Insights
AI Automation and the Death of Workplace Apprenticeship
I have been thinking about a question that has no obvious owner inside any organisation I work with: who is responsible for producing the next generation of experts? Not hiring them. Producing them.
For most of industrial history the answer was embedded in the work itself. Juniors were given the small, unglamorous, correctable tasks — the routine fix, the first draft of the memo, the reconciliation nobody enjoyed. Those tasks were economically marginal. That was never their real function. Their function was to let someone struggle, get it slightly wrong, and be corrected by a person who had struggled before them. Expertise was a by-product of inefficiency.
We are now removing that inefficiency at speed, and I do not think we have noticed what we are removing with it.
The Disappearing Entry-Level Gap
Stanford’s Digital Economy Lab, working with ADP payroll data covering millions of American workers through mid-2026, found that employment of 22-to-25-year-olds in AI-exposed occupations now sits roughly 19% below where it would be had it tracked their less-exposed peers. Experienced workers in the same occupations show no comparable gap. The divergence is concentrated where AI automates rather than augments.
That is usually read as a displacement story. I think it is a transmission story, and the more interesting evidence is qualitative. A study published this year of software engineers — juniors at the threshold of the profession and the seniors who once trained them — describes a pattern the authors call absorption. Entry-level work does not disappear into a queue waiting for a junior. It is redirected into a senior-plus-AI workflow, where it is completed faster and better. Nobody makes a decision to stop training anyone. The training simply has nowhere to happen.
The researchers also describe a perceptual asymmetry. Seniors evaluate juniors on output, which now looks fine. Juniors experience the loss privately, as an absence of something they cannot name because they never had it. Neither side is positioned to see the problem, so neither side raises it.
An organisation can lose its apprenticeship system without a single meeting acknowledging that it happened.
The Hidden Cost of Efficiency
What makes this hard is that no individual decision in the chain is wrong. Giving the routine task to the AI is correct this quarter. It may be correct every quarter for four years. The cost arrives in year five, in a cohort of people who can supervise output they were never able to produce — able to accept or reject an answer, but not to sense when the question was wrong. Judgment, as far as we can tell, is not transferable by explanation. It is residue left behind by having been responsible for something.
I do not have a framework for this yet. I have a suspicion about where it starts.
Organisations measure the productivity of work. Almost none measure the developmental yield of work — whether a given task, done by a given person, left behind a capability that did not exist before. That number has never needed to exist, because apprenticeship was free and automatic, a side-effect of how work was distributed. It is now neither free nor automatic. It has become a thing that must be paid for deliberately, in slowness, in tasks assigned to the person who will be worse at them, in seniors spending hours correcting work an AI would not have got wrong.
Replenishing the Stock of Expertise
That is a real cost, and it should be named as one rather than smuggled into a development budget. A firm that declines to pay it will look more efficient than its competitors for several years. Then it will discover it has been consuming a stock of expertise it stopped replenishing, and that this particular stock cannot be bought back at any price, because everyone else stopped replenishing theirs at the same time.
Execution is becoming abundant. Judgment is still made the slow way — by people who were allowed to be responsible before they were ready.