The junior work was the training

Matt Beane spent years watching surgical residents learn, and found that in robotic operations most of them had stopped touching the patient. They watched a screen while the attending worked the console. On paper the training was intact. The hours where the skill is actually built had gone. Most of what I read last week had the same shape, just in offices rather than operating theatres.

A few of the things I read last week:

• Mustafa Seref Akin modelled the decision and found that a firm automating junior tasks never pays for the learning pathway it removes, which he calls a skill-ladder trap. (SSRN)
• In a field study of 523 early-career professionals given a purpose-built AI agent, critical thinking, AI literacy and domain knowledge did not predict who got value from it and who did not. (HBR)
• About a thousand students scored 48% better on a maths unit while they had GPT-4, and 17% worse on the test once it was taken away; a version that gave hints instead of answers did no such harm. (Wharton)
• PwC is planning to cut entry-level audit hiring by a third to two fifths through 2028, while expecting the people it does hire to do manager-level work within three years. (Financial Times)
• Two case studies found freelancers' work moving from producing to validating AI output, with nothing left to rebuild the skill that move consumes. (arXiv)

None of that is a question about how many people we hire. It is a question about where, in an ordinary week, a person gets better at their job.

The saving is ours, the shortage is everyone's

Akin's paper is short, and it explains a lot. When a firm automates the tasks its juniors used to do, it collects the saving in full. The cost, which is one fewer senior person in eight or ten years, lands on a market rather than on a balance sheet. So every firm has a good reason to do it and no firm has a reason to stop. Nothing in that is careless. It is what a sensible finance director does with a line item, and the learning pathway was never a line item.

We have been reading this as a hiring story, and I am not sure that is where it lives. As we posted on Monday, employment for 22 to 25 year olds in the most AI-exposed occupations sits about 19% below where it should be. On Wednesday we shared the other side, the New York Fed finding firms retraining rather than cutting, roughly eight to one. Both of those are about headcount. Neither tells us whether anybody is still learning anything, and that is what I would want to know first.

What the study found is stranger than the headline

The field study of 523 early-career professionals is the piece I would hand to a colleague. The researchers, at the University of Texas at Austin and KPMG, watched people do real business tasks with an agent built for the job. What surprises me is what did not predict who came out ahead: not critical thinking, not AI literacy, and not how much they knew about the domain.

If that holds up, it takes the problem out of hiring and out of the curriculum and puts it in the design of the work. The Wharton experiment points the same way from the other end. Same students, same model, and the harm disappeared when the tool was built to give hints rather than answers. What differed was what the tool in front of them would do for them, and somebody chose that.

The freelancers in the arXiv paper show the cost of not choosing. Their work moved from making things to checking things, which sounds like a promotion and is not, because checking only builds judgement in someone who already has some. Ethan Mollick put the general case well at the end of August: if agents make every interesting decision and leave people with the approvals, the exceptions and the failures, we will have automated the wrong half of the job.

Teaching a team to actually check a model's output is what the Verify playbook is for, and it is a better use of an afternoon than another tool demo.

The fair objection is that we are blaming AI for something else entirely. Martha Gimbel and the Yale Budget Lab keep looking at the labour data and finding no meaningful macroeconomic effect from AI; she thinks much of what gets announced as AI restructuring is cover for the ordinary aftermath of over-hiring, rate rises and tariffs. She may well be right about why the hiring numbers moved, and I would rather she were. It does not settle this one. Whatever is driving the graduate hiring line, the model is doing the first draft and the first reconciliation in a great many places, and those were the repetitions.

Some organisations are treating it as a design job. IBM's decision to triple entry-level hiring got a lot of attention on Monday, and the less-quoted part interests me more: Cognizant has opened entry-level recruiting to liberal arts and non-STEM graduates. Two thirds of the CEOs in that same piece said AI is increasing their entry-level headcount rather than shrinking it. None of it is proven, but they are arguing about the right thing.

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

We are usually brought in after the decision, once the tooling is bought and the junior roles have been redrawn, and asked to build a programme around it. Two things worth doing instead, both conversations rather than projects. Take one role and write down what a person in it spent their first year doing three years ago, then mark what a model now does. Whatever is marked was doing the teaching, and the next question is where those hours live now. Then find someone senior and ask which judgements they got badly wrong early on, and what it was that taught them. In my experience the answer is never a course, and it is usually something we have since made more efficient.

If you would rather work this out with other people who are stuck on the same thing, that is what we do together in the AI Capability Academy programme at aicapability.academy.

The provocation

Name the person in your organisation who is two years in and will be running something in eight. What did they learn this month, and who watched them learn it?