Transmission // 09
The ladder lost its bottom rung
2 min
No mass unemployment, no collapse — but hiring for 22-to-25-year-olds in exposed fields has slowed, and entry-level roles now ask for senior judgement.
The honest summary of the 2026 labour data is boring and worth reading anyway: there is no systematic increase in unemployment in AI-exposed occupations. The apocalypse did not arrive on schedule. What did arrive is narrower and harder to fix.
Hiring has slowed specifically for workers aged 22 to 25 in exposed fields. And the roles that remain have mutated: PwC’s 2026 AI Jobs Barometer finds AI-exposed entry-level positions are seven times more likely to demand traditionally senior skills — judgement, leadership — with those roles up 35% since 2019 while other entry-level roles fell 10%.
What employers actually did
Not “replaced juniors with AI”. Something more specific:
- 38% shifted basic data processing off entry-level workers and onto models.
- 31% raised experience requirements for entry-level jobs as a direct result.
- More than a third of entry-level postings now require AI skills — nearly triple the share from autumn 2025.
Read those three together. The tasks that used to be how you learned the job got automated, and the requirement to already know the job stayed. That is the whole mechanism, and it has nothing to do with headcount.
The apprenticeship problem
Every skilled profession has run on the same quiet subsidy: juniors do low-value work badly, seniors correct it, and the correction is the training. It was never efficient. It was never supposed to be — the inefficiency was the tuition.
Automate exactly that tier and the ledger looks better this quarter and the pipeline stops. In five years the industry discovers it has no mid-level engineers and no idea where they were supposed to come from, because the answer was “from the boring work we deleted”.
For anyone hiring
The trade is legible if you are willing to name it. You can have the productivity now or the seniors later. Taking both requires deliberately spending on training that no longer happens by accident: pair juniors on the work the model does, make them review its output rather than compete with it, and treat “can tell when the machine is wrong” as the actual entry-level skill.
That skill is learnable. It is just no longer free.