Companies Are Deleting the Judgment Layer. Then Renting It Back From a Chatbot.

Four business school academics, led by Dirk Lindebaum at the University of Bath, published a paper this month in the Academy of Management Review with a title only an academic could love: "A Process Model of Managerial Phronesis in the Age of Generative AI." Strip the Greek and it says something every operator already suspects. The more managers lean on ChatGPT and its cousins to do their thinking, the less capable they become of doing it themselves.

Phronesis is Aristotle's word for practical wisdom. Not book knowledge. The stuff you accumulate by making decisions, living with the consequences, reading a room, and getting it wrong enough times that you eventually get it right. Lindebaum and his coauthors argue that generative AI cannot manufacture this, because it has no experience of the world, no grasp of what a decision costs, and no sense of the social and emotional wiring of a workplace. It produces plausible text from patterns in existing text. That is a useful thing. It is not judgment.

The paper lays out two opposing processes. The first they call epistemic de-skilling: managers under time pressure treat AI as a shortcut, stop asking hard questions, stop seeking other perspectives, stop learning from real interactions, and gradually lose the capacity to think critically at all. The second they call epistemic up-skilling: the same tools, used to challenge assumptions and stress-test reasoning, actually sharpen judgment. The variable that decides which path you end up on is accountability. When managers know they will have to explain and defend a decision to real people, AI becomes a sparring partner instead of a substitute.

One honesty note before I go further, because this site does not do hype. This is a conceptual paper, a process model built from theory, not a randomized trial with a control group. It does not contain a number that proves managers are getting dumber. What it does is give the profession a precise vocabulary for a pattern that other research is already measuring. A Microsoft Research and Carnegie Mellon survey of 319 knowledge workers, presented at CHI in 2025, found that the more confidence people placed in AI's ability to do a task, the less critical thinking they reported doing themselves. The Bath team is explaining the mechanism. The empirical crowd is starting to count the bodies.

Here is where the paper gets uncomfortable, and where I part ways with the polite framing in the press release. The researchers describe de-skilling as a risk to be managed. In my mind it is closer to a strategy being executed.

Look at what corporate America has done with its management layer over the past two years. Gartner predicted in late 2024 that by 2026, one in five organizations would use AI to flatten their structure, eliminating more than half of their current middle management positions. That prediction is now the operating plan at a large share of the companies I talk to. "Unbossing" got its own vocabulary. Spans of control that used to be eight are now fifteen or twenty. The managers who survived inherited the work of the ones who did not, and the tool they were handed to absorb that load was a chatbot.

So walk through the Bath model with that picture in your head. The paper says de-skilling is most likely when managers face substantial time pressure and treat AI as a quick fix rather than engaging with the difficulty in front of them. Which population of managers is under more time pressure than one that just absorbed two colleagues' teams? Companies did not stumble into the conditions for epistemic de-skilling. They built them, deliberately, on a slide labeled "efficiency."

And then there is the other half of the equation, the one my industry is directly responsible for. Phronesis is cultivated through experience. That is the whole definition. It cannot be hired out of a bootcamp or prompted into existence. It lives in people who have run a P&L through a downturn, managed a layoff badly once and well the second time, and learned which customer complaint is a fire and which is noise. Those people are disproportionately over fifty. And those are exactly the people the hiring infrastructure filters out before a human ever reads the resume. The applicant tracking system, the recruiter screen, the "digital native" job description: all of it selects against the one asset the study says AI cannot replicate.

The result is a company that has thinned its judgment layer, loaded the survivors with enough work to guarantee they outsource their thinking, and closed the front door to the people who could rebuild the capacity. Then leadership reads a headline about AI eroding managerial judgment and calls it a technology problem.

The most useful sentence in the paper is this one: "It is that which Gen-AI cannot satisfactorily explain that managers must explain to themselves and others." That is the entire design principle. AI systems are notoriously bad at explaining why they produced a given output. In an organization where nobody has to justify anything, that opacity gets waved through. In an organization where a manager must stand in front of a client, a board, or a team and say "here is why we did this," the gap in the machine's reasoning becomes a forcing function for the human's.

This is the good news, and it is not small. The same tool that hollows out a manager in one shop sharpens one in another. The difference is not the model, the vendor, or the prompt library. The difference is whether the organization still expects people to own decisions.

Which brings me back to my industry, because accountability is also a function of who you put in the seat. Experienced professionals bring their own accountability with them. Someone who has spent twenty-five years being answerable for outcomes does not read an AI draft and hit send. They read it the way a veteran editor reads a rookie's copy: useful starting point, now let me tell you what it missed. That reflex is not a personality trait. It is phronesis, and it is the thing the study says you cannot download.

For employers, the move is not complicated, but it does require reversing two decisions that felt smart eighteen months ago. Stop measuring management productivity by how many direct reports a person can survive. That number is a de-skilling index, not an efficiency metric. And stop letting the screening infrastructure quietly reject the workers who already have the judgment you are now paying consultants to teach.

Design the roles so that explaining is part of the job. The paper's own recommendation is that organizations must deliberately structure roles, responsibilities, and workflows so that people keep developing the human competencies AI cannot match. Translate that out of academic: make people defend their calls out loud, in front of other people, on a regular cadence. Then hire the people who are already good at it.

For the experienced professional reading this, the message is simpler and more pointed. Your thirty years of practical wisdom is the scarcest input in this whole system, and it is the one input the machine cannot generate. Do not throw it away to save ten minutes. Use the tool the way the up-skilling half of the paper describes: to argue with, not to obey. The day you stop being able to explain why the AI is wrong is the day you become interchangeable with it.

Companies spent the last two years deleting the judgment layer to hit a margin target, and they are about to spend the next two years discovering that judgment does not regenerate on its own. The Bath paper hands them the diagnosis and, quietly, the cure. Accountability plus experience produces wisdom. Time pressure plus a chatbot produces the opposite. The people with the experience are already in the market. Most of them are over fifty, and most of them cannot get past your screen.

Sources: Lindebaum, Balasubramanian, Ashraf & Haack, "A Process Model of Managerial Phronesis in the Age of Generative AI," Academy of Management Review (2026); University of Bath announcement (Sept 4, 2026); Lee et al., "The Impact of Generative AI on Critical Thinking," CHI 2025 (Microsoft Research / Carnegie Mellon); Gartner, Top Predictions for IT Organizations and Users in 2025 and Beyond (Oct 2024).