A new peer-reviewed model argues that managers who use generative AI as a shortcut under time pressure may gradually weaken the judgment they develop through experience. The same technology may have the opposite effect when managers must explain and defend their decisions.
Researchers from the University of Bath, Ohio State University, Cardiff University and the University of Lausanne developed the model for the Academy of Management Review. The paper was published online on August 25 and publicized by the University of Bath on September 4.
The work is theoretical, not an experiment measuring what happened to a group of managers. Its authors set out two possible paths: AI can reduce the effort managers put into learning from a situation, or it can prompt them to examine their reasoning more closely.
Which path is more likely may depend less on the software than on the workplace surrounding it.
What managers may lose when AI does the first thinking
The researchers focus on managerial phronesis, an academic term for the practical wisdom used to make moral and context-specific decisions. It includes knowing what deserves attention, reading the circumstances around a problem and applying lessons learned through experience.
Those abilities are difficult to reduce to a checklist. A manager dealing with a supplier dispute, a safety concern or an employee problem must usually weigh facts alongside relationships, local knowledge and the possible consequences for other people.
Generative AI works differently. It produces an answer by finding patterns in data and generating new content from them. It does not experience the workplace or live with the outcome of a decision.
The authors propose that repeated reliance on synthetic answers can create what they call “epistemic de-skilling.” In plain language, a manager may become less capable of forming knowledge independently because too much of the thinking has been handed to a machine.
The risk is broader than receiving an inaccurate answer. A manager may stop gathering firsthand information, asking colleagues for different views or examining why a proposed action appears sensible.
Time pressure can turn assistance into substitution
Time pressure is central to the model. A manager facing a deadline has a strong incentive to use AI for a quick recommendation and move on. The immediate benefit is visible: the task is finished sooner.
The possible cost develops more slowly. If the system repeatedly supplies the first interpretation, the manager has fewer opportunities to encounter the uncertainty and disagreement through which judgment is formed.
Related evidence supports treating this concern seriously, while stopping short of proving the new model. A 2025 study presented at the ACM CHI conference surveyed 319 knowledge workers about 936 examples of using generative AI at work. Greater confidence in the AI was associated with less reported critical thinking, while greater confidence in the worker’s own abilities was associated with more.
That earlier study relied on workers describing their own behavior, so it did not establish that AI caused their critical thinking to decline. It did show why trust, confidence and the demands of a task deserve attention alongside the accuracy of the technology.
Accountability may produce the opposite result
The Bath-led team also describes a more productive route, which it calls “epistemic up-skilling.” AI can present alternatives, expose assumptions or produce an explanation that a manager then has to question.
Accountability is the deciding condition in this part of the model. A manager who knows a decision must be justified to colleagues, employees, customers or a regulator has more reason to examine the AI’s answer, compare it with direct evidence and explain any judgment the system cannot make.
For businesses, this points to a gap between placing a person at the end of an automated process and making that person responsible for the reasoning. A routine approval click provides little protection if the manager cannot explain which evidence was checked, which alternatives were rejected and who may be affected.
Professor Dirk Lindebaum of the University of Bath said managers should use AI to challenge assumptions and test their reasoning instead of accepting its output at face value. “It is that which Gen-AI cannot satisfactorily explain that managers must explain to themselves and others,” he said.
Speed is easier to count than lost capability
The paper arrives as AI is moving from occasional workplace assistance into routine processes. As we recently reported, software companies are beginning to charge businesses for work completed by AI agents, tying commercial models more closely to automated output.
Companies can measure the time an AI system saves, the number of tasks it completes and the cost of each interaction. A gradual decline in employees’ ability to frame a problem, challenge an answer or recognize an unusual situation is much harder to see in a performance dashboard.
This creates a measurement problem for employers. An AI-assisted decision may look efficient today even if the process gives managers less opportunity to develop the experience needed for a harder decision tomorrow.
The danger should not be exaggerated. The new paper offers a process model that still requires empirical testing, and it explicitly allows for AI to strengthen judgment under the right conditions. Its contribution is to identify the workplace incentives most likely to push use in either direction.
Accuracy checks remain necessary, as our earlier coverage of inaccurate AI shopping advice and customer trust also showed. For managerial decisions, one further question may be revealing: can the person responsible explain the decision without asking the system to explain it for them?
