The bench can be the same size and still form nobody.

The argument about AI and early-career work almost always arrives as a headcount. Graduate intake down a percentage. Entry-level postings falling. A hiring freeze in the function that used to run an intake every autumn. It is the natural evidence to reach for, because headcount is the one number about junior staff that every institution already computes, and it arrives in a board pack without anyone having to commission anything.

It is also weak evidence, and the weakness shows quickly under context. The most-quoted British figure is Adzuna’s finding that entry-level postings fell about 32 per cent. Over the identical window, total UK vacancies fell 36.1 per cent, on the ONS AP2Y series, which anyone can pull. Entry-level postings fell less than the market they sit in. Read against its own denominator, the number describes a hiring recession that entry-level weathered slightly better than average.

The most-cited American figure carries a different problem. “Canaries in the Coal Mine”, the Stanford Digital Economy Lab paper on ADP payroll records that reported a roughly 16 per cent relative employment decline for 22-to-25-year-olds in the most AI-exposed occupations, remains an unpublished working paper. Its own Appendix L reports the original within-firm specification, applied to current data, returning p = 0.26. Several published critiques are on the record: timing artefacts, the choice of employment rate over unemployment rate, deterioration that predates ChatGPT, and remote work absorbing the coefficient.

There is also a counter-literature, and a board that has read one side of this would be surprised by it. The Yale Budget Lab’s synthetic difference-in-differences is the strongest null in the set. BIS Working Paper 1325, Humlum and Vestergaard bounding the effect at plus or minus two per cent, and the California Policy Lab’s administrative data run in the same direction. The Economic Innovation Group found unemployment rising faster in the occupations least exposed to AI.

The headcount question is live, and it will stay live for years, which is thin ground for a board that has to decide something this year. There is also a reason it may never settle into anything usable, and it sits in the theory.

Enrique Ide’s model of the intergenerational transmission of knowledge, circulated as arXiv 2507.16078 and CEPR Discussion Paper DP20940, shows entry-level automation reducing growth and welfare even where entry-level employment holds perfectly steady. The mechanism runs through the mentor pool. Automate the tasks by which a novice used to sit beside the most productive experts, and the novice is reallocated: same job title, same seat count, further from the practice that was worth absorbing. Best-practice diffusion slows. Ide names investment banking and urological surgery as the illustrative trades.

That is the sentence we would take into a meeting. The bench can be the same size and still form nobody. The headcount will show none of it, because the headcount is holding, which is why arguing from headcount is a weak move even in the cases where its numbers turn out to be right.

The content question has better-designed evidence behind it, because it can be tested by taking the tool away. Bastani and colleagues, in PNAS in 2025 (122(26):e2422633122), put about a thousand high-school mathematics students through a randomised design with a genuine withdrawal arm. With access, a standard chatbot tutor lifted performance by 48 per cent, and a version built with learning safeguards lifted it by 127 per cent. Then access was removed. The unguarded group scored 17 per cent worse than students who had never had the tool at all. The safeguarded group did not.

Same model, same task, same students. What separated the group left worse off from the group left intact was how the tool had been built into the work. That turns the subject into something a board can specify: a deployment design, with conditions, and someone who owns them.

The professional version of the same test exists. Budzyń and colleagues, in The Lancet Gastroenterology and Hepatology on 12 August 2025, examined adenoma detection on standard colonoscopies, the ones performed without AI assistance, across four Polish centres before and after AI was introduced in those centres. Detection fell from 28.4 per cent to 22.4 per cent: 6.0 percentage points, with a 95 per cent confidence interval of minus 10.5 to minus 1.6. The study is retrospective and observational, covers a three-month window, and has no concurrent control.

A larger trial points the other way on its own question. Dominitz and colleagues, in Gastroenterology in 2026, found detection rising where AI was available, and their own conclusion states that endoscopist deskilling “remains undetermined”. The two studies measure different outcome variables. Dominitz measures detection across all colonoscopies once the tool was there, which is the direct effect of having it. Budzyń measures detection on the colonoscopies done without it, which is the carry-over. Both belong in the pack.

None of this settles into a number anyone can put on a register this quarter. It moves the question, which is enough for now. Headcount arrives in the pack whether or not it was asked for. The content question has to be asked, and it costs very little: take one entry-level role, take a week of its actual work, and sort the tasks by whether the person doing them found out afterwards which parts were wrong, and why.

A junior who drafts a credit memo and learns within the week which lines were wrong is being formed by the work. A junior whose drafts go into the system and come back approved is producing it. That sort can be run this month, from records the institution already keeps, and it moves independently of headcount, which is the reason to run it. It can also come back clean. If the week’s work is coming back marked, that is good news, and it is worth having on the record while it is still true.