Some work is valuable because of the result.
Some work is valuable because it changes the person doing it.
The trouble is that these are often the same task, and we have only been measuring the first one.
The junior analyst rebuilding a model is not just making a model. The lawyer checking the cases is not just checking the cases. The designer making ten bad versions before the good one is not just producing options. They are learning what a wrong answer feels like before it has a polite explanation attached to it.
AI is very good at making the visible artefact arrive sooner. That is useful. It can also remove the repetitions through which a person earns the instinct to know when the artefact is wrong.
This is not nostalgia for busywork. A great deal of junior work was tedious, badly managed, and ripe for removal. Nobody should have to spend three years manually changing the formatting in a slide deck to prove they deserve an opinion.
But we should be honest about what is being removed with it.
A job teaches more than its job description says
Most professions have an unofficial curriculum.
It is made of small embarrassments: the first time your number does not reconcile; the first time a senior person asks the question you did not think to ask; the first time you realise the client was not asking for the thing you spent all night producing. This is not efficient. It is how a person develops judgment.
The official story about AI is that it will free people from drudgery so they can do higher-value work. Sometimes it will. But “higher-value work” is not a room people can simply walk into because the low-value work vanished.
They need a route there.
The best available evidence is mixed in an interesting way. In a large customer-support study, AI assistance helped less experienced workers improve rapidly, partly by making the practices of stronger workers available to them. But the same study found smaller gains for the most experienced workers and some evidence of reduced quality at the top end. AI can spread a pattern. That is not the same thing as teaching someone when the pattern no longer fits. Brynjolfsson, Li and Raymond
That distinction matters because the valuable part of expertise is rarely the standard case. It is knowing when the standard case is lying to you.
The risk is not only deskilling
There are three different failures, and they are easy to confuse.
Deskilling is when someone who once knew how to do a task loses fluency because the system now does it for them.
Never-skilling is when a new person never gets enough practice to develop the capability in the first place.
Mis-skilling is worse: someone learns to recognise a polished answer as a good answer. They become fluent in the performance of judgment without developing judgment itself.
That last one should make us uncomfortable. It creates people who sound ready before they are ready. The work looks calm because the uncertainty has been hidden upstream in the model.
No company would knowingly hire a pilot who had only flown in good weather. Yet plenty are building knowledge-work roles where the person is asked to supervise work they have never learned to do.
The model produced the memo. The junior person checked it. Great. Checked it against what?
A better question than “what can we automate?”
Ask this instead:
Which parts of this work are a training ground, even when the output is disposable?
That question changes the design.
Some work should still be done manually at first, not because manual is morally superior, but because the person needs to experience the shape of the problem. Some work should be done with AI but reviewed against a prior attempt. Some people need access to failure cases, not just examples of perfect output. Some work needs to become simulation: a safe place to make decisions, see consequences, and explain the reasoning before the stakes are real.
That is more deliberate than the old apprenticeship. It may even be better. The old system wasted plenty of human time. But it cannot be replaced by giving every new hire a very articulate first draft and hoping that competence will enter by osmosis.
The point is not to preserve the work that disappeared.
It is to preserve the person the work was making.