I think the most expensive mistake in AI productivity is counting the work that disappears without asking what people used to learn by doing it.
An assistant can produce an answer for a junior employee. That isn't the same thing as producing a more capable employee. Confuse the two and you can end up with a business that's quicker at finishing work and worse at understanding it.
The sales pitch naturally favours the first number. AI vendors can sell completed tasks. An executive can put the saving in a board presentation. Neither gets much applause for preserving an awkward half-hour in which someone struggles with a problem and finally understands it.
Call me cynical, but the official story about freeing everyone for higher-value work skips a rather useful question: where are the people qualified to do that work supposed to come from?
The answer arrived. The learning didn't.
Anthropic ran a randomised study with 52 mostly junior software engineers learning an unfamiliar programming tool. One group could use an AI assistant. The other worked without it. Afterwards, both took a quiz on what they'd just been doing.
The AI group averaged 50%. The other group averaged 67%. The assisted group finished about two minutes faster, but that speed difference wasn't statistically significant.
That's an uncomfortable exchange rate. Less understanding, without even a convincing saving on the clock.
This was a small coding study measuring immediate comprehension, not proof that AI ruins careers. The result still deserves more attention than another demonstration of software completing a task while a delighted presenter watches.
The detail I find most useful isn't that AI users did worse on average. It's that some used the assistant to ask questions, request explanations and understand the material, and those patterns were associated with stronger quiz results. That doesn't establish that the questioning caused the better scores. It does make the lazy distinction between using AI and not using AI look fairly useless.
Getting an answer and learning how to reach one are different products.
If the tool does the whole exercise while you watch, calling it training is generous. It's a driving school that sends you a taxi receipt and congratulates you on passing.
Don't turn this into a ban
I'm not arguing for compulsory suffering. Nobody needs to earn the right to use a calculator by spending a winter with an abacus.
There's good evidence for the other side too. A field study covering 5,179 customer support agents found that access to an AI assistant increased issues resolved per hour by 14% on average. The gains were larger for novice and lower-skilled workers, and the researchers found suggestive evidence of learning.
That matters. AI can help less experienced people perform better, not merely make experienced people faster. A blanket ban would throw away precisely the assistance that could help someone get started.
My judgement is that we need to separate two questions. Can someone complete this job with the machine beside them? And are they getting better at knowing when its answer is wrong?
Those aren't rival ambitions. They're both part of building a business that won't fall apart when the easy cases run out.
Take a hypothetical junior colleague preparing a customer quote. AI could turn the brief into a polished proposal. Lovely. But the useful learning might sit in the awkward bits: which request changes the scope, which assumption makes the price nonsense, which promise the business can't safely make.
A flawless-looking document doesn't tell you whether the person can spot any of that. Nor does asking them to approve the document. A signature isn't a substitute for understanding. It's just understanding's considerably cheaper stunt double.
I'd judge that workflow by whether the colleague can explain the assumptions and handle a changed brief, not simply whether the proposal arrived before lunch. That's a management choice, not a feature you can switch on in a subscription.
Someone has to grow the experts
The commercial temptation looks obvious to me. Sell the output, count the hours removed and leave the development of judgement off the invoice. The vendor gets a persuasive productivity story. The business gets an immediate saving to discuss. The bill for missing expertise, if it comes, arrives later.
That isn't evidence of a plot. It's an incentive problem. What gets counted this quarter can crowd out what makes the company competent next year.
I don't think the answer is to preserve every junior task. Some work deserves to disappear. But removing drudgery and removing practice aren't automatically the same thing. The useful distinction is whether the task teaches someone how to make a consequential decision, rather than merely how to survive a spreadsheet.
The companies I'd take seriously here aren't just promising to finish the exercise. They're making it easier to question the answer, inspect the reasoning and find out what the person actually understands. That's a harder proposition to demonstrate than a stopwatch. It's also a more interesting business than renting out confidence.
We can't keep promising human oversight while quietly dismantling the route by which humans become qualified to provide it.
An expert isn't a junior with a better subscription.
Who gets paid to grow the experts when everyone else is paid to skip the apprenticeship?