Roles change from the bottom up
AI does not remove roles. It removes the bottom of them: the first draft, the first pass, the reconciliation nobody enjoyed. Those were the tasks people learned the job on. A company that automates its junior work without rebuilding the training path stops producing seniors, and finds out about three years later.
Decision rights
The decision to automate a task is also a decision about how the next person learns that job. They arrive as one choice and get made as one, whether anyone says so or not. Treat them as two and the second one goes unmade.
So it belongs with whoever automated the task, not with HR eighteen months later. The person who can see what the tool absorbed is the person who knows what nobody will practise any more.
Most companies decide this by drift. Nobody resolves to stop hiring juniors. The requisition just does not get raised, because this quarter the work is covered. Drift is the default outcome and naming it is most of the defence.
Memory and context
What leaves with the bottom rung is tacit knowledge, and it leaves quietly because it was never written down anywhere to notice missing.
The reason a senior can review AI output is that they did the work by hand for years and know what wrong looks like before they can explain why. That instinct was built by doing exactly the tasks now being automated. The skill that qualifies someone to supervise the machine was earned by doing the work the machine now does, and no documentation substitutes, because if it could be written down it would already have been automated. That is the trap in one sentence and everything below is a partial answer to it.
Execution
The role does not shrink by a percentage. Its shape changes: the doing contracts, the specifying and checking expand, and those are different skills wearing the same job title.
Someone excellent at producing work is not automatically good at judging it. Reviewing output you could have produced yourself is one skill. Reviewing output you could not have produced is a much harder one, and it is what the job becomes at the top.
Two things help and both cost something. Keep part of the manual work deliberately, as training, and be honest internally that you are trading efficiency for capability. And hire for judgment earlier than feels comfortable, because the people who would have grown into it are the ones you stopped hiring.
Governance
Some work should not be automated even where it can be, because it is the training path. That is a governance call precisely because no individual manager will make it: it costs their quarter and pays for someone else's third year.
The other half is fairness, and it decides whether any of this survives contact with the people it happens to. Nobody chose to have their job changed. A company that automates the interesting half of a role and leaves a person the exception queue has not made a smaller job. It has made a worse one. Exceptions are, by construction, the hard cases the automation could not handle. If the queue is the job now, it needs to be resourced and recognised as the harder work it is.
Observability
Watch the ratio of people who can supervise the work to people who can only operate the tool. It moves slowly, in one direction, and it is invisible on any dashboard that counts output.
Time-to-competence for new hires is the other number, and it lengthens quietly. There is less to practise on, so the on-ramp gets longer while nobody changes the expectation.
The tell arrives late and is unmistakable: every senior you have joined before the automation did. By then the fix takes years.