Roles of Artificial Intelligence in Collaboration with Humans: Automation, Augmentation, and the Future of Work

Source: Andreas Fügener, Dominik D. Walzner and Alok Gupta; Management Science 72(1) · 2026

Abstract. Humans will see significant changes in the future of work as collaboration with artificial intelligence (AI) will become commonplace. This work explores the benefits of AI in the setting of judgment tasks when it replaces humans (automation) and when it works with humans (augmentation). Through an analytical modeling framework, we show that the optimal use of AI for automation or augmentation depends on different types of human-AI complementarity. Our analysis demonstrates that the use of automation increases with higher levels of between-task complementarity. In contrast, the use of augmentation increases with higher levels of within-task complementarity. We integrate both automation and augmentation roles into our task allocation framework, where an AI and humans work on a set of judgment tasks to optimize performance with a given level of available human resources. We validate our framework with an empirical study based on experimental data in which humans classify images with and without AI support. When between-task complementarity and within-task complementarity exist, we see a consistent distribution of work pattern for optimal work configurations; AI automates relatively easy tasks, AI augments humans on tasks with similar human and AI performance, and humans work without AI on relatively difficult tasks. Our work provides several contributions to theory and practice. The findings on the effects of complementarity provide a nuanced view regarding the benefits of automation and augmentation. Our task allocation framework highlights potential job designs for the future of work, especially by considering the often-ignored, critical role of human resource reallocation in improving organizational performance.

Introduction. The future of work will be characterized by the diffusion of artificial intelligence (AI) into almost all areas of the workplace. Indeed, recent developments in areas of deep learning (LeCun et al. 2015) have allowed AI to perform judgment tasks, such as image recognition, object detection, or speech recognition, on par with or even better than humans. Furthermore, AI can make decisions autonomously, which means that it has agentic qualities formerly exclusive to human decision makers (Baird and Maruping 2021).

In most environments that require complex decision making, human-AI collaboration can achieve even better performance than humans or AI working alone (Dellermann et al. 2019). One example is skin cancer detection. Although modern AI outperforms even experienced dermatologists (Esteva et al. 2017), studies indicate that combining AI with physician estimates can yield even better results (Hekler et al. 2019). However, a possibility that has not yet been discussed is that of a heterogeneous approach, in which physicians are sometimes substituted with AI, sometimes work together with AI, and sometimes perform cancer detection without help from AI. The human resources made available by automation in some cases could be reallocated so that groups of humans could solve more challenging cases.

In this paper, we explore AI’s role in collaborating with humans on judgment tasks, assuming that humans and AI have complementary skills. We differentiate between automation, where AI substitutes for humans, and augmentation, where AI supports humans by providing advice. To our knowledge, there is a lack of structural investigation into the benefits that AI could provide through these roles, the effect of complementarity on the benefit of collaborating with AI, and the distribution of work between humans and AI.

We approach this investigation in two ways. First, we propose a task allocation framework in which tasks can be performed by one of three work constellations: by humans, by AI (automation), or by humans with AI (augmentation). Compared with the status quo setting, in which each task is performed by one human, we analyze the benefits that AI can provide through automation and augmentation. One of the primary benefits of automation is that AI can replace humans in tasks on which it performs better (substitution benefit). Another benefit is that the human resources released through automation can then generate value by working on other tasks (reallocation benefit). Through augmentation, AI provides benefit by interacting with humans and thus, improving their performance (augmentation benefit). We theorize that these benefits are dependent on two types of complementarity: on the potential complementarity because of performance differences between humans and AI on different tasks (between-task complementarity) and on the potential complementarity because of the interaction of humans and AI on the same task (within-task complementarity). In the second part of our investigation, we validate our findings with data from an experimental study in a context where real humans performed a set of image classification tasks either alone or with the help of AI. The setting of image classification was selected because it is a generic, idealized judgment task that is highly suitable for validating our theoretical findings. Because state-of-the-art AI solutions exist and both between-task complementarity and within-task complementarity arise, the context is suitable for validating the findings of our theoretical discussion. Our paper aims to answer the following research questions.

1. How does the type of complementarity between humans and AI in judgment tasks affect the different benefits that AI can provide through automation and augmentation?

2. How does the introduction of AI for automation and augmentation in judgment tasks affect the distribution of work between humans and AI?

Our study yields three main findings. First, we theoretically and empirically show that the benefits of AI depend on the type of complementarity. The benefits that AI can provide through automation (substitution and reallocation benefit) increase with between-task complementarity between humans and AI because humans and AI are allocated to those tasks where each can provide the greatest benefit. In contrast, the benefit that AI can provide through augmentation increases with within-task complementarity between humans and AI. Second, our framework, which combines automation, augmentation, and reallocation of the human resources released through automation, yields superior results to those obtained with only automation or augmentation. In our empirical study, humans achieved an accuracy of 68%, automation achieved 77%, and augmentation achieved 80%. With the same human resources, our framework yielded an accuracy of 88%, greatly outperforming the other options.

Related work. Our research contributes to two streams of literature. The first literature stream explores effective collaboration designs for humans and AI that can be viewed as constituting a “multiagent, goal-oriented system” (Puranam 2021, p. 76). In this system, different work constellations can be considered for a task where AI works alone (automation), works together with humans (augmentation), or is not used at all. Given the different types of complementarity discussed above, a crucial F ̈ugener, Walzner, and Gupta: Roles of AI in Collaboration with Humans Management Science, 2026, vol. 72, no. 1, pp. 538–557, © 2025 The Author(s) 541 question for organizations is when to prioritize which work constellation. Empirical evidence supports the idea that organizations lose potential when they rely only on a single work constellation. Kesavan and Kushwaha (2020) conducted a field experiment at a spare automobile parts retailer, where humans had the option to override a data-driven decision-making tool. They found that humans overriding the tool (augmentation) performed better than the tool (automation) for growth-stage products, for which little data are available. In contrast, augmentation performed worse than automation for mature- and declining-stage products, for which large amounts of data are available. Their results indicate that tasks with little data are difficult for AI tools, which means that these are tasks for which humans can provide complementary knowledge. In contrast, tasks with large amounts of data may be relatively easy for AI tools.

To determine which work constellation to use, different frameworks have been proposed in the literature. So-called triage models have been introduced as a way to decide ex ante who should perform a given (prediction) task. A triage model is a type of metamodel that is trained to predict when an intervention by a human is beneficial for task performance (Raghu et al. 2019). Many recent papers discuss the learning-to-defer approach, where the AI learns to decide whether it should perform a task on its own or should defer the task to a human (Madras et al. 2018, Mozannar and Sontag 2020). Chen et al. (2023) presented a framework to decide when to rely on human adjustments to algorithmic demand forecasts (augmentation) and when to rely only on the algorithmic forecast (automation). Although all of these frameworks provide helpful insights into effective human-AI collaboration design, structural investigation regarding the factors that drive the effectiveness of each work constellation is lacking.

Method. In this section, we present a novel task allocation framework that allocates humans and AI to a given set of judgment tasks. The framework is inspired by Michelucci and Dickinson (2016), who present a conceptual overview of distributed systems in a crowdsourcing context. We use our framework to analyze the distribution of work between humans and AI as well as the benefits that AI can provide through automation and augmentation. Here, benefit is defined as the improvement in task performance relative to the baseline case, in which each task is performed by one human. Figure 1 illustrates the framework.

To describe the general characteristics of the type of tasks that we deem appropriate for our framework, we focus on two conditions. First, the Polanyi condition states that the procedure for optimally solving the tasks cannot be formalized; therefore, these tasks cannot be solved—either by AI or by humans—with absolute certainty (Autor 2015). Thus, a human or an AI will solve those tasks differently. This condition ensures the existence of complementarity, which is necessary to generate benefits through human-AI collaboration. An example of a task that does not satisfy this condition is a deterministic optimization problem, for which the optimal solution can simply be calculated. Here, no complementary knowledge of humans and AI exists. Second, the ground truth condition requires that the tasks follow a certain structure. More specifically, the task needs to have one true answer (ground truth), which the decision maker must ascertain from a set of options (McGrath 1984). Thus, humans and AI can solve the task correctly with a certain probability. The condition also states that a meaningful aggregation mechanism for crowd-based decision making exists, which enables the combination of individual choices into an aggregated choice: for example, through majority voting for classification tasks or through averaging for regression tasks.

This formulation is an extension of that of Boland et al. (1989), who provided a model for wisdom-of-crowds decision making with dependent decision makers. Our formulation adapts their model to the case of advice taking and extends the model by considering the heterogeneous effects of correct and incorrect AI advice. The effect of augmenting a crowd of humans with AI depends on whether the correct choice is the most likely choice before and after receiving advice. In the first case, humans without advice select the correct choice with the highest probability, and their performance converges to one as the crowd size increases. If the correct choice remains the most likely choice after receiving incorrect advice, the performance converges to one as well. If an incorrect choice is the most likely choice after receiving incorrect advice, the performance converges to that of AI. In this case, advice is detrimental for a large number of humans, and providing AI advice decreases performance by 1 pAI t . In the second case, an incorrect choice is the most likely choice without advice, and the performance converges to zero. If the incorrect choice remains the most likely choice after receiving correct advice, the performance converges to zero as well. If the correct choice is the most likely choice after receiving correct advice, the performance converges to the performance of AI. In this case, providing advice is beneficial for a large number of humans but is limited to the benefit of automation as providing AI advice increases performance by pAI t . Note that we refrain from discussing cases in which the probabilities of the correct choice and the most likely incorrect choice are exactly equal before and after receiving advice.

We make the following assumptions. Each task consumes the same amount of human resources if it is allocated to a human. A sufficient number of individual human workers exist such that multiple humans can be allocated to a task. Thus, when a task is allocated to multiple units of human resources, we assume that these resources come from different humans, which is a requirement for our wisdom-of-crowds logic.

Discussion. The empirical results validate the theoretical statements on the effects of between-task and within-task complementarity on the distribution of work between humans and AI and on the benefits that AI can provide through automation and augmentation. Based on these results, we derive meaningful theoretical and managerial implications. We believe that these implications hold for any set of (judgment) tasks that satisfy the two conditions specified in Section 3 (the Polanyi condition and the ground truth condition).

Although the general notion of previous human-AI research seems to be that augmentation is preferable to automation, our analysis suggests that a more nuanced view is necessary to understand the benefits of AI. Our task allocation framework highlights the central tenet of Raisch and Krakowski (2021), who argue that strong interdependencies between the roles of automation or augmentation exist and that these roles, therefore, need to be analyzed simultaneously. Our work demonstrates that the benefits that AI provides through automation and augmentation depend on the type of complementarity between humans and AI.

Our results also corroborate the findings of Dell’Acqua et al. (2023), who argue that the capabilities of AI can be defined along a “jagged technological frontier” (Dell’Acqua et al. 2023, p. 1). For tasks that lie within the frontier (e.g., summarizing ideas), AI can provide benefits by automating the task or by augmenting humans who work on the task. For tasks that lie outside the frontier (e.g., providing strategic recommendations to a company), AI cannot provide any benefits, which means that humans should work on these tasks without AI support.

Applying our framework to the skin cancer detection example from Section 1, this could lead to the following setup. In the first step, all of the cases are evaluated by the AI. All cases with high AI certainty are automated by the AI. In cases with medium AI certainty, physicians review the AI suggestions, and a team of physicians is allocated to cases that are highly uncertain from the AI perspective. In those cases, the physicians who are not required for the automated cases can help improve the detection performance in the most challenging cases.

In our task allocation framework, the contribution of humans lies in improved performance on tasks where AI may be inferior. The degree to which the human contribution is meaningful depends on the existence of complementarity between humans and AI. Complementarity is a decisive factor that determines whether humans can provide benefits in the future of work. As complementarity is driven by the knowledge gap between humans and AI, it is crucial to critically evaluate whether this knowledge gap will persist in the future of work. We believe that developments in three areas will affect this knowledge gap: (1) improvements in AI, (2) changes in task and process design, and (3) loss of tacit knowledge.

Conclusion. It is very exciting to witness the ever-growing importance of AI in the discussion of the future of work. With this paper, we contribute to this discussion by theoretically outlining the benefits that AI can provide when collaborating with humans. Our task allocation framework and our empirical analyses show that the benefits of AI depend on the existence of between-task and within-task complementarity. AI can support humans in many capacities by taking over tasks, augmenting humans, or allocating humans to tasks where AI cannot provide benefits. In summary, humans and AI will collaborate closely, and the division of labor and their respective contributions will be heavily influenced by how well they complement each other.

Limitations. With our task allocation framework, we aim to provide a starting point for discussing the allocation of tasks between humans and AI under different types of complementarity. Many exciting research opportunities exist to extend or adapt the current framework.

The first set of extensions concerns the adaptation of the framework’s objective function. In our current version, we maximize the average task performance while keeping human resources (i.e., costs) and the number of tasks fixed. Alternative approaches could choose a different objective; one could minimize the human resources (costs) while keeping performance and the number of tasks fixed. Further, one could maximize the number of completed tasks while keeping performance and human resources fixed.