Human-AI Ensembles: When Can They Work?

Source: Vivek Choudhary, Arianna Marchetti, Yash Raj Shrestha, Phanish Puranam; Journal of Management 51(2):536-569 (online 2023) · 2025

An “ensemble” approach to decision-making involves aggregating the results from different decision makers solving the same problem (i.e., a division of labor without specialization). We draw on the literatures on machine learning-based Artificial Intelligence (AI) as well as on human decision-making to propose conditions under which human-AI ensembles can be useful. We argue that human and AI-based algorithmic decision-making can be usefully ensembled even when neither has a clear advantage over the other in terms of predictive accuracy, and even if neither alone can attain satisfactory accuracy in absolute terms. Many managerial decisions have these attributes, and collaboration between humans and AI is usually ruled out in such contexts because the conditions for specialization are not met. However, we propose that human-AI collaboration through ensembling is still a possibility under the conditions we identify.

Introduction. There is considerable interest today in the potential for collaboration between humans and artificial intelligence (henceforth, AI) technologies. Contemporary AI is based on machine learning (henceforth, ML) techniques that allow computers to learn solutions from data rather than be explicitly programmed (Bishop & Nasrabadi, 2007; Goodfellow, Bengio, & Courville, 2016; also refer to Csaszar & Steinberger, 2021 for a review of the literature). While the core concept is not new, recent algorithmic advances coupled with expansion in processing power and the ever-increasing availability of digital data have made AI viable in ways that were not possible before, resulting in a wave of enthusiastic adoption with various applications spanning research and practice. Organizations have started exploring how to improve managerial performance by employing a combination of humans and AI—rather than either alone—to tackle a variety of problems, and scholars have begun to study the antecedents and consequences of such efforts (e.g., Murray, Rhymer, & Sirmon, 2020; Shrestha, Ben-Menahem, & Krogh, 2019). A characteristic of prior attempts to combine humans (henceforth, H) and AI to perform a task1 has been the emphasis on division of labor with specialization. Specialization implies that H and AI perform different (sub-)tasks and their distinct outputs are then combined to generate the final output.2 Division of labor with specialization involves redefining the task division and task allocation between agents to exploit their respective advantages, in terms of superior output and/or lower cost of labor (Canetti et al., 2019; Holzinger, 2016; Murray et al., 2020; Dastin, 2018). Consider, for instance, a hiring task that involves two sub-tasks: screening the application pool and interviewing selected candidates. AI can be used to automate the first sub-task by screening applicants’ resumes and shortlisting candidates, and H to conduct in-depth interviews of the selected few (Pessach et al., 2020). A more subtle form of specialization involves H checking the work of AI or training it (on the presumption that H has superior capabilities to do so). In sum, division of labor with specialization arises when an original task can be decomposed into sub-tasks and those can be allocated across actors based on their relative competence. As we will show, the wide variety of H-AI collaboration literature (including work on the “human in the loop” configuration, e.g., Holzinger, 2016; Ostheimer, Chowdhury, & Iqbal, 2021) implicitly assumes a division of labor with specialization. However, division of labor with specialization is not the only possibility for H-AI collaboration. In this paper, we consider an alternative that has relevance when the task involves decision-making based on a prediction: A division of labor without specialization. The idea of aggregating the predictions of multiple decision makers tackling the same decision problem has a rich heritage in the literatures on human decision-making (Nisbet, Elder, & Miner, 2009; Page, 2010, 2014; Tumer & Ghosh, 1996) and in computer science on aggregating multiple prediction models (e.g., Polikar, 2012; Sagi & Rokach, 2018; Zhang & Ma, 2012). In keeping with the latter, we use the term “ensembling” to denote aggregation of predictions from multiple models applied to the same prediction problem (Sagi & Rokach, 2018).3 In this form of division of labor, H and AI tackle the same prediction problem, but their different predictions are aggregated in some way (e.g., by averaging for estimation problems; quorum, plurality, or unanimity for classification problems) to arrive at a final output (Puranam, 2021). Whereas specialization requires H and AI to perform distinct (sub-)tasks, ensembled H and AI tackle the same task (Csaszar & Steinberger, 2021: 20). The concept of ensembling H-AI is orthogonal to the ideas of “augmentation” and “automation” (Raisch & Krakowski, 2020) and can result from either, as we also show in this paper. Augmentation of a set of human workers occurs when an AI is added to the humans to improve the performance of a task. Automation, on the other hand, implies that algorithms replace (at least some) humans in performing the task. Ensembling can be achieved through augmentation if AI is added to (a group of) humans, or thorough automation if AI replaces (some but not all) humans in the group (refer to Table 1 for an illustration). The appeal of ensembling is the potential for improving decision accuracy when decision makers vary in the errors they make on the same task (Nisbet et al., 2009; Page, 2010, 2014; Steyvers, Tejeda, Kerrigan, & Smyth, 2022; Tumer & Ghosh, 1996). This diversity in errors can (although need not always) be beneficial when two agents make opposite prediction errors: one underestimates and the other overestimates the outcome value.

Related work. Prior Literature on H-AI Collaboration in Decision Tasks In this section, we review the relevant literature on how H and AI can be combined for prediction-based decision-making. We first highlight the close link between prediction and decision-making, and why AI algorithms, despite their considerable power at prediction tasks, do not prove universally superior to human decision makers. Next, we review research pertaining to the gains from specialization of H and AI to sub-tasks they are each superior at, as well as the literature on the use of H as superior decision makers that can act as gatekeepers or trainers of algorithms. We conclude this section by noting that the possibility of ensembling between H and AI—which involves neither specialization nor AI superiority to H in predictive accuracy—has so far remained under-explored.

AI as an Aid to Prediction and Decision-making All decisions under uncertainty ultimately require prediction (Agrawal et al., 2018), although a prediction may not be enough to make a decision (Bertsimas & Kallus, 2020). For instance, when evaluating projects to select between alternatives—for example, which candidate to hire, which project to invest in, or which course of action to pursue for a given strategy—the decision maker predicts the corresponding outcome of each available option and selects the alternative most likely to yield the best outcome. Decisions may involve predictions in the form of (i) estimation (i.e., deciding on the value of a variable, for instance how much to invest in a new project), or (ii) screening (e.g., accepting or rejecting a proposal, such as hiring a candidate for a particular job). Estimation is also referred to as a “regression” problem, and screening as a “classification” problem in the AI literature (Hastie, Friedman, & Tibshirani, 2009). Existing research documents that AI algorithms can be helpful in tackling decision problems involving regression and classification.5 In contrast to traditional optimization algorithms, such as branch and bound, dynamic programming, and integer programming used in the development of expert systems (a form of “traditional” AI) based on exact predetermined rules to identify a solution (Kanet & Adelsberger, 1987; Lee, Chen, & Wu, 2010), contemporary AI algorithms that are based on ML (which we focus on in this paper) do not require as much a priori knowledge of the problem structure, or clearly defined rules, to generate a solution. Instead, they can approximate it based on models fit to data.

Method. Theory: The Distinctive Benefits of H-AI Ensembles The distinctive feature of ensemble decision-making is that all its members perform the same prediction task, that is, there is no specialization (Anderson, 2019: 23; Brown, 2010: 1). Ideas relating to ensembles have been extensively studied in both ML (e.g., boosting, bagging, stacking, cross-learning algorithms) and the social sciences (e.g., Condorcet’s jury theorem, wisdom of the crowd, pooling of experts). The existing literature identifies benefits not only from ensembling estimation tasks (e.g., Page, 2007), but also screening decisions (e.g., through voting systems) and probabilities or quantile estimates in various ways (Becker, Guilbeault, & Smith, 2021; Lichtendahl, Grushka-Cockayne, & Winkler, 2013; O’Hagan, 2006; Ranjan & Gneiting, 2010; Thomas & Ross, 1980). However, an attempt to understand how and when ensembling H and AI can be useful in (managerial) decision-making is novel. We build on Steyvers and colleagues’ (2022) discussion of the benefits of hybrid systems where H and AI work together for predictive accuracy, as well as on the differentiation of H’s and AI’s errors in prediction. We develop our theoretical arguments in three stages. First, we explain why diversity in predictions is necessary but not sufficient for ensembles, rather than just H or just AI decision makers, to be useful. Next, we explain the source of diversity between H and AI predictions. This leads us to identify the precise conditions under which H-AI ensembles will dominate pure H or pure AI decision-making in project evaluation (Proposition 1). Finally, given that an ensemble between H and AI can be formed either by AI replacing one of the existing H (automation) or by adding to the existing H in a team (augmentation), we also derive the conditions under which the former is more likely to add value than the latter (Proposition 2). While we describe our arguments in terms of predictive accuracy, its most direct application in the domain of managerial decision-making is for project evaluation in data rich contexts, such as hiring employees, investing in projects, selecting suppliers and strategic partners, acquiring firms, or launching new products (Christensen & Knudsen, 2010; Csaszar & Ostler, 2020).

Why Diversity in Predictions Is Necessary (But Not Sufficient) for Ensembles to Be Useful Diversity in predictions (and therefore in prediction errors, i.e., the difference between the predictions made and the actual outcome) made by agents is necessary but not sufficient to improve the accuracy of their ensembled prediction. The intuition is most simply expressed for the case of a point prediction task, based on a result known as “ambiguity decomposition” (Krogh & Vedelsby, 1995). This theorem exists in multiple fields related to statistics, with the version given by Scott Page as the “diversity prediction theorem” being the most accessible (Page, 2010), and widely popularized (by Surowiecki, 2005) as the “wisdom of the crowd.” It posits that the error of a crowd’s estimate (which is the average of its members’ estimates) is systematically lower than the average individual error (where the error is expressed as the square of the difference between estimate and truth) as long as diversity (i.e., the sum of the squared difference in estimates) across individuals is positive. The key expression can be written as follows:

The ensemble (in this case, the crowd’s prediction) can thus be expected to always perform better than the average accuracy of its members in a single estimation task, which represents the right benchmark if one cannot hope to learn which of the ensemble members is likely to be the most accurate (Page, 2010). However, we might be interested in knowing whether ensembling can help us improve on the best individual predictor, or how we could improve the ensemble’s accuracy. Unfortunately, the identity in (1) neither guarantees that the crowd (i.e., the ensemble) always beats the best (i.e., most accurate) individual, nor that increasing the diversity of crowd predictions will necessarily increase its accuracy.

Discussion. Mirroring the ever-increasing reliance on AI witnessed across various industrial domains, including the recent breakthroughs with generative AI and the ChatGPT applications, organizational researchers have begun to analyze it in theoretical terms (e.g., Balasubramanian et al., 2022; Lindebaum, Vesa, & Hond, 2020). Many have focused on identifying task and system characteristics and dimensions under which the human or algorithmic component may be best suited to complete the task (e.g., Dellermann et al., 2019; Jarrahi, 2018; Murray et al., 2020). Others have studied how to optimally include superior human judgement as an input to algorithmic processes, either in the form of gatekeeping (Canetti et al., 2019; Dastin, 2018), or human-in-the-loop configurations, where H train AI (Bhardwaj, Yang, & Cudré-Mauroux, 2020; Holzinger, 2016; Jain et al., 2019). Yet, these do not account for the possibility of ensembling decision-making by H and AI— which does not involve specialization, or an advantage of H over AI, or vice versa. Moreover, it can be useful even when neither H nor AI on their own attain satisfactory accuracy in their predictions. Ensembling, because of the observability of the counterfactual and the relatively easy reversibility, offers the luxury of being easily discarded when we feel reasonably confident that it will no longer be beneficial, and allows for that belief to be checked continuously. This observability of the counterfactual is particularly useful in dynamic environments due to the chances of AI model decay (i.e., reduction in accuracy) caused by data shifts, where managers can continuously evaluate AI models’ accuracy as compared to their own decisions. Additionally, as H are involved in the exact same decision-making, H-AI ensembling also avoids the risk of H losing decision-making capabilities, which is the case with replacement-based approaches (Balasubramanian et al., 2022). Our paper contributes to the literature on H-AI collaboration in three ways. First, we delineate the possibility for collaboration between H and AI through ensembling, a form of division of labor without specialization. Second, we identify the precise data availability conditions that are likely to make H-AI ensembling attractive to project evaluation tasks performed by managers in data-rich contexts. Distinguishing between AI-accessible data and H-only accessible data, we argue that the human value in ensembles must ultimately rely on unique H-specific and machine inaccessible data. The existence of such H-specific data need not be static and can evolve with every decision. Further, H so far have the lead in the ability to learn from unrelated tasks and transfer their learning to the task at hand, thereby constantly enriching the H-only accessible data. This may change if research in AI makes significant progress on the problem of transfer learning (Zhuang et al., 2020). Third, we also explain why, under the constraint that H must always be part of an ensemble, the H-AI ensemble can be composed either by augmentation or automation (Raisch & Krakowski, 2020). In automation, a human agent is replaced or substituted by an AI, whereas in augmentation the AI is added to the ensemble to complement the human agents (Tschang & Almirall, 2021). Somewhat surprisingly, in this situation H are most at risk of replacement when they exhibit less diversity relative to each other in terms of prediction errors, and not because they have lower accuracy than AI. In fact, it is not necessarily the H with the lowest accuracy who is at greatest risk of being replaced in the ensemble but, rather, the one who adds the least diversity of prediction error relative to the AI.

Conclusion. The possibility of ensembling H and AI algorithms for prediction-based project evaluations that, currently, neither perform well individually, expands the frontiers of H-AI collaboration beyond what can be accomplished through division of labor with specialization (including relying on human superiority to act as a gatekeeper or trainer). Since the properties of many important managerial decisions concerning project evaluation (in particular, their unknown underlying structure and the limits of data on past decisions) make it difficult for AI to attain satisfactory stand-alone performance in the near term, we suggest that ensembling AI with managers is a possible avenue worth exploring.

Limitations. We acknowledge that our theory on H-AI ensembling does not account for all possible interactions between H and AI. First, not all forms of collaboration between H and AI involve prediction-based decisions. For instance, as noted above, our theory does not directly apply to ensembling of qualitative outputs between H and generative AI. Combining the creative outputs (e.g., text, images, music) of H with that of AI is not identical to static error cancellation, but may be closer to the dynamics of learning by doing and coupled search in complex search environments (e.g., Knudsen & Srikanth, 2014; Puranam & Swamy, 2016), which requires further investigations.

Second, we acknowledge that not all managerial decision-making is suitable to H-AI ensembling. While project evaluation has been used to model the overall decision-making process of managers and firms (e.g., Christensen & Knudsen, 2010; Csaszar & Ostler, 2020), and encompasses a wide array of managerial decision-making scenarios, this class of problems does not comprise everything managers do, and excludes, for instance, motivating employees, or complying with legal requirements.