Chapter 20
The Missing On-Ramp
The Utopia of the Agentic Enterprise, Part IV. Doing Less
Some of the work that stops or goes to an agent was also how people learned the job. Drafting routine documents was how juniors found out what a document can get wrong, though nobody paid them for that part. Take the drafting away and the documents still get written. The finding out has to be put somewhere on purpose, or it stops with the drafting.
A cost experiment produces results within a quarter. What it does to talent takes years to become visible. So when AI is modelled as a way to reduce headcount costs while the pipeline that builds future expertise is hollowed out, no single function can resolve the conflict. Finance can’t see talent effects on the balance sheet, because there is no line for them. HR can, but has no authority over pay or pricing.
The career architecture of knowledge professions assumes a pipeline. Junior roles are on-ramps where novices learn by doing the ground-level work.
When AI handles more of the drafting and first-pass modelling that juniors learn on, the on-ramp narrows. The seniors’ judgement was built on years of those cases, and the next generation can’t build theirs if machines handle the first years.
Bainbridge’s “Ironies of Automation” describes this problem exactly.1 The designer automates whatever can be automated and leaves the person to monitor it and take over when it fails. But skill is kept by practice, and monitoring isn’t practice, so the person is least able to take over at the moment the design needs them to. One of her ironies is that the designer, having decided the human is unreliable, still relies on the human for everything the design couldn’t anticipate. An AI deployment can still be designed as if the paper had never been written.
Of course, AI can also speed up expertise when the conditions are right. Seeing good work earlier, and having to explain why their own first attempt differs from it, can build a junior’s pattern recognition faster than waiting years for enough cases to go wrong.2
So whether an organisation gets deskilling or faster expertise depends on whether anyone protects unassisted practice, and checks whether expertise is developing or being handed to the tool. Where that is nobody’s job, the default wins. Without any unassisted work, a junior may never learn to recognise what they don’t know.
Reskilling can fix a skill. Expertise still comes from apprenticeship. In practice that means the junior’s assessment first, then the AI’s, then a senior practitioner on the gap. It means practice time protected in the schedule, and milestones that test whether the junior can spot AI output that is subtly wrong in their domain.
An organisation can employ experienced professionals today while running an operating model that won’t produce their replacements. Only the trend shows the gap, and nobody is asked to track it.
Business cases treat routine work as residue. Some of it is. But some of it is training and resilience that just happens to look repetitive from the floor above.
When a model takes over the routine work, weak signals go first. A delay too small to trip a metric got noticed because people ran into it repeatedly and talked about it, which the model doesn’t do.
Learning by proximity goes next, and nobody notices, because it was never a task. The junior no longer sees which numbers a senior distrusts. If everyone works alone with a generated first draft, fewer people get to watch each other think.
None of this is an argument for keeping dead reporting or duplicate entry. Ask whether the routine still teaches calibration or keeps a manual fallback alive. Where it does neither, automate it or delete it, and deleting is cheaper. Where it does, rebuild the capability somewhere else before removing the work: review sessions that show reasoning, and deliberate exposure to cases ordinary enough to teach without being catastrophic.
Whoever makes these choices is unlikely to be around to pay for them. Budget cycles run annually and pipeline erosion takes several years to become visible, so AI-assisted juniors look like capacity nobody had to budget for. The safe move for any one manager is to take the throughput gain and leave the expertise question to whoever inherits the team, since the manager is reviewed on four quarters and none of them will contain the cost.
Organisations that want a different outcome have to put the pipeline question on a ledger the manager actually looks at. That means a staffing assessment that grades how the juniors develop alongside throughput, and a review cadence that follows managers into their next role, to credit or debit the trajectory they left behind.
Checklist
- How many juniors did you hire last year for this kind of work, compared with three years ago?
- Before automating or stopping a routine, what else does it do: skill, calibration, relationships, recovery, weak signals, a manual fallback, spare capacity? If none, automate or delete it. If any, rebuild that capability somewhere else first.
- Do juniors make their own assessment before seeing the AI’s, then talk through the gap with a senior?
- Is protected practice time in the budget as a deliverable, or does it only count as lost throughput?
- Does a manager’s review grade junior development alongside throughput?