The University of Bath has highlighted research on how generative AI may alter managerial judgment. Its September 4 announcement describes two possible paths. Managers can outsource enough thinking to weaken practical wisdom, or they can use AI to challenge assumptions and deepen reflection. The difference lies in how work and accountability are designed around the tool.
The paper, A Process Model of Managerial Phronesis in the Age of Generative AI, appears in the Academy of Management Review. The public University account presents a conceptual model rather than a controlled workplace trial. It proposes mechanisms for "epistemic de-skilling" and "epistemic up-skilling" without reporting a universal rate of skill loss or a measured causal effect from any particular AI model.
That boundary is important because the managerial risk is easy to overstate. The research does not show that every use of generative AI makes a manager less capable. It asks how repeated patterns of use might shape the practical wisdom that develops through experience, reflection, social interaction and responsibility for consequences.
A completed answer can conceal an unfinished decision
Imagine a manager preparing to explain a change in a team's schedule. A generated draft arrives within seconds. It is clear, tactful and apparently complete. Yet the manager may still lack the facts that determine whether the decision is sound: who faces the greatest disruption, which earlier commitments matter, what alternatives were considered and whether the explanation matches what actually happened.
The model has completed a writing task. The managerial decision remains unfinished.
That gap makes phronesis , often translated as practical wisdom, relevant to AI-assisted work. Phronesis concerns judgment about how to act in a particular situation. A fluent general answer cannot supply firsthand knowledge of these employees, these obligations and these consequences merely by sounding reasonable.
An AI-assisted decision therefore needs at least two evaluations. The first asks whether the generated content is accurate and useful. The second asks whether the responsible person has examined enough context to justify acting on it. A strong answer to the first question cannot substitute for the second.
This is why fluency can become a managerial hazard. Language models are optimized to produce coherent continuations, and their polished output can create a premature sense of closure. The document looks finished, so the decision feels finished. The missing work is often less visible: checking the local facts, weighing competing obligations and deciding which consequences are acceptable.
Accountability changes the role AI plays
Bath's announcement associates time pressure and uncritical reliance with epistemic de-skilling. When managers use generative AI as a shortcut, they may ask fewer questions, seek fewer perspectives and learn less from direct interaction. Over time, the process model suggests, this can weaken the knowledge and practical judgment needed for complex decisions.
The constructive path also starts with workflow. The researchers describe epistemic up-skilling when managers use AI as a tool for reflection, test its reasoning, explore alternative scenarios and fill the explanatory gaps themselves. Accountability matters because a person who expects to justify a decision has a reason to inspect the output instead of merely forwarding it.
The decisive design question is therefore where explanation occurs. If explanation is requested only after a decision fails, it becomes a reconstruction. If it is required before action, it can reveal a missing fact or unsupported assumption while there is still time to change course.
In the scheduling example, a useful review would separate three things: facts the manager knows directly, suggestions introduced by the model and uncertainties that still need investigation. That record is more informative than a generic declaration that a human reviewed the output. It shows whether review produced a judgment or simply added a signature.
More paperwork is not more judgment
Organisations can respond to AI risk by adding approval forms, disclosure labels and mandatory checkpoints. Those controls may help, but they can also become another layer of text generated without reflection. A completed template proves that fields were filled. It does not prove that anyone understood the situation.
A better checkpoint asks the responsible person to explain one consequential assumption in ordinary language. What evidence supports it? Which affected perspective is absent? What observation would change the decision? These questions make uncertainty visible and create a reason to gather information beyond the model's response.
The same principle applies to model outputs that include citations or extended reasoning. Supporting material can improve traceability, but the manager still has to decide whether the sources are relevant, whether the reasoning fits the organisation's circumstances and whether the proposed action treats people fairly. Accountability cannot be delegated to formatting.
There is also a role for deliberate friction. The fastest workflow is not always the one that preserves the most learning. For low-stakes, reversible tasks, automatic drafting may be entirely appropriate. For decisions affecting employment, workload, safety or access to opportunity, slowing down long enough to surface assumptions is part of competent management.
Feedback must reach the person who made the choice
Judgment develops through consequences as well as deliberation. Suppose the new schedule is implemented and an overlooked problem appears. If the manager receives only a summary designed to show completion, the opportunity to learn is weakened. Direct feedback from affected colleagues gives the decision-maker a concrete mismatch to understand and incorporate next time.
This turns AI-assisted judgment into a sequence: understand the situation, generate or compare options, choose, observe the consequences and revise the mental model. Generative AI can contribute at several points, but the learning loop breaks when the person responsible is insulated from context or feedback.
Our related article on Tina Huang's AI productivity feedback loop examines iteration as a way to improve outputs. Managerial judgment adds another requirement. The goal is not only a better document or faster result. It is a decision that someone can explain, defend and revise in light of what happens.
Testing the model requires longitudinal evidence
The Bath framework points toward empirical questions that future studies can test. Useful evidence would identify the managerial task, the participants' experience, the AI system and version, the comparison condition and the period over which learning is assessed. Immediate speed and output quality should be measured separately from retained ability to reason without assistance.
Researchers could also compare different accountability structures. One group might receive AI output with no explanation requirement. Another might have to document key assumptions before acting. A third might receive structured feedback after the decision. Measuring both task performance and later unaided judgment would help distinguish productivity gains from changes in capability.
Until that evidence exists, two broad conclusions remain unsupported. Routine AI use has not been shown to make all managers less capable, and more capable models have not removed the need for contextual judgment. The process model instead supplies a disciplined hypothesis: the way people use and answer for AI-assisted decisions may determine whether the technology erodes or exercises practical wisdom.
For organisations, the immediate lesson is concrete. Identify which part of a decision remains human, what facts must be checked outside the model and what explanation the responsible manager should be able to give before action. Generative AI can accelerate analysis and writing. Accountability keeps that acceleration connected to judgment.
