We are automating the work people learn on

Florian Herrengt wrote about an engineer who was asked where the data in his own system comes from. The answer was: "Hmm... actually I don't know. Let me ask Claude." He published that on 11 August, about his own trade, software. I think the same sentence is being said this month in finance teams, policy units and legal departments, and in those places nobody is writing it down. What worries me is not that the engineer did not know. It is that there is no longer an obvious hour in his week when he would have found out.

A few of the things I read last week:

• In a randomised experiment, AI assistance cut the education-based performance gap among 1,174 adults from 0.548 to 0.139 standard deviations, but the gain survived the tool being removed only where heavy use came with sustained effort. (arXiv)
• A study of 523 early-career professionals sorted them into AI apprentices, delegators and amplifiers, and found that critical thinking, AI literacy and domain knowledge did not predict who landed where. (HBR)
• Luis Garicano and Luis Rayo model apprenticeship as a trade in which juniors buy knowledge with routine work, and find it holds only while a trained graduate produces more than about 2.72 times what a novice does. (CEPR)
• J.P. Eggers, Sarah Ryan and Asha Dinesh write that AI has made execution cheap across the innovation cycle, so the scarce work is now deciding what to build. (HBR)
• Judah Adeniyi, surveying 361 workers and reviewing 121 studies, finds that AI training which is badly timed or unsupported raises burnout and early-retirement intentions among workers over 55. (The Conversation)

Read together, they are all circling the same question: where is a person supposed to get good now.

Nobody designed the old apprenticeship

It was never a programme. The junior did the tedious work because it was tedious and somebody had to, and doing it taught them the shape of the thing. The reconciliation taught you where the numbers came from. The first draft of the memo taught you what the partner actually cared about. None of it was on a curriculum and none of it had a budget line, which is exactly why it can disappear without anyone taking a decision.

Garicano and Rayo make that exchange explicit. In their model the junior's routine work is the currency they pay with, and the senior's knowledge is what they buy. Rayo puts it about as bluntly as an economist can: now that AI does that work essentially for free, the currency apprentices used to buy knowledge with is vanishing. Training survives where there is still a real gap between what a trained person can do and what a novice with a good model can do.

The randomised experiment says the same from the worker's side. Take the tool away and most of the levelling goes with it, and what stays belongs to the people who used it hard and kept working at the problem themselves. The levelling is real, and it is mostly a loan.

Hiring juniors is not the same as teaching them

There is a good case against all of this and I want to put it properly, because the doom version of this argument is lazy. Companies are not shutting the door. IBM said it would triple its US entry-level hiring this year, Salesforce opened a thousand graduate and intern places, and Dropbox widened its internship intake by a quarter. People have been announcing the end of the junior role since the spreadsheet arrived, and they have been wrong every time.

I take that seriously. But hiring a junior and teaching one are two different acts, and only the first shows up in those numbers. The study of 523 early-career professionals is the part I find hardest to argue with. What separated the people who grew alongside AI from the people who simply handed work to it was not their critical thinking, or their AI literacy, or how much they knew about the domain. It was how they engaged with it. That is not something you can screen for at the door, which means it has to be built inside the building, in the way the work is arranged.

This is also not only about the young. Adeniyi's piece is the one I would most want a head of L&D to read, because training that arrives on top of a full workload lands as one more demand, and among workers over 55 it pushes people toward the exit. We could lose the top of the ladder and the bottom of it in the same year for the same reason, which is that nobody adjusted the work around the learning.

As we noted on Friday, three in four employers now screen for AI skills while about one in five trains everyone. Eggers, Ryan and Dinesh say execution has become cheap and judgement scarce, and I think they are right. But judgement about what is worth building is built by having built things.

If you work in L&D, HR, or transformation

We are usually asked for a programme when what is missing is a decision about who does which work, and I know that is a harder thing to sell. Two things worth trying this week anyway. Take one workflow an agent has taken over and name the part of it a person still does by hand on purpose, then write down who does it and how often; if there is no such part, you have found something worth raising. Then find the person on a team who is a year or two ahead of the others, and give them two protected hours a week to sit beside someone behind them, inside the workload rather than on top of it.

The provocation

Think of the most junior person you work with. What did they do last month that they will still be better for in three years, and if you cannot name anything, who in your organisation is supposed to be responsible for that?