Here's a trap that looks like a win on a spreadsheet. AI is genuinely good at the work we used to hand to junior people — the first-draft code, the research summary, the basic analysis, the routine document. So the obvious move, the one finance loves, is to use the tool instead of hiring the junior. Same output, lower cost, no onboarding. The maths is clean.
The maths is also lying to you, because it only counts this year.
Seniors are grown, not procured
You cannot buy a senior. You can buy a senior — poach one from someone else — but as an industry, as an economy, the supply of senior people exists only because somewhere a junior person spent years doing unglamorous work badly, then less badly, then well, and became the senior. There is no other path. Expertise is the residue of having done the work, including the boring and the wrong parts of it.
The junior tasks we're so keen to automate aren't just output. They're the apprenticeship. The analyst grinding through a tedious dataset is learning what clean data feels like and developing the instinct that something's off before they can say why. The junior developer fixing trivial bugs is building a map of the codebase in their head. That "low-value" work is where judgement is manufactured. You're not paying for the output. You're paying for what doing the output does to the person.
Take it away entirely and you get a cohort who can prompt a model to produce a passable first draft but have never built the underlying judgement to know when the draft is subtly, dangerously wrong. They skipped the years that grow the instinct. And nobody notices the gap for about five years — right up until your current seniors retire and there's nobody behind them, because you stopped making any.
The pipeline problem is everyone's, and no one's
This is what economists call a collective action problem, and it's worth naming because it explains why it'll happen by default. For any single company, cutting junior hiring is rational — let everyone else bear the cost of training people, then hire the finished article. But if everybody reasons that way, nobody trains anyone, and in a decade there's a shortage of mid-level and senior talent that no salary can conjure, because the people simply don't exist. Each firm optimised itself into a problem none of them can buy their way out of.
I'm not appealing to anyone's better nature here. I'm pointing out that the firm which keeps growing its own people, while competitors strip-mine the pipeline, ends up with the one thing that becomes genuinely scarce. That's not charity. That's the smart position, held with a longer time horizon than next quarter.
Redesign the role; don't delete it
The answer isn't to protect junior work by banning the tools and having bright twenty-three-year-olds do by hand what a model does in seconds. That's just expensive nostalgia, and they'd hate it. The answer is to redesign what junior means.
Build the role around judgement instead of rote production. If the model writes the first draft, the junior's job becomes critiquing it, finding where it's wrong, improving it — which, done with a good mentor, actually develops judgement faster than producing slop from scratch ever did. They get to see more cases, across more situations, sooner. Used deliberately, AI could be the best apprenticeship accelerator we've had.
Concretely:
- Protect the learning, not the task. Ask what a given junior activity was teaching. If a tool removes the activity, you have to deliberately replace the lesson — through review, through harder problems brought forward, through exposure — or you've cut the apprenticeship without noticing and kept the headline.
- Make seniors teach, and count it as work. If AI absorbs the rote, senior time frees up. Aim that freed time at mentoring rather than at simply shipping more. The transfer of judgement from senior to junior is now the scarce, valuable thing, so resource it like it matters and measure people on it.
- Give juniors real responsibility earlier, with a safety net. When the mechanical floor is handled, push them up the stack sooner — judgement calls, design decisions, talking to actual users — with the support to fail safely. Done right, you grow people faster than the old model did.
- Hire for trajectory, not just current output. A junior who's slower than the model today but learning fast is an appreciating asset. One who leans on the tool and never builds underlying skill is a liability you won't spot for years. Hire and develop for the curve.
The organisations that win the next decade won't be the ones that cut deepest and fastest on entry-level roles. They'll be the ones who understood that AI changes what junior people should do without changing the brute fact that you still have to grow your seniors from somewhere. Automate the junior tasks and skip the junior people, and you're eating your seed corn. It tastes like a saving right up until planting season, when you find you have nothing left to plant.
