Rebuilding the Apprenticeship

We argued in The Missing Rung that automating the first rung of the ladder quietly spends down a stock of judgement — the senior engineers we rely on were made by doing the early-career work we are now handing to agents. That was the diagnosis. This is the harder question: how do you rebuild the thing on purpose? It is really a question most organisations have never had to ask, because judgement used to arrive as a free by-product of the work. It doesn’t any more. And underneath the engineering specifics sits a much bigger problem — how an organisation reproduces judgement across generations at all.

Judgement was never free. The work was paying for it.

Judgement was never free — the work was paying for it

It is tempting to assume seniority simply accrues with time, but the learning science is blunt that it does not: beyond roughly the first two years, length of experience is only weakly related to actual performance. Expertise comes instead from deliberate practice — practice with a clear goal, real motivation, immediate feedback and repetition that refines — and from apprenticeship, where a novice gradually moves from the edges of real work toward full responsibility. Here is the thing nobody noticed: ordinary engineering work used to bundle all of that, invisibly. The boilerplate, the first draft, the well-specified ticket, the review you sat in — they delivered the feedback, the repetition and the modelling for free, smuggled inside the act of producing. So when AI removes that work, it is not removing typing. It is removing the hidden curriculum that production had been quietly paying for.

The target shifts: from producing to judging

What is the apprenticeship actually for? Its end-state, in the research, is that the learner internalises the master’s evaluative function — the ability to monitor their own work, notice their own errors, and judge quality without being told. That has always been the real prize. And it is exactly the capability the AI era now puts front and centre: generative AI shifts knowledge work from execution toward oversight — from producing an answer to verifying one, from doing the task to stewarding it. The work itself now rewards the evaluative skill the apprenticeship was always secretly building. So the target of a modern apprenticeship is explicit: teach people to judge — to review, to accept, to decide — using the AI’s output as the material they learn on.

This is not an argument for skipping the build

A necessary caveat, and stating it plainly makes the rest more credible, not less. This is not an argument for learning to review without learning to build. Judgement still emerges through production — you cannot reliably judge work you have never done. The difference is that judgement is now the explicit target rather than the accidental by-product. Juniors must still build; we are simply no longer able to assume that building, on its own, will quietly manufacture the judgement. So they produce and judge, with the judging made deliberate — not produce and hope.

The Designed Apprenticeship

If the apprenticeship will no longer happen by accident, it has to be designed. The Designed Apprenticeship is the deliberate version of the thing the work used to do for free. Like any architecture it has components — the seats a junior occupies, and the method by which a senior transfers judgement into them:

• The Reviewer Seat — put juniors in the position of judging AI output early: deciding whether a plausible change is actually right, and learning to feel when it is confidently wrong.
• The Decomposer Seat — let them break problems down and decide what to delegate to an agent and what to keep, which is where a great deal of senior judgement actually lives.
• The Acceptance Seat — give them real (supervised) ownership of what ships, because judgement forms under genuine stakes, not in a sandbox with no consequences.
• Cognitive Apprenticeship — the method that ties the seats together: make the expert’s hidden thinking visible, coach against it, scaffold the support, then fade it as the judgement takes hold.

The method: make the reasoning visible

The engine of all of this is older than AI and well evidenced: cognitive apprenticeship works by externalising the reasoning that experts normally keep tacit. A master models the thinking aloud, coaches the apprentice through attempts, scaffolds with hints and structure, and gradually fades that support as the apprentice internalises it. In practice this is what a good review always was — a senior pausing on a change that looked entirely reasonable and explaining why it made them uncomfortable. The lesson was never “here is the bug”; it was “here is why this looked correct and wasn’t”. That is the moment judgement transfers, and it is precisely the moment that disappears when the agent writes the change and no one narrates the call. Rebuilding the apprenticeship means staging those moments on purpose: reviews run as teaching, the senior’s reasoning said out loud, the junior in the seat where they have to make the judgement themselves.

The most valuable part of a review was rarely the approval or the correction — it was the explanation. A senior would pause on a change that looked entirely reasonable and explain why it made them uncomfortable. The apprentice was not learning what was wrong; they were learning how somebody experienced sees the problem. The systems disappeared; the judgement survived because it was transferred inside the work — and that is exactly the work we are now automating.

The AI complication

AI does not make this easy, and honesty matters more than a clean story. There is real evidence that leaning on the tool can erode the very thing we are trying to build: higher confidence in AI correlates with less critical thinking, and studies find people produce better outputs with AI yet show no greater learning or transfer — the work gets done, the judgement does not form. But the opposite also appears: novices frequently under-use or misuse AI and lock in poor strategies, because using AI well is itself an undeveloped skill. The truth is that whether AI-assisted work builds judgement faster or slower is genuinely under-measured — it is a risk to manage deliberately, not a settled law. What tips it one way or the other is scaffolding: AI left to do the thinking offloads it; AI used as something to interrogate and judge can sharpen it. The design is the difference.

Fund it as infrastructure

The last move is the one that makes it real: treat the apprenticeship as infrastructure with an owner, a budget and a measure, not as a hope. The economics support it — apprenticeship investment tends to return positively even during the training itself, so growing judgement is not pure cost deferred against future benefit. Measure the stock you are trying to refill: track whether your people are getting better at the calls — the accuracy of their accept/reject decisions on AI output, the rate at which they catch the confidently-wrong change. And resist the two easy escapes. “AI tutors will teach them” is not yet supported by the evidence. “We will just hire seniors” is the build-versus-buy fallacy at civilisational scale: you cannot buy your way out of a judgement shortage you have stopped manufacturing, because everyone is drawing on the same shrinking pool.

Who creates the next senior?

Almost everyone in this debate is asking how AI changes senior engineers. Far fewer are asking the question that decides the next decade: how future senior engineers get created at all. The Missing Rung named the risk — that we are spending down a reserve we have stopped refilling. The Designed Apprenticeship is the answer: rebuild, on purpose, the mechanism that the work used to provide for free. Make judgement the explicit target, make the senior’s reasoning visible, put juniors in the seats where they have to exercise it, and fund the whole thing as the infrastructure it has quietly always been. Because judgement was never free. The work was paying for it — and now you have to.