Some of the margin consulting firms are gaining from AI is borrowed from the future. Junior work was never only a source of profit. It was how a firm produced the people who would one day lead it. Remove that work, or let the tools do it unexamined, and today's numbers improve while the judgement the firm will need tomorrow fails to form.

Junior work was the training

Nobody designed the pyramid as a school, but it worked as one. A first-year analyst in a due diligence team rebuilt the target's numbers and found the ones that would not reconcile. A junior actuary ran valuations until the assumptions felt familiar. A graduate on a strategy project wrote the first draft, had it torn apart and learned why. A new engineer on an implementation traced the defect nobody else wanted to touch.

That work was billable, so clients paid for it. It was also how people learned what a wrong answer looks like. Ten years of it produced a manager who could sense when a number was off. Fifteen produced a partner a client would trust.

AI separates the output from the learning

AI now does much of that work well. A junior consultant can produce analysis and drafting that reads like a manager's. The output improves. The judgement behind it need not.

Recent research points the same way. In a three-month study of patent drafting, AI assistance helped experienced lawyers when they later worked without it. Junior lawyers showed no average gain, and their results became more erratic. Separate experiments found that people given a confident AI answer tend to defer to it, and grow more confident even when it is wrong.

That second finding matters for review. A manager checking a polished draft starts from an answer that already looks finished, and is less likely to rebuild it from scratch. The old safeguard, an experienced eye on junior work, weakens just as junior work becomes harder to read.

This is not inevitable. Where tools are used to coach rather than replace, less experienced staff can learn faster than before. But that happens by design, and most firms have not designed for it.

Healthy numbers can hide a weakening firm

Every measure firms usually watch still looks healthy: utilisation, margin, output per head, client feedback on deliverables. The weakness shows up elsewhere: a manager who cannot defend a number in front of a client, an error that gets past review, a promotion round with nobody ready. A large firm has enough experienced people to absorb this for years. A small one does not, and its clients notice within a year or two. Left alone, it ends in senior people who are hard to replace and a succession plan that depends on a few of them. These are the same people a firm needs to tell a material finding from a merely interesting one, the skill on which "If the client can write the brief, you are already late" rests.

Firms that use AI savings to cut junior intake make this worse. Fewer juniors, each learning less, will produce far fewer people fit to lead.

The firms that keep training will pay for everyone else

If most firms cut back on training, those that keep investing become suppliers to the rest. Their best people, trained at their expense, will be hired away by rivals that saved the cost. The firm that trains well may be the one losing value, unless it gives its people reasons to stay. For an investor, that turns a firm's training record into a question about retention, earn-outs and key-person risk, not only about culture.

What to test, and what to build

For an investor, a margin improvement built on thinner junior ranks should not be treated as permanent. We would look at how the grade mix is changing, whether promotion is slowing, and how much of the firm's client trust rests on a handful of senior people. We would also test the team directly. Take a strong piece of recent work, introduce a fact that should change the answer, and see whether they notice and can explain what follows. The tools are allowed. The question is whether the firm as a whole can catch its own mistakes.

For leaders, the answer is not to rebuild the old pyramid. The firm of the next decade may need fewer partners, whose judgement was formed by challenging AI output rather than by years of producing work by hand. The question is what that apprenticeship looks like, and it has to be designed on purpose. Some work should be done, or at least critiqued, without the tools. Juniors should be asked both to defend an answer and to find what is wrong with one. Promotion should reward people who catch errors, not only those who produce polished work.