performing complex models. Our results diverge from the widespread assumption of a tradeoff curve and indicate that the tradeoff between model performance and explainability is much less gradual in the end user's perception. This is a stark contrast to assumed inherent model interpretability. Further, we found the tradeoff to be situational for example due to data complexity. Results of our second experiment show that while explainable artificial intelligence augmentations can be used to increase explainability, the type of explanation plays an essential role in end user perception.
... An empirical quantification of end user explainability is necessary to provide first-hand knowledge to the engineers of intelligent systems for the development of intelligent systems (Jauernig et al., 2022;. Despite this apparent deficiency, they are commonly referenced as a motivation for a user-or organization-centered XAI research or intelligent system deployment (e. g., Asatiani et al., 2021;Guo et al., 2019;Rudin, 2019). ...
Article
Full-text available
Machine learning algorithms enable advanced decision making in contemporary intelligent systems. Research indicates that there is a tradeoff between their model performance and explainability. Machine learning models with higher performance are often based on more complex algorithms and therefore lack explainability and vice versa. However, there is little to no empirical evidence of this tradeoff from an end user perspective. We aim to provide empirical evidence by conducting two user experiments. Using two distinct datasets, we first measure the tradeoff for five common classes of machine learning algorithms. Second, we address the problem of end user perceptions of explainable artificial intelligence augmentations aimed at increasing the understanding of the decision logic of high-performing complex models. Our results diverge from the widespread assumption of a tradeoff curve and indicate that the tradeoff between model performance and explainability is much less gradual in the end user’s perception. This is a stark contrast to assumed inherent model interpretability. Further, we found the tradeoff to be situational for example due to data complexity. Results of our second experiment show that while explainable artificial intelligence augmentations can be used to increase explainability, the type of explanation plays an essential role in end user perception.
Chapter
The definition of the smart city has changed over time and is expected to continue to change. For example, Japan stepped into the Society 5.0 vision, which encompasses robotics, artificial intelligence, the Internet of Things, and big data. Society 5.0 is Japan’s national vision within the fifth Science and Technology Fundamental Plan, which aims to realize a data-driven, human-centered, next-generation society (Kobashi et al., 2020; Zhang et al., 2020). This vision means that everyone, regardless of location, will benefit from innovation and advances in technology, including the elderly population in rural areas. Community 5.0 emphasizes solving all kinds of social problems while harmonizing sustainability and economic growth. Therefore, Community 5.0 guides cities of all sizes on how to build smart infrastructure. It is no longer just about new technologies, but also about understanding a community’s problems and providing appropriate solutions through these technologies. In the context of the smart city, this approach is called “problem-solution-oriented smart city” (Neirotti et al., 2014), and it is the basis for achieving the greater goals of Community 5.0.
Preprint
Full-text available
Our recent research shows that the design philosophy of human factors research in the intelligence age is expanding from "user-centered design" to "user-centered design 2.0" and "human-centered AI", and the human-machine relationship presents a trans-era evolution from "human-machine interaction" to "human-machine teaming". These changes have raised new questions and challenges for human factors research, compelling us to re-examine the paradigm and agenda of human factors research that was traditionally based on non-intelligent technologies. In this context, this paper reviews the cross-generational expansion of the human factors research paradigm and summarizes the new conceptual models and frameworks we proposed to enrich the human factors research paradigm, including a human-agent teaming model, a human-agent joint cognitive ecosystem framework, and an intelligent sociotechnical systems framework. This paper further enhances these concepts and looks forward to the corresponding application of these concepts and future research agenda. This paper also looks forward to the future agenda of human factors research from three aspects: "human-AI interaction", "intelligent human-machine interface", and "human-machine teaming". It analyzes the role of human factors research paradigms on future research agenda. We believe that the research paradigms and the research agenda influence and promote each other. Human factors research in the intelligence age needs diversified and innovative research paradigms, thereby further promoting the development of human factors science.
Conference Paper
Full-text available
Artificial Intelligence systems are spreading to multiple applications and they are used by a more diverse audience. With this change of the use scenario, AI users will increasingly require explanations. The first part of this paper makes a review of the state of the art of Explainable AI and highlights how the current research is not paying enough attention to whom the explanations are targeted. In the second part of the paper, it is suggested a new explainability pipeline, where users are classified in three main groups (developers or AI researchers, domain experts and lay users). Inspired by the cooperative principles of conversations, it is discussed how creating different explanations for each of the targeted groups can overcome some of the difficulties related to creating good explanations and evaluating them.
Article
Full-text available
Screening mammography aims to identify breast cancer at earlier stages of the disease, when treatment can be more successful1. Despite the existence of screening programmes worldwide, the interpretation of mammograms is affected by high rates of false positives and false negatives2. Here we present an artificial intelligence (AI) system that is capable of surpassing human experts in breast cancer prediction. To assess its performance in the clinical setting, we curated a large representative dataset from the UK and a large enriched dataset from the USA. We show an absolute reduction of 5.7% and 1.2% (USA and UK) in false positives and 9.4% and 2.7% in false negatives. We provide evidence of the ability of the system to generalize from the UK to the USA. In an independent study of six radiologists, the AI system outperformed all of the human readers: the area under the receiver operating characteristic curve (AUC-ROC) for the AI system was greater than the AUC-ROC for the average radiologist by an absolute margin of 11.5%. We ran a simulation in which the AI system participated in the double-reading process that is used in the UK, and found that the AI system maintained non-inferior performance and reduced the workload of the second reader by 88%. This robust assessment of the AI system paves the way for clinical trials to improve the accuracy and efficiency of breast cancer screening. An artificial intelligence (AI) system performs as well as or better than radiologists at detecting breast cancer from mammograms, and using a combination of AI and human inputs could help to improve screening efficiency.
Article
Full-text available
There is widespread agreement that there should be a principle requiring that artificial intelligence (AI) be ‘explicable’. Microsoft, Google, the World Economic Forum, the draft AI ethics guidelines for the EU commission, etc. all include a principle for AI that falls under the umbrella of ‘explicability’. Roughly, the principle states that “for AI to promote and not constrain human autonomy, our ‘decision about who should decide’ must be informed by knowledge of how AI would act instead of us” (Floridi et al. in Minds Mach 28(4):689–707, 2018). There is a strong intuition that if an algorithm decides, for example, whether to give someone a loan, then that algorithm should be explicable. I argue here, however, that such a principle is misdirected. The property of requiring explicability should attach to a particular action or decision rather than the entity making that decision. It is the context and the potential harm resulting from decisions that drive the moral need for explicability—not the process by which decisions are reached. Related to this is the fact that AI is used for many low-risk purposes for which it would be unnecessary to require that it be explicable. A principle requiring explicability would prevent us from reaping the benefits of AI used in these situations. Finally, the explanations given by explicable AI are only fruitful if we already know which considerations are acceptable for the decision at hand. If we already have these considerations, then there is no need to use contemporary AI algorithms because standard automation would be available. In other words, a principle of explicability for AI makes the use of AI redundant.
Conference Paper
Full-text available
Knowledge work organizations are increasingly leveraging automation to enhance and transform their business processes. Many types of automation tools are being deployed in a large variety of information processing tasks, requiring effective management of human-automation cooperation. Yet, conceptual understanding of human-automation hybrid work remains thin and current literature lacks practical recommendations for managers. To address this gap, we synthesize findings from our three earlier case studies with organizations pursuing a wide array of automation tools and examine them through the lens of distributed cognition. We demonstrate how distributed cognition informs about the organizing for human-automation interaction when deploying automation. Our contribution lies in the presentation of six recommendations on three issues: human-automation task allocation, mitigation of the risk of deskilling, and management of collective knowledge across human and automation.
Article
Huge increases in computing capacity and data volumes have spurred the development of applications that use artificial intelligence (AI), a technology that is being implemented for increasingly complex tasks, from playing Go to screening for cancer. Private and public businesses and organizations are deploying AI applications to process vast quantities of data and support decision making. These applications can help to reduce the costs of providing various services, deliver new services and improve the safety and reliability of operations. However, unlike conventional information systems, the algorithms embedded in AI applications can be "black boxes." Previously, those who developed applications could completely explain how an algorithm worked. Given an input, they could tell you what the output would be and why, because the systems applied human-made rules. That is no longer true for AI-based applications. The application creates internal structures that determine outputs, but these are inscrutable to outside observers, and even the programmers cannot tell you why a specific output was generated. Many AI systems leverage machine learning, There are many examples of problems resulting from inscrutable AI systems, so there is a growing need to be able to explain how such systems produce their outputs. Drawing on a case study at the Danish Business Authority, we provide a framework and recommendations for addressing the many challenges of explaining the behavior of black-box AI systems. Our findings will enable organizations to successfully develop and deploy AI systems without causing legal or ethical problems. 1,2
Article
Metahuman systems are new, emergent, sociotechnical systems where machines that learn join human learning and create original systemic capabilities. Metahuman systems will change many facets of the way we think about organizations and work. They will push information systems research in new directions that may involve a revision of the field’s research goals, methods and theorizing. Information systems researchers can look beyond the capabilities and constraints of human learning toward hybrid human/machine learning systems that exhibit major differences in scale, scope and speed. We review how these changes influence organization design and goals. We identify four organizational level generic functions critical to organize metahuman systems properly: delegating, monitoring, cultivating, and reflecting. We show how each function raises new research questions for the field. We conclude by noting that improved understanding of metahuman systems will primarily come from learning-by-doing as information systems scholars try out new forms of hybrid learning in multiple settings to generate novel, generalizable, impactful designs. Such trials will result in improved understanding of metahuman systems. This need for large-scale experimentation will push many scholars out from their comfort zone, because it calls for the revitalization of action research programs that informed the first wave of socio-technical research at the dawn of automating work systems.