Organizational decisions have become more data-driven and collaborative, with the increasing utilization of artificial intelligence, machine learning, and analytics to support decision making. While humans and machines are each bounded in their own rationalities, their collaboration has enabled a new, collaborative rationality by augmenting the intelligence and capabilities of each. New research is required to explore the degree and collaboration, with the aim of enhancing collaborative rationality, and its effect on decision making. Furthermore, the resulting rationalization, and sensemaking from the decision outcomes. However, data-driven decisions are complex in nature, and current theoretical developments fall short in accommodating for their multi-faceted nature and changing context, and there is lack of theoretical support on how, when, and why to evaluate such decisions. Accordingly, we follow a design science research methodology to develop and evaluate a model for data-driven decision evaluation. This model depicts the relationship and role of the multiple elements involved in data-driven decision making, and provides a feedback loop which inputs the results of evaluating decision outcomes back into the process/system, thus enabling learning from the past through experience, and ultimately enhancing collaborative rationality and decision making. Data-driven decision making, decision evaluation, collaborative rationality, human-machine BIR 2022 Workshops and Doctoral Consortium, 21st International Conference on Perspectives in Business Informatics Research (BIR 2022),
collaboration, design science research
For centuries, decisions have inevitably defined the future of organizations and societies, with
decision makers constantly striving to understand and deliberate the complicated process of how to
decide. In recent decades, the coexistence of artificial intelligence (AI), or machine learning (ML),
systems with human decision makers has ignited an interest in machines augmenting human intelligence
and capabilities, which has led to different, and unprecedented, dimensions of “intelligent” data analysis
in order to support and enhance decision-making and learning [
2022 Copyright for this paper by its authors.
There are three main modes of collaboration between humans and machines in decision-making.
The first category is full human to AI delegation, where the machine makes the algorithmic decisions,
and the role of the human is simply supervisory. The second category is of a hybrid, sequential form.
Either the machine provides suitable alternatives to the human, who then selects the most appropriate,
or the human decision maker selects a set of alternatives and then passes them to the machine for
evaluation. Here, the role of the machine is mainly to provide recommendations or insights. Finally, the
third category is aggregated human-machine decision making where different aspects, or elements, of
the decision are allocated to each of humans and machines based on their respective strengths, and then
aggregated into a collective decision [
We utilize the term data-driven decision making to define this collaboration and encompass the
relationship between the human decision maker, machine, data, decision-making process, and decision
outcome [
However, to argue that such decisions are indeed better, evaluation is necessary. Evaluation clarifies
options, reduces uncertainties, and generates information and knowledge about the results within
contextual boundaries to make more informed decisions in the future [
In recent years, research has started perceiving the relationship between organizational learning and
machine learning [
Due to the complexity of data-driven decisions and the interrelationship between the various
elements involved, the evaluation and monitoring of the resulting decisions requires additional research
of its own for building experiences [
Accordingly, the main research question we aim to study is:
RQ 1: “How can we support data-driven decision making and enhance collaborative rationality between humans and machines through decision evaluation?”
We intend not just to study collaborative rationality and data-driven decision evaluation, but also to design a solution which is theoretically sound and practically feasible. Accordingly, we adopt a design science research (DSR) methodology to develop and evaluate a model (artifact) which supports evaluating data-driven decision outcomes, enabling collaborative rationality, and creating a feedback loop to enhance human and machine learning from past decisions. Hence, we attempt to research the following sub-questions:
RQ 1.1: “Why do organizations need to evaluate collaborative rationality-based decisions?” and “How can collaborative rationality-based decisions be evaluated?”
Here we research the theoretical underpinnings behind data-driven decision making and evaluation by reviewing the literature, in order to pave the road for developing more comprehensive evaluation solutions for human-machine collaborative rationality-based decisions.
RQ 1.2: “What are the requirements and design objectives for ex-post evaluation of data-driven decisions in organizations?”
Here we define the requirements and design objectives (DOs) for a data-driven decision evaluation
solution, contributing to the first two stages of the DSR process (cf. [
RQ 1.3: “How can data-driven decisions be evaluated to enable feedback and learning loops, as well as enhance the quality of human-machine collaborative rationality-based decisions?”
Here we develop and test a model, based on theory and practical case studies, as a design science
artifact for evaluating data-driven decisions, thus completing stages 3-6 (design and development,
demonstration, evaluation, and communication) of the DSR process (cf. [
This paper is structured as follows. In Section 2, we elaborate the research methodology followed and explain what has been, or will be, done in each of the DSR stages. Finally, in Section 3 we describe the expected contribution of this research to knowledge and practice.
The research methodology applied is design science, which consists of a rigorous process to design
new artifacts intended to solve observed problems, make research contributions, evaluate the designs
of the artifacts, and communicate results to appropriate audiences [
In addition to perusing the knowledge base and the available literature, we work closely with
decision makers and practitioners in multiple organizations to gather the business needs from the
environment, as well as develop, demonstrate, and evaluate the artifact, hence achieving research
relevance and rigor [
Peffers et al.’s [
Define Objectives of a
Solution • Literature • Interviews • Design objectives
Design and Development • Design based on literature review and interviews • Static and behavior models
Process Iteration
Demonstration • Instantiation in companies and application on use cases
Evaluation • Ex-ante evaluation • Ex-post evaluation
Communication • Dissemination and publication of findings
We start with a problem-centered initiation by identifying the problem and motivation. This is
mainly done by analyzing the literature and related research. The defined problem and motivation for
research were evaluated and supported in practice through the expert interviews conducted and
explained in the following subsection. While human-machine collaboration can potentially enhance
decision making [
Moreover, it is perceived throughout research and practice that data-driven decision making can lead
to the optimizing of decisions and result in more informed, quality decisions which could have
otherwise been unattainable. Controversially, decision-making can also become more difficult when
different combinations of data or analytics show different patterns, sometimes conflicting with the
preferred choice of the decision-maker, and hence the decision-maker does not know how to proceed
with the results [
While classical decision-making research focuses on the decision-making process, the decision
maker, and the decision itself, the emergence of big data analytics has led to an evolution of modern
data-driven decision making. Consequently, two new elements, the (big) data and the analytics need to
be incorporated and integrated with the classical decision-making elements [
However, a great deal of effort and research is required in order to meaningfully integrate these elements. Human and machine decision makers can coexist and collaborate to reach a higher level of rationality and maximize collaborative intelligence, unattainable by either one individually, which necessitates innovative models of decision making. Thus, the degree of collaboration between the human and the machine, the selection of quality (big) data, the appropriate use of analytics methods and tools, the definition of the decision-making process, and how to integrate all of these elements together, along with the resulting information and knowledge, are all imperative aspects to study for optimizing data-driven decisions and enhancing collaborative rationality between humans and machines.
Several recent studies have elaborated different modes of collaboration between humans and
machines in organizational decision making for combining human and machine intelligence and
capabilities, according to the context and types of the decision [
This leads to collective intelligence, where future research on cognitive computing has been
highlighted, with the goal of building a rational, combined, and collective mechanism motivated by the
capability of the human mind and strengths of AI systems [
Nevertheless, no comprehensive, or holistic, solutions accommodating for the multifaceted nature
of collaborative data-driven decisions were found in the literature. The interaction between humans and
machines and their roles in decision making is still not clear, the concept of collaborative rationality
between humans and machines and its effect on decision making requires thorough elaboration, and
further research is necessary to evaluate the resulting decisions and determine the benefit, impact, and
learning that consequently occur from this collaboration [
Yet, with the various number of elements involved in complex, data-driven decisions, evaluation
considering individual metrics or perspectives is insufficient. Accordingly, we can neither rely solely
on the machine’s evaluation of its models, nor the human’s evaluation of their choices or judgements.
Conversely, evaluation should consider the decision as a whole. However, many solutions focus on
evaluating machine output rather than the entire decision, or evaluate with individual metrics, despite
studies showing that even decisions with high accuracy are not necessarily correct and suffer in other
dimensions [
The objectives of a solution were defined both from theory and practice in [
The concepts were then exemplified through an industrial example of a chemical production plant to foresee how ex-post evaluation of data-driven decisions could be done in practice and outline the initial requirements for a design solution. Two expert interviews were conducted to gather data on the plant’s data-driven decisions and discuss the needs and requirements for a data-driven decision solution, from the viewpoints of both the production planner and process operator roles.
From the production planner’s perspective, the data-driven decision involves determining and planning the production targets and capacities for a specific time interval and to schedule production. The purpose is to optimize the production rate and product portfolio to meet market demand. While these operational decisions are made short-term, they have a long-term impact. The human determines the objectives and constraints, and the ML tool supports the decision by simulating scenarios and suggesting alternative schedules. However, the human selects the best schedule to meet the designated criteria and makes the final decision (mode of collaboration: AI-recommends, human decides). Commonly, the human overlooks the output of the ML tool and decides not to use it, due to lack of trust in the reliability of the results.
From the process operator’s perspective, the decisions include choices on set points for the process (such as feed rates and temperature), steering the process, and avoiding/overcoming fault situations. The purpose is to optimize the process efficiency (energy, material), avoid faults, and solve possible problems. Decisions are continuous and may also be triggered whenever there is a new target for the process, or a disturbance occurs. The ML tool provides outputs, insights, and predictions based on the data and process parameters. However, the inference and final decision are made by the human, who may use their own knowledge and expertise, along with additional monitoring methods (mode of collaboration: AI recommends, human decides and AI generates insights, which human uses in decision process).
The need for an ex-post evaluation solution was mainly to assess the reliability of the ML tool and increase trust, enhance both human and machine learning from evaluation feedback, evaluate decisions at different time intervals to indicate if the reliability of the tool increases, evaluate the monitoring methods, and evaluate uncertainties in the measurement data and their effect on the decision.
By comparing the current and desired approaches for decision evaluation stated in the interviews, the results were summarized into a set of evaluation requirements for each of the theoretical evaluation concepts previously defined from the literature. According to their functional similarities, the requirements were further thematized and mapped to more abstract and implementable DOs for an evaluation solution.
Accordingly, this study resulted in four DOs for a data-driven decision evaluation solution, shown
in Table 1. The first DO for an implementable data-driven decision evaluation method is that it should
be comprehensive and incorporate multi-faceted criteria ranging across different contextual levels. The
second DO is that a processual evaluation is performed across different stages in time to accommodate
changing contexts and consequences of the decision. The third DO suggests incorporating into the
evaluation the mode of collaboration between humans and machines and the consequent effect on
decision making, the decision outcomes, and achieving collaborative rationality. Finally, the fourth DO
would be to enable a potentially automated feedback loop which ensues from the evaluation. Decision
makers then learn a new set of lessons from experiences and from evaluating the outcomes of their
decision (link between actions and outcomes), which then leads to learning within the organization, as
well as amongst other organizations [
The next stage in the research process is to design and develop an artifact based on the previous DOs. Our artifact, in this research, is a model providing a holistic, multifaceted, and multidisciplinary view for evaluating collaborative human-machine decisions. The essential constructs and attributes which are relevant to the model, as well as the underlying assumptions, were defined from the literature and previous theories.
An additional expert interview was conducted in another organization, focusing on data-driven decisions for predicting and preventing customer churn. This interview helped externally validate the design objectives and add new perspectives from a different case and decision context. Utilizing the theoretical concepts, and the knowledge gained from the three expert interviews, an initial version of a structural model for evaluating data-driven, or human-AI centric decisions, shown in Figure 4, was developed.
For structuring the model with a conceptual modeling grammar [
The model thus accommodates for the multifaceted and changing nature of data-driven decisions across contextual levels and provides a holistic perspective by depicting the relationship between the different concepts, such as the decision maker, machine agent, data, collaborative rationality, decision, criteria to which the decision should conform, ex-ante evaluation of the decision, ex-post evaluation of the decision, and the resulting feedback loop and types of learning enabled.
We assume recurring data-driven decisions, taking place within an organizational context, further embedded within an external, environmental context. A human decision maker and artificial agent (machine, analytics, etc.) are informed by a set of available data, of a certain quality. Both are characterized by a level of rationality, have a particular role in the decision-making process, have perceived models of the environment under which they operate, and a set of reference points which serve as baseline criteria, or governing variables, to which their decisions must conform. They are also prone to errors and biases. The human is further driven by emotions, which is a crucial attribute which the machine lacks.
Together, the decision maker and machine, within an organizational context, engage in collaborative rationality, which can be utilized within the context of the decision, and in other contexts as well. The collaborative rationality is defined by the mode of collaboration, its characteristics, and its supported models of rationality. This collaborative rationality facilitates, and is facilitated by, the ex-ante evaluation, as new insights are generated through the evaluation of choices and alternatives, and it further supports ex-post evaluation, from which it is supported through the feedback loop. The ex-ante evaluation, utilizing its own set of metrics to evaluate alternatives, and the perceived outcomes leads to a decision. This is defined by the time of the decision, its characteristics, and the actions involved for implementation.
Subsequently, certain events resulting from the decision trigger an ex-post evaluation, utilizing another set of metrics and available data regarding the decision, at a specific time, to evaluate the actual outcomes and their impact and consequences. The ex-ante evaluation, the decision outcome, and the ex-post evaluation are within a particular decision context; however, they are governed by a set of conformance criteria spanning across multiple contexts. These criteria include reference values to which the entire decision should conform (e.g. benchmarks, KPIs, objectives, etc.).
The ex-post evaluation provides different types of feedback to enable experiential and organizational learning, either pertaining to the context of the decision or providing information and knowledge to other contexts as well. Such feedback serves as input to the other constructs and processes. First of all, it supports single-loop learning by updating the actions to resolve any errors. The feedback also provides knowledge for updating the ex-ante evaluation, and the opportunity to compare predicted outcomes with actual outcomes, thus making sense of the consequences and learning through experience. This could enhance future evaluations and predictions, as well as provide insight for revisiting the evaluation metrics and assumptions. Furthermore, learning from the feedback could entail updating the models themselves with which the alternatives were evaluated, which could lead to double-loop learning. Double-loop learning can occur when the feedback results in human or machine agents updating their reference points or perceived models of the environment, and when it leads to updating the conformance criteria for the decision.
Furthermore, this feedback can provide knowledge of the data required, and help detect discrepancies which may need to be solved. It can support rationalization and result in retrospective sensemaking. Finally, the feedback resulting from ex-post evaluation can enhance collaborative rationality by contributing to deutero learning. Here the decision makers can reflect on their collaborative, human-machine learning processes, and create new learning strategies for improving the collaborative rationality between them.
This model was then presented to an expert panel of four practitioners in the second organization
and validated. An example instantiation of the model was created for the chemical manufacturing plant
case, and an instantiation of the model was created for the customer churn case, together with two of
the experts, as a prototypical evaluation. Finally, the model was presented to four experts in a third
organization, with which we still aim to implement the model on their data-driven decision making
case, thus achieving additional validation and fulfilling the Eval 2 stage in Sonnenberg and vom
Brocke’s [
Since the structural model is static, and data-driven decision making is dynamic and requiring processual evaluation due to constantly changing contexts and consequences, a behavioral model depicting the relationship between the elements and their resulting behavior is still necessary and is the next stage in future work. It is intended to develop the behavioral model through collaboration with the third organization implementing data-driven decision making in order to observe the processual nature of decision making and evaluation, collaborative rationality between humans and machines, and the interaction between the elements and their impact on the decision. 2.4.
The models should then be demonstrated through a prototypical instantiation in the context of an
organization and evaluated (ex-post) in the Eval 3 and Eval 4 stages of Sonnenberg and vom Brocke’s
[
Future work thus includes practically implementing the model in a different case in the third
organization and studying the relationship between the constructs. This will lead to developing the
associated behavioral model to represent the dynamic aspects and the activities, sequences, flows, and
implementable relationships between the elements and their resulting behavior, in addition to the
behavior of collaborative rationality in each of the decision making stages through demonstration in a
practical setting. By understanding these process, we can then influence their change in the desired
directions [
Consequently, methods may be developed as a set of steps for manipulating the constructs so that
the solution statement models are realized. Instantiations can further be used to operationalize the model
and method, as a realization of the working artifacts in the environment [
Gregor and Hevner [
The design objectives contribute by guiding the development of a data-driven decision evaluation solution, based on knowledge gained from theory and practice. The static and behavioral models would serve as level 2 contributions, or nascent design theories. The instantiations of the models in organizations would serve as level 1 contributions. This would allow us to theorize and develop level 3 theories on data-driven decision making and collaborative rationality between humans and machines.
Hence, the main expected contributions of the research are as follows: • Suggesting and elaborating the theoretical concept of collaborative rationality, as well as enhancing it through deutero learning which results from evaluating data-driven decisions. • Highlighting the importance of a holistic, multi-faceted, and processual ex-post evaluation, while considering the different contexts, changing nature of the environment, and constantly challenging the evaluation criteria. • Creating a feedback loop for enabling learning, rationalization, and sensemaking from datadriven decisions to enhance future decisions. • Providing a modular structure for the practical implementation of the parts of the models and their relationship, resulting in a data-driven decision evaluation solution.
If the scope of the research allows, future work may also include developing a method for performing data-driven decision evaluation and design principles for potentially automating evaluation. The contribution and benefit of this research in science and industry is a new approach to evaluating, and possibly automating the evaluation of, data-driven decisions, further supported by rigorous research methods. Accordingly, the impact of this research can help develop good practices in data-driven decision making, and enhance learning from previous data-driven decisions, both for the decision maker and the machine. It can aid decision makers and policy makers in utilizing analytics to extract insights, driving them to make better quality and more informed decisions, and helping them overcome the problems and challenges faced with current practices. Finally, it can provide a better understanding of the collaboration between humans and machines in decision-making across various levels and contexts, as well as enable a collaborative rationality between both.
I would like to acknowledge my supervisors:
Professor Tero Päivärinta, M3S, Faculty of Information Technology and Electrical Engineering, University of Oulu, Finland
Professor Ahmed Elragal, Department of Computer Science, Electrical and Space Engineering, Luleå University of Technology, Sweden
I would also like to acknowledge the ITEA3 project Oxilate (https://itea3.org/project/oxilate.html), and the organizations and experts who have willingly collaborated in this research.