The trouble with training talent

A few years ago I spent real energy fighting to get the best young talent into my teams. I was confident in the deal on offer. I could give them an exciting mission, onboard them fast, and convert their energy and openness into disproportionate value while they built a career. This was a deal that was good for both sides, and meant I got to work with a disproportionate number of talented, energetic, interesting young people. I can’t make the deal work as well any more, and it’s made me think.

The junior role in most knowledge businesses was never really about the output. It was a translation layer. An experienced professional had the judgment, the context, and the relationships, but turning that into a working model, a clean analysis, or a new tool required hours of skilled-but-junior labour. The senior specified, the junior built, and the loop went round until the thing worked.

The handoff and interaction between senior, experienced professionals and junior, smart, apprentices is where juniors learned. It is also where they were economically useful. The two were the same activity, which is the point almost everyone misses when they talk about training budgets. Firms did not fund this type of learning. Learning was a free byproduct of low-value production. The apprenticeship was subsidised by the so-called ‘grunt’ work that senior professionals didn’t want to do.

Agentic AI has collapsed the handoff. The experienced professional who is willing to sit at a keyboard can now go directly from judgment to output, iterating faster and more accurately than the specify-build-review loop ever allowed, and capturing the upside personally. I have watched this shift my own hiring instincts. I am now (selfishly, and in the short-term) better off finding one rare senior person with deep domain experience who is willing to get hands on than hiring three bright graduates, and I do not think my instincts are unusual. I understand from people involved in graduate recruitment across several firms that intake numbers are going down, not as policy, just as a series of individually sensible decisions.

Why this wave is different

The obvious objection is that we have been here before. Spreadsheets destroyed the junior analyst’s manual modelling. CAD destroyed the draughtsman. Word processing destroyed the typing pool. Each time, the entry rung moved rather than disappeared, and a new generation learned on the new tools.

The difference this time is what happened to the curriculum. In every previous wave, the new grunt work was still grunt work: valuable enough that firms would pay juniors to do it, educational enough that doing it built judgment. The subsidy survived because low-value production survived. This time the production itself has gone to the machine. What remains for a junior to do is either too trivial to teach anything or too senior for them to do yet.

Which turns training from a free byproduct into a pure cost. A cost with a three-to-five year payback, no retention lock (the trained employee can leave the day they become valuable), and no offsetting production. Every finance director who looks at that line item will reach the same conclusion, and individually they will all be right. Collectively it is a catastrophe, because every firm is now free-riding on a talent pipeline that no firm is refilling. This is a well understood problem in economics in general, and it does not get solved by asking firms to be nicer (cf carbon emissions).

The opportunity we are mining

The rational firm response, the one I described in my own hiring instincts above, is a depletion play. The hands-on senior with twenty years of accumulated judgment is a stock, not a flow. That stock has maybe a ten-to-fifteen year half-life through retirement alone, and by my logic above nothing is flowing in behind it. We are mining a seam and calling it a talent strategy.

Industries that paused building people for a generation have run this experiment already. Nuclear did. Failure to keep building nuclear plant and submarine meant that the UK lost SQEP altogether in several domains. Aerospace did in several countries. Both lost tacit knowledge of how the plumbing actually works, the unwritten stuff that never made it into documentation, and both spent decades and enormous sums partially reconstructing it. The knowledge that matters most is exactly the knowledge that cannot be rehired from the market, because the market stopped producing it too.

There is a second-order effect as well. Frontier AI models are converging: trained on overlapping data, distilled from each other, increasingly similar in what they produce. In a world of converged models, competitive edge comes from what the models do not have, which is off-distribution / rare event human experience. The pool that generates that experience, people spending years close to real problems, is precisely the pool we have stopped funding. We are not just cutting a talent pipeline. We are cutting the pipeline for new ideas, at the exact moment the machines make everyone’s old ideas identical.

Where this lands

The honest answer is that this was never a firm-level problem, and the historical fixes were never firm-level either. Articled clerkships in law, pupillage at the bar, the actuarial examination system - every one of these exists because professions worked out, some of them centuries ago, that individual employers cannot be trusted to fund the pipeline, so the cost had to be mutualised and enforced by institutions. The apprenticeship was rebuilt as infrastructure.

I suspect that is where this ends up again. Professional bodies, regulators, or industry consortia will eventually rebuild structured entry paths that decouple learning from immediate productivity, because the alternative is running the seam dry and finding out what a knowledge industry looks like with no one left who knows how anything works. The uncomfortable part is the timing. Institutions move in decades and the depletion is happening in years.

There is one genuine counter-argument that I’ve heard. It is possible that AI compresses experience acquisition itself: that a junior who can run a hundred simulated deal cycles, or re-underwrite a decade of decisions against actual outcomes, builds judgment faster than the apprenticeship era ever allowed. If that is true, the pipeline problem is a transition problem rather than a structural one. I will take that argument seriously in a follow-up piece, because I think it is the only lever that attacks the problem rather than managing the decline. I am cynical though about the impact on innovative thinking in the medium to long term.

My prediction: within five years, the graduate intake numbers in knowledge-intensive industries will have plummeted, and the first serious institutional responses will come from the regulated professions, because they are the only ones with bodies capable of imposing mutual cost. I could be wrong about the timing and the response. I do not think I am wrong that the old deal has stopped working on the same terms, and that graduates in the next few years will find they are not in as strong a position as the graduates of the last 20 or so years.

As usual, I look forward to being told where I’m wrong.