Human-AI collaborative decision-making as an organization design problem

Source: Phanish Puranam; Journal of Organization Design 10:75-80 (open access) · 2021

Abstract The promise of collaboration between humans and algorithms in producing good decisions is stimulating much experimentation. Drawing on research in organization design can help us to approach this experimentation systematically. I propose typologies for considering different forms of division of labor between human and algorithm as well as the learning configurations they are arranged in, as basic building blocks for this endeavor.

Introduction. In this brief Point of View article, I offer some thoughts on how we may conceptualize collaborative decision-making between humans and AI algorithms as a problem in organization design. While there are many possible forms of interaction between humans and AI algorithms, the arguments here are most relevant to knowledge work in which humans and AI algorithms through some form of collaboration, together produce a decision that is implemented by a third party (for instance stock picking, investment, sentencing, screening candidates). I refer to these as situations of “Human–AI Collaborative Decision-Making” (or HACD). The arguments may also apply to situations which involve a human training an AI algorithm (e.g., self-driving cars that learn from observing humans drive) or vice versa (e.g., chatbot-based language learning applications), or the use of algorithms to improve matches between humans (e.g., friend suggestions on social media platforms), but will require additional considerations that I do not address here. In what follows, I use the terms “AI” and “algorithm” interchangeably with “machine learning” (ML). I am aware that not all algorithms are AI, and not all AI is machine learning (Broussard 2018; Raj and Seamans 2019), but my usage avoids tedium. Human–ML collaboration for decision-making is different from other forms of human interaction with/adoption of technology for at least two reasons. First, the outputs of decision tasks unlike physical production tasks can sometimes be aggregated in a manner that improves their accuracy through error cancellation. This means that division of labor without specialization is feasible in HACD. Second, there is the potential for mutual adjustment: both humans and AI based on machine learning algorithms are adaptive systems that change how they make decisions over time through learning from experience (i.e., past data). Organizational scientists understand the dynamics of mutual adaptation and know them to be distinctive from one sided-adaptation or static collaboration (Knudsen and Srikanth 2014; Lave and March 1993; Lounamaa and March 1987; Puranam and Swamy 2016). I make two main points in this article. First, I note that from the perspective of organization design, there are multiple possible configurations of division of labor in HACD besides the one that is most intuitive and currently dominates popular discourse, namely one based on specialization (human and AI each do different sub-tasks they are relatively best at). Second, organization design research also suggests that there are multiple learning configurations to consider in which humans and AI may “learn together”. I express these possibilities in terms of simple typologies. Together they describe a design space which may not yet be completely or densely populated, but which may serve to guide our explorations in terms of both practice and theory. * Phanish Puranam Phanish.puranam@insead.edu

Method. Taking an organization design perspective on the problem of human–AI collaborative decision-making (henceforth HACD) requires us to view the combination of the human and the algorithm as an organization—i.e., a multi-agent, goal-oriented system. The goal of an HACD organization is to produce a decision. The design of the organization constitutes the choices about division of labor (task division and task allocation) and integration of effort (information and reward provision, exception handling) that characterize the organization (Puranam 2018). Why should an organization that involves HACD ever be superior to an organization comprising only humans or only algorithms? A first cut at the problem is shown in Fig. 1. At any point in time, given the prevailing state of technology, we can speak of three types of tasks: Type A tasks are those in which algorithms equal or outperform humans (e.g., today that would include image or handwriting recognition). Type B tasks are the ones that humans outperform algorithms on (e.g., evaluating a job applicant’s integrity remains a Type B task, even though reading the CV might be a Type A task) and must remain in the hands of humans. However, by appropriate division of labor—including breaking up aggregate tasks into smaller ones which can be differentiated into Type A and Type B1—the human and algorithm each can do what they are better at. This unlocks gains from specialization (net cost of coordination between agents) through HACD. This logic lies at the heart of a lot of contemporary workflow and process automation. It is not fundamentally different from the calculus of outsourcing or offshoring or gains from trade as set out by Adam Smith and David Ricardo; and it applies to tasks in general, not only decisions. The more interesting case is Type C, where despite no clear superiority of either human or algorithm, the combination through aggregation may outperform either alone. A distinctive feature of decisions is that their accuracy can sometimes be improved through pooling and error cancellation (Larrick and Soll 2006; Rokach 2010; Surowiecki 2004). Having a human and an algorithm (or indeed several algorithms—as is the case with ensemble learning models) make the same decision and then aggregating their outputs can produce improved quality in terms of greater decision accuracy. This is not possible usually with physical products. For decisions that involve predicting a continuous variable (e.g., quality), the “wisdom of crowds” provides the intuition. For decisions that involve predicting a discrete category (e.g., hire or reject), Condorcet’s jury theorem provides a foundation, which illustrates how increasing the size of a jury of identically and modestly accurate members can increase the jury’s aggregate accuracy.

Building on this intuition of the differences between types of tasks, and by drawing on basic ideas in organization design, we can give a more comprehensive picture of the possible divisions of labor in HACD. Figure 2 illustrates the arguments below by giving hypothetical divisions of labor between humans and AI in the context of HACD for stock picking, such that an equity analyst and an algorithm might jointly make a recommendation on whether to buy a stock.

A division of labor involves (a) decomposition of the goal (the final decision) into tasks (decisions) that aggregate into the final decision and (b) allocating sub-clusters of these tasks across the organization’s members. The resulting allocation of tasks to agents—a division of labor—can be described in two ways. First, there will exist a structure of interdependence between tasks (and therefore between clusters of tasks allocated to different agents). Two tasks are interdependent if the value created when both tasks are performed is different from the sum of values created by performing each task alone. For instance, this could be because they draw on common rivalrous inputs, the value of outputs is super or sub-additive, or one is an input to another (Burton and Obel 1984; Milgrom and Roberts 1990; Thompson 1967).

Discussion. Considering HACD organizations with a joint emphasis on the nature of division of labor and the learning configuration can help us understand and design them better, both in terms of expanding the space of possibilities, as well as the precision with which we characterize particular points within them. Some models already exist in the organization design literature for each of the types of possible division of labor in HACD in at least some of the possible learning configurations. These models typically use adaptive reinforcement learning algorithms to simulate human decision-makers— but that should not prevent us from re-interpreting them as models of HACD, particularly once heterogeneity between agents is added to the picture. For instance, when the division of labor in decision-making involves specialization, feedback is often on group level output. Models of coupled learning have highlighted that the key design challenge in such situations is to avoid superstitious learning from false negatives and false positives (Lave and March 1993; Lounamaa and March 1987). Common priors and vicarious learning in the case of parallel specialized decisions (Puranam and Swamy 2016; Knudsen and Srikanth 2014; Aggarwal et al. 2017), and the stability of personnel in the case of sequential specialized decisions (Denrell et al. 2004) have been argued to mitigate the challenge. Coupled learning might also arise without specialization. Piezunka et al. (2020) study learning by participation, in which parallel unspecialized decision-makers receive feedback only on their aggregate decisions derived from voting. They point out that the quality of decisions over time depend on how the contrarians—those whose beliefs did not align with the majority vote at a point in time—influence future decisions. We also know that serial and parallel architectures lead to different learning dynamics even with isolated learners, because the inputs and therefore opportunities for learning are censored in serial architectures (Christensen and Knudsen 2013, 2020). However, these hardly exhaust the combinatorial space obtained by crossing possible divisions of labor with learning configurations. There is much to do in terms of completing our conceptual understanding of these possibilities, and even more to do in confronting the models with data. A partnership between organization design researchers and practitioners interested in HACD seems ripe with opportunity.