The cheapest work to hand to an agent is the early-career work — the boilerplate, the first draft, the well-specified ticket. It is also, awkwardly, the work that made seniors. The judgement that lets someone hold the accountable core was not taught in a course; it was earned by doing those tasks, getting them wrong, and learning what “wrong” feels like before it ships. Automate the first rung of the ladder and you save money today against a cost that arrives years later: the people who can own the hard call.
How seniority is actually made
Senior judgement is tacit. You acquire it by repetition at the coalface — writing the unglamorous code, debugging your own mistakes, sitting in reviews where someone more experienced explains why a plausible change is quietly dangerous. The foundational work was never just output; it was the apprenticeship, bundled invisibly into the act of doing the tasks. That bundling is what agentic delivery breaks: the tasks can now be done without a junior ever doing them.
What the evidence says — suggestively, not conclusively
The early data points the worrying way. Stanford’s “Canaries in the Coal Mine” finds early-career workers (22–25) in the most AI-exposed occupations seeing a roughly 16% relative decline in employment, while experienced workers in the same fields hold steady. A controlled trial of an AI coding assistant found the largest productivity gains went to juniors and recent hires — the rung most easily automated is the one AI most readily replaces.
But intellectual honesty matters more than a scary number. The causation is genuinely contested: the Economic Innovation Group attributes the junior dip to the sharpest monetary-tightening cycle in decades rather than to AI, and other analyses find no discernible aggregate disruption yet. Crucially, no study has measured the thing that actually matters here — whether AI-assisted juniors build judgement faster or slower than those who learned the hard way. So this is a risk to manage deliberately, not a law to cite. The mechanism is plausible and the stakes are high; that is reason enough to act, and not enough to overclaim.
The hidden line on the P&L
Treat judgement as a stock, not a flow. For decades it replenished itself for free, as a byproduct of juniors doing tasks. Automate those tasks and the stock stops topping up — invisibly, because nothing breaks this quarter. The bill comes due the year you go to staff the accountable core and find no one ready to hold it. That is the line the AI-delivery P&L does not show: the cheapest thing to automate and the most expensive thing to lose can be the same thing.
The people we trust with the hard calls today learned them by doing the work we are now handing to agents. Stop rebuilding that apprenticeship on purpose and we are drawing down a reserve we have stopped refilling — and it will not show up on a dashboard until the day it is already gone.
Rebuilding the rung on purpose
If the apprenticeship will no longer happen as a side-effect of foundational work, it has to be designed in. The shift is to teach the roles, not the typing:
- Put juniors in the decomposer and reviewer seats early — deciding what to delegate and judging what comes back — rather than only the doer seat that agents now fill.
- Give them real accountability, with a senior beside them, not a sandbox with no consequences — judgement forms under genuine stakes.
- Make reviews teaching, not gatekeeping: the explanation of why a plausible agent output is wrong is the lesson the foundational work used to provide.
- Treat apprenticeship as infrastructure with a budget and an owner, not a hope.
The point is not to slow agents down to keep juniors busy — that would be theatre. It is to recognise that the pipeline which produced senior judgement ran on the very work we are automating, and to rebuild it deliberately before the stock runs low.