The University of Bath has highlighted research asking what happens when managers hand too much thinking to generative AI. Its September 4 announcement describes two possible directions: dependence that weakens judgment, or deliberate use that stimulates reflection. The distinction puts the design of work, rather than access to a tool, at the centre of the question.
The university identifies the paper as A Process Model of Managerial Phronesis in the Age of Generative AI, published in the Academy of Management Review. Bath's account presents a conceptual model of how judgment may develop or deteriorate. It does not report a measured rate of skill loss or establish that a particular model caused harm in a controlled workplace trial.
This first look draws on the university's public account of its researchers' work. Newsroom could not access the full journal text. The interpretation and hypothetical workplace example below are our own, not experimental findings from the paper.
A completed answer can hide an unfinished decision
Imagine a manager preparing to explain a change in a team's schedule. A generated draft arrives quickly. It reads smoothly, sets out reasons and anticipates objections. Yet the manager may still not know which employees face the greatest disruption, which earlier commitments matter or whether the proposed explanation matches what actually happened.
The hypothetical draft has completed a writing task. The decision remains unfinished. That gap is the practical concern raised by the new research: a polished output can make work appear ready before the person responsible has examined its context.
The background concept is phronesis , often translated as practical wisdom. Its relevance here is judgment about action in a particular situation. An organisation does not only need a plausible general answer. It needs someone able to explain why that answer fits these people, these obligations and these circumstances.
A model's answer can therefore be evaluated along at least two separate dimensions. Is the content accurate enough for its intended use? And has the human decision-maker done the work necessary to take responsibility for using it? Passing the first test does not automatically answer the second.
Accountability needs a visible place in the workflow
Bath's announcement says the researchers associate time pressure and uncritical reliance with a risk of knowledge-related de-skilling. It also describes a possible constructive path when people must explain their reasoning and use AI to challenge assumptions. These are proposed mechanisms, not a guarantee that adding an approval step will preserve expertise.
Our inference is that the important design question is where explanation happens. If it is requested only after a decision fails, it becomes a reconstruction. If it is required before action, it can expose a missing fact while there is still time to find it.
In the scheduling example, a useful record might distinguish what the manager knows directly, what the model suggested and what remains uncertain. That separation is more revealing than attaching a generic statement that a human reviewed the output. It identifies whether the review changed anything and why.
There is a risk on the other side too. An elaborate review template can become another document generated without reflection. More paperwork is not evidence of more judgment. A practical check would ask the person to explain a consequential assumption in ordinary language and identify the observation that could change their mind.
Feedback should reach the person who made the choice
Suppose the schedule is implemented and an overlooked problem appears. Who hears about it? If the manager receives only a summary prepared to show completion, the opportunity to learn may be lost. If affected colleagues can explain the mismatch directly, the decision-maker has a more concrete basis for revising the approach.
This is a proposed organisational implication, not a tested intervention reported here. It follows from treating judgment as something exercised over a sequence: understanding a situation, choosing, hearing consequences and reconsidering. An AI-assisted process can be examined at every point in that sequence.
The Newsroom's recent coverage of Tina Huang's AI productivity feedback loop offers a related discussion of iteration. The distinction in this article is responsibility. Improving an output and learning to justify a decision may overlap, but they are not the same outcome.
What would make the claim stronger
Useful follow-up evidence would describe the work being performed, the people involved, the AI system and version, the comparison condition and the time over which learning was assessed. It would also distinguish immediate speed from retained ability to reason without assistance.
Those questions prevent two equally premature conclusions. One is that routine AI use necessarily makes managers less capable. The other is that a more capable model necessarily removes the need for contextual judgment. Neither conclusion is established by the university announcement.
For readers, the immediate value is a sharper way to assess a workflow: identify the decision that remains human, the facts that must be checked and the explanation the responsible person should be able to give. The new research supplies a question worth testing. It does not yet supply a universal measure of how much managerial wisdom an AI tool adds or takes away.
