The School Hidden Inside the Job

AI is not coming for entry-level cognitive work. It is coming for the only way anyone has ever learned to become senior at it.

Every serious piece of research on the future of cognitive labour keeps circling the same shape and refusing to see it. Microsoft’s Work Trend Index says 83% of leaders expect AI to let employees do more complex work earlier in their careers. KPMG reports that adaptability now outranks technical skill for entry-level roles. Deloitte describes the human future in three words that reappear in every enterprise AI report written this year: judgement, investigation, intervention. The consensus view is that routine cognitive work compresses into agents and the humans above it rise into more valuable territory.

This is true in the narrow sense and catastrophic in the wide one.

The wide sense is this. The routine cognitive tasks that are about to be automated were never really just routine tasks. They were the curriculum. The first-pass research memo was how the junior analyst learned which senior partners cared about which numbers and why. The scaffolded code was how the developer built the sense that lets her, six years in, smell a memory leak through a screen. Tier-one support was how the product manager learned what customers actually meant when they used the word “slow”. Document review was how the lawyer learned to feel the architecture of a case through her hands. The drafts, the corrections, the boring repetitions, the slightly wrong first attempts at something important: this is not noise around the apprenticeship. It is the apprenticeship. Cognitive labour teaches itself by doing, badly, at a survivable scale, under someone who has already done it badly and survived.

The trades understand this. A carpenter’s apprentice is not training for the moment when she will supervise a table saw. She is building the proprioceptive memory of what wood does under pressure, and she can only do it by being wrong in small ways until her body knows better. This is why the phrase “learning curve” originally referred to manual work. The body kept the ledger.

Cognitive labour has no body. Its ledger is kept in the drafts.

Remove the drafts and something very specific happens. The graduate who would have written five bad memos before writing one good one now watches an agent produce twenty decent memos in an afternoon and her job is to check them. She is told this is more dignified work. In a sense it is. But she has no mental model of the memo. She has never felt a paragraph resist her. She has never been embarrassed in front of a senior partner by a miscounted basis point. The embarrassment was the education. The slow hot rising of the face at one’s own error is how a human being installs a standard. She is being asked to audit the thinking of a system that has done the thing she has never had to do, and her judgement is expected to track the quality of work she has never produced.

This is not a transitional problem. It is a generational one. And it is specifically a cognitive-labour problem, because cognitive labour has no subsidiary apprenticeship hiding inside it the way the trades do. There is no equivalent of the carpentry student’s three years of sanding. You do not get to the senior partner’s office through some separate formative layer underneath the drafting work. The drafting work is the formative layer. Take it out and there is nothing beneath.

The firms moving fastest into agents will hit this wall at different speeds depending on how textual and deterministic their work is. Law firms will hit it first, because discovery, precedent search, and first-draft memo writing are exactly what current-generation agents handle well and exactly what the early career was built around. Consulting will follow. Investment banking will take longer than you would expect because the relationship layer has a long half-life, but only until the junior analyst pool thins out. Software will hit it in a strangely bifurcated way: today’s senior engineers will be fine for another decade, then suddenly be the last ones who know how to debug from first principles, because nobody under thirty will have been forced to do it the hard way. Medicine and academic research will resist longest because their practice still has embodied components and the stakes of error are too legible to outsource quickly. None of them are safe from the underlying issue. The issue is structural, and it is already legible in the enterprise reports. Deloitte’s Gartner-sourced prediction that at least 15% of daily work decisions will be made by digital colleagues by 2028 is usually read as a productivity number. It should be read as a pedagogy number. Every decision made by an agent is a decision a junior human did not make, did not get wrong, did not learn from.

The optimistic counter arrives on cue. New ladders will appear. Agent operations lead. Evaluation engineer. Workflow architect. AI risk owner. Human-in-the-loop auditor. Every enterprise AI essay ends in this place, with a list of plausible-sounding new jobs. These are lateral roles for people who already have craft. They are not entry. You cannot evaluate an agent’s legal reasoning without having reasoned about law yourself. You cannot audit a clinical recommendation without clinical judgement. You cannot design a workflow until you have inhabited one for long enough to know where its informal exceptions actually live. Every one of these new roles presupposes a decade of the exact work being automated. If the decade stops happening, the roles become unfillable by anyone under thirty-five roughly a decade from now, which is to say, next Tuesday in organisational time.

Nobody is seriously planning for this. Some firms have begun to talk about what they call learning pathways, which is consultant language for we know this is a problem and we have no solution. A few are experimenting with deliberate friction, asking graduates to write the document by hand before seeing the agent’s version. A few more are trying reconstructed apprenticeship, slow structured practice against a standard the market no longer pays the firm to produce. These are honest efforts. Most of them will be quietly killed by the next budget cycle. A two-tier training system running underneath a one-tier production system is only sustainable if the firm is willing to pay a permanent tax for deliberately slower work, and public markets do not reward that tax.

There are harder counters. One is that the work itself will become easier, so less training is needed. This is plausible for the tasks and false for the judgement. Judgement about when an agent is wrong, when a client’s stated need is not their real need, when the model has confabulated its citations, when the decision the system recommends is legally correct but wrong for this client: this kind of judgement cannot be read from a manual. It is acquired through the slow friction of having been wrong oneself. A second counter is that simulation will replace experience, that synthetic practice environments will let graduates accumulate something like the ten thousand hours without the hours. Maybe. Simulators work for pilots because the cockpit is a closed physical system with legible feedback. A simulator for cognitive judgement inside a messy firm has to simulate not just the work but the politics, the client, the senior partner’s mood, the quarterly pressure, and the specific moment when a junior realises she is about to be humiliated and chooses to push through. Nobody is close to building that, and the people claiming to be are selling something.

The deepest counter is that perhaps cognitive labour was never really trained this way. Perhaps the apprenticeship story is romantic and what was actually happening was survivorship bias plus institutional cover for cheap labour. There is something to this. Much of traditional white-collar apprenticeship was exploitative, and most of what juniors produced was indeed discarded. But exploitation and pedagogy were running on the same track. You cannot burn the exploitation without also burning the pedagogy unless you have built a replacement, and the replacements being proposed are uniformly worse than what they replace. A formal training programme is not the same as the slow accumulation of having been wrong in real conditions with real consequences. Watching the agent be right is not the same as having been wrong yourself.

Imagine the law firm that gets this right. It decides, in 2028, that the apprenticeship is a strategic asset rather than a cost. It pays for graduates to do the document review by hand while the agent does it twenty times faster. It absorbs the margin hit as an investment in having senior partners in 2040. Now imagine the firm next door that does not. For five years their margins look better. For ten years their margins look better. By 2041 the first firm has a generation of senior partners who know what they are doing. The second firm has a generation of forty-year-olds who have spent their careers auditing machines and are now being asked to exercise judgement in situations the machine has never seen, and they do not know how. Which firm gets there first? Which shareholders tolerate the first firm long enough to find out?

This is what the enterprise AI reports are dancing around when they use the phrase “talent pipeline”. The pipeline was never a pipeline. It was a pedagogy. The pedagogy was built into the tasks. The tasks are being automated. No replacement pedagogy exists at anything like the scale required. The most honest version of current corporate AI strategy, written in the voice of its own implications, would read: we are harvesting the last generation of seniors trained under the old conditions and we have not yet decided what to do when they retire.

There is one small mercy in this picture. The problem is visible. It is not emerging from some hidden corner of the economy, it is staring out of every enterprise adoption report written in 2025 and 2026. Anyone reading Microsoft’s Work Trend Index carefully can see the fracture. Anyone reading Deloitte on judgement, investigation, intervention can feel the absence underneath those three words. The reports are not hiding it. They are simply not yet willing to name what it means, because naming it would require firms to do the one thing quarterly capitalism cannot easily do, which is pay now for something that matters in fifteen years.

The generation currently twenty-two will find this out first. They will not find it out from essays. They will find it out from the quiet moment, some day in 2032, when a senior partner asks them to make a call on a matter the agent cannot handle, and they realise they have never actually done any of the work the judgement is supposed to rest on, and the partner realises it at the same moment, and neither of them has a name for what has been missing for years.

The name for it is the school inside the job. We are about to close it. And nothing is yet ready to take its place.