The Human Factor

As humans, we are frequently confronted with more options than unaided cognition can handle. To navigate this complexity, we increasingly rely on algorithms that augment our capacity for complex decision-making. A common assumption when deploying an algorithm for a decision-aid system is that its recommendations will be faithfully implemented.¹

In practice, however, the algorithm’s outputs can be accepted, adjusted, or rejected by the human decision-maker.² Neglecting this gap – both in algorithm design and evaluation – means overlooking the last, and arguably most important, step: real-world deployment. In other words, an algorithm is only as effective as the decisions it shapes, which may differ from the ones it prescribes.

Human-algorithm interaction takes two broad forms. In input-adjusting interactions, the decision-maker modifies the algorithm’s parameters (such as objective function weights or risk profiles), which in turn shape the algorithm’s output. In output-adjusting interactions, the algorithm’s recommendation is directly modified by the human before implementation. The former is arguably preferable, but the latter is very common in practice.³ 

Why Do Humans Ignore Algorithms?

Humans deviate from algorithmic recommendations for different reasons. Sun et al. categorize these deviations into two types: information deviations and complexity deviations.⁴ Information deviations occur when humans possess information that the algorithm lacks and consequently identify better solutions. This type of deviation is typically beneficial to operational performance. Complexity deviations, by contrast, stem from humans’ aversion or inability to fully implement algorithmic prescriptions, often because they are difficult to interpret, leading humans to prefer simpler but inferior solutions. 

Algorithmic deviation depends not only on the algorithm, but also on the human using it. Allen and Choudhury link algorithmic aversion to domain expertise.⁵ While domain knowledge can complement algorithmic advice and lead to superior outcomes, it can equally increase aversion, as more experienced users are more likely to deviate. Liu et al. find that drivers are less likely to follow algorithmic route recommendations when the recommendation conflicts with their past experience at a given location and time and when peers behave differently from the algorithmic suggestion.⁶ 

Martínez-de-Albéniz and Krigul examine the factors that lead couriers to deviate from algorithmically proposed routes, finding a preference for myopic decisions that yield immediate rewards but are suboptimal in the long run.⁷ This tendency toward myopic, locally scoped decisions is not unique to couriers. Humans in general tend to favor decisions restricted to their immediate area of responsibility.⁸ If workers use algorithms that seek a global performance metric while being themselves evaluated on narrow, local metrics, misalignment is inevitable and may even be reinforced by incentive structures that reward local performance.⁶ 

How Do We Promote Algorithmic Compliance? 

There are three main strategies to promote algorithmic compliance: control, explainability, and human-algorithm integration. 

Control

The most direct approach is to design systems that enforce compliance, removing or constraining worker discretion. This is the spirit of the Japanese concept of poka-yoke (mistake-proofing), which eliminates the possibility of error by design. Voice-picking systems in warehouses, for example, issue instructions and require spoken confirmation before releasing the next task. Yet workers routinely find ways to game these control systems.⁹  Control mechanisms can also take softer forms, such as incentive structures and penalties tied to adherence.

Explainability

A rapidly growing research agenda focuses on making algorithms transparent to the humans who use them, aiming to build trust among decision-makers. Explainability is especially important when recommendations emerge from opaque algorithms whose outputs may be accurate, but at the cost of invisible reasoning.

Explainability can be pursued in two ways: as a post-hoc layer applied to an existing algorithm (e.g., SHAP values for feature importance analysis), or by design, through the development of inherently interpretable algorithms (e.g., genetic programming or decision trees).¹⁰ Balakrishnan et al. distinguish several dimensions of algorithmic transparency, from an algorithm’s internal logic to the features it uses as input, noting that different applications call for different transparency angles.¹¹ The authors show that when noncompliance stems from the worker possessing genuinely superior information, algorithmic transparency alone does not resolve the issue, and feature transparency is more useful instead. 

Human-Algorithm Integration 

The least-explored but most promising effort to promote algorithmic compliance is to incorporate human knowledge and preferences directly into the algorithm’s design rather than explaining after the fact why the algorithm’s answer is correct.¹² This approach treats the gap between human and algorithm not as a compliance problem to be solved post-hoc, but as information to be captured from the start. 

Grand-Clément and Pauphilet developed an adherence-aware Markov Decision Process framework in which decisions are made sequentially over time, explicitly accounting for the fact that not all recommended decisions will be followed.¹ The likelihood of adherence is modeled as a source of uncertainty, integrated directly into the optimization. Sun et al. propose a “human-centric bin-packing algorithm” that uses machine learning to anticipate worker deviations and incorporate this anticipation into the prescriptive plan, thereby reducing deviations and improving overall performance.⁴

Where Do We Go From Here? 

When evaluating an algorithm, evaluation metrics should align closely with the goal of real-world impact. When a forecasting system informs a decision-making algorithm, rather than evaluate it solely by accuracy, it is preferable to judge it by the quality of the decisions it induces.¹³ This moves the algorithm’s evaluation closer to real-world deployment. Similarly, rather than assessing a decision-making algorithm on theoretical benchmark instances, it is better to measure its real-world impact after deployment. This shifts the moment of evaluation to after the human has decided, rather than judging the prescribed decision before the human can interact with it. 

Algorithmic design and human-factor management remain largely separate activities: an algorithm is designed, and an implementation strategy follows. To pursue their integration, it is necessary to connect fields that rarely speak to one another.^2,4 Operations research and machine learning provide the computational tools for making decisions at scale; behavioral and organizational science provides the frameworks for understanding why humans adopt or reject algorithmic guidance. 

Davis et al. identify doctoral training as an important lever for promoting this integration, encouraging students from each field to take courses in the other, ideally creating curricula that directly combine both.15 This research area is gaining traction: a recent special issue on human-algorithm interaction attracted 319 submissions, signaling an unprecedented level of interest from the management science community.² 

The Compliance Component 

The effort invested in theory generation should be matched by an equivalent effort in theory integration, so that the developed knowledge is translated into real-world practice.14 In operations research and management science, it is standard practice to subject a proposed algorithm to extensive computational experiments, such as varying parameters, stress-testing assumptions, and demonstrating robustness across configurations.

What if algorithmic evaluation also considered the compliance component? Is it truly outside the scope of our research to assess whether a recommendation will actually be used? We routinely test robustness to different assumptions, yet we almost never question the assumption that prescribed decisions will be faithfully implemented. 

References 

1. Grand-Clément J., Pauphilet J., 2026, “The best decisions are not the best advice: Making adherence-aware recommendations,” Management Science, Vol. 72, No. 11, pp. 667–92.
2. Caro F., Colliard J. E., Katok E., Ockenfels A., Stier-Moses N., Tucker C., et al., 2026, “Introduction to the Special Issue on the Human-Algorithm Connection,” Management Science, Vol. 72, No. 1, pp. 1–13.
3. Käki A., Kemppainen K., Liesiö J., 2019, “What to do when decision-makers deviate from model recommendations? Empirical evidence from hydropower industry,” European Journal of Operational Research, Vol. 278, No. 3, pp. 869–82.
4. Sun J., Zhang D. J., Hu H., Van Mieghem J.A., 2022, “Predicting human discretion to adjust algorithmic prescription: A large-scale field experiment in warehouse operations,” Management Science, Vol. 68, No. 2, pp. 846–65.
5. Allen R., Choudhury P., 2022, “Algorithm-augmented work and domain experience: The countervailing forces of ability and aversion,” Organization Science, Vol. 33, No. 1, pp. 149–69.
6. Liu M., Tang X., Xia S., Zhang S., Zhu Y., Meng Q., 2026, “Algorithm aversion: Evidence from ridesharing drivers,” Management Science, Vol. 72, No. 1, pp. 193–203.
7. Martínez-de-Albéniz V., Krigul C., 2025, “Human Agency in Last Mile Delivery”, under review, https://blog.iese.edu/martinezdealbeniz/files/2025/02/Human_Machine_Last_Mile_Deliveries_web.pdf.
8. Read D., Loewenstein G., Rabin M., 1999, “Choice Bracketing,” Journal of Risk and Uncertainty, pp. 171–97.
9. Cheon E., Erickson I., 2025, “Fulfillment of the Work Games: Warehouse Workers’ Experiences with Algorithmic Management,” Proceedings of the ACM on Human-Computer Interaction, pp. 1–30.
10. Rudin C., 2019, “Stop explaining black box machine learning models for high-stakes decisions, and use interpretable models instead,” Nature Machine Intelligence, pp. 206–15.
11. Balakrishnan M., Ferreira K., Tong J., 2022, “Improving human-algorithm collaboration: Causes and mitigation of over-and under-adherence,” https://www.smeal.psu.edu/lema/documents/jordan-tong-improving-human-algorithm-collaboration-pdf.pdf.
12. Powell W. B., Towns M. T., Marar A., 2000, “On the value of optimal myopic solutions for dynamic routing and scheduling problems in the presence of user noncompliance,” Transportation Science, pp. 67–85.
13. Elmachtoub A. N., Grigas P., 2022, “Smart ‘Predict, Then Optimize’,” Management Science, Vol. 68, No. 1, pp. 9–26.
14. Burton J. W., Stein M. K., Jensen T. B., 2020, “A systematic review of algorithm aversion in augmented decision making,” Journal of Behavioral Decision Making. Vol. 33, No. 2, pp. 220–39.
15. Davis A. M., Mankad S., Corbett C. J., Katok E., 2024, “OM Forum – The best of both worlds: Machine learning and behavioral science in operations management,” Manufacturing and Service Operations Management, Vol. 26, No. 5, pp. 1605–21.