Take a graduate intake from ten years ago and follow it forward. A couple are still with the firm, one on partner track. Most left inside four years, which everybody treats as leakage. Look at where they went. Several are senior in industry now, and a handful sit on the other side of the table deciding which firms get appointed.
Some of them have already hired you.
The graduate programme is a client-development pipeline, and it has been generating revenue for decades while the accounts treat it as a cost.
The research bears this out. Studies of corporate alumni find that well-managed departures reliably convert former employees into customers, referrers and ambassadors. The consulting industry's most celebrated alumni network has worked as a client-generation engine for decades. Every partner in a large firm can confirm the mechanism from their own contact list.
Which makes the current efficiency case badly mispriced. A firm that hollows out its intake to capture an AI productivity gain is shrinking its own future client base, on a delay long enough that nobody will connect the two events. For a chief financial officer that may be the strongest argument available, and it will never appear on a dashboard.
There is a second cost, and it compounds with the first. The entry-level tasks AI most readily absorbs are precisely the tasks through which juniors built pattern recognition and domain intuition: the summarising, the first drafting, the formatting, the document review. Lave and Wenger called it legitimate peripheral participation: novices become experts by doing real but low-stakes parts of real work, watched by people who know the difference between good and not good. Nobody learns judgment from a curriculum. They learn what good looks like by repeatedly producing work that is not yet good.
Whether that matters depends on something no dashboard tracks.
In the 1840s, ready-made paint in metal tubes ended the apprentice's years of grinding pigment by hand. The craft did not collapse. Apprentices started painting sooner, at a higher level of abstraction, and art historians credit the tube with helping to make Impressionism possible. That is the optimistic case, and it is often correct.
The power loom is the other case. Skilled hand weavers were bypassed, and their accumulated expertise was stranded. The difference between the two was never the technology. It was whether anybody redesigned the apprenticeship on purpose.
The early data is more consistent with the loom than the tube, though I would not overstate it. Researchers at Stanford, using payroll records, found workers aged 22 to 25 in the most AI-exposed occupations have seen a relative employment decline of around 16 percent, while older workers in the same occupations held steady. The finding is contested, as any early finding should be: under the authors' most stringent controls it becomes clearly significant only from 2024, and they warn against reading AI as the sole cause. Set it beside the reported graduate intake cuts at the large accounting firms and the direction of travel is not in dispute. Only the speed is.
The moment that stays with me came from a conversation rather than a dataset. At a conference last year, a vice president at one of the Big Four told me plainly that he fears for his firm's business model within five years if they stop taking in graduates and training them properly. What struck me was not the fear but how hard it is to act on. Nothing is on fire. Every individual reduction looks sensible. The damage will surface on somebody else's watch. That is what this looks like from the inside: a series of reasonable decisions rather than a crisis.
I run a consultancy whose service is judgment-preserving AI adoption, so I am professionally disposed to find this problem everywhere. Weigh that. I have watched too many graduate programmes quietly shrink this year to conclude I am only seeing what I am paid to see.
Something to try this week. Find the most recent business case that proposed reducing an entry-level function and run it back through three questions. What capability was being built in the people who did that work? Where will that capability come from now? And what was that intake earning us commercially, in future clients and future leaders, that nobody counted?
If the case cannot answer all three, the saving is not coming out of cost. It is coming out of capital.