The bottom was the training programme

In plain terms. The junior work AI takes first is how people used to learn the job. If that's right, the saving shows up this year and the cost shows up later, as experienced people you can't find. Decide which work you keep for people because it teaches, and check how many agents your plan expects one person to supervise. None of this is a reason to slow down, but it does mean deciding, on purpose, what people keep doing.

Where this starts

When the work gets cheap sets out the short version: the technology lag has all but gone, the organisational lag hasn't, and what a business is worth moves to its judgement. This piece is about the first of the three things to plan for, and the one I'd put in front of any leadership team first: what AI does to how people learn.

AI is doing one of two things to your people, and for the first few years they look the same

There are two different things AI could be doing to the people in your firm.

The first is a productivity shock. People do the same work faster. Firms hire fewer juniors for a while, the shape of the organisation distorts, and then it settles back, because the way people learn the job hasn't changed. This is what most plans assume, whether or not they say so.

The second is a learning shock. The entry-level work being automated was also the curriculum: the thousand small, dull, correctable tasks through which a junior became someone whose judgement you could bill for. Automate the curriculum and the organisation doesn't settle back, because the supply of judgement dries up a few years later, when the juniors you didn't train would have become the managers you can't find.

For the first few years the two look the same. Fewer graduates, better margins, a thinner base. The difference only shows when the missing year-groups fail to arrive at mid-level. The distinction comes from "Pyramids, Diamonds, and Oscillations" (Friebel, Huang, Li, Shukla and Zhang, April 2026), and its value is that it turns a metaphor into something you can test. The paper doesn't say which shock AI is. On the evidence so far I think it's a learning shock, for three reasons this piece works through: the junior work really was the curriculum, the middle of the firm isn't what the plan assumes, and the cost and the benefit never appear in the same year's accounts.

The junior work was the curriculum

Ask anyone senior in a professional firm how they learned judgement and they will describe work that looked like drudgery. First-pass reviews. Month-end checks. Bundling. Drafting the version a partner would tear up. Nobody designed that work as a curriculum, but that is what it was, and the tearing up was the teaching.

The same is true outside professional services. The graduate in a finance team who spent a year matching invoices learned what a normal supplier looks like, which is the knowledge that later lets them spot an abnormal one. The junior in a claims team who processed a thousand routine claims learned, without anyone teaching it, the shape of a fraudulent one. Take the routine work away and the pattern recognition that came free with it goes too.

It holds in product businesses as well. A junior engineer who spends a year fixing small bugs in other people's code learns how the system really fits together, which is what later lets them change it without breaking it. A year on the support queue teaches what customers actually struggle with, and that's where a lot of good product managers come from. If agents take the small bugs and the routine tickets, decide what you're keeping for people, or you'll end up with a team that can ask for code and can't debug it.

The learning shock is sharpest where judgement is the product and people learn it by doing the work: professional services, finance, operations, support. A twenty-person software company meets it later and differently, in who its first hires learn from. It's the same question at a different size.

This is why the graduate intake decision is a bigger decision than it looks. A firm that halves its intake this year because the drafting can be done by software is making a bet that the drafting was only drafting. If it was also how people learned, the saving arrives this year and the cost arrives later, somewhere else in the business.

Sort the work by what it taught, before you automate any of it

Most automation decisions sort tasks by one thing: what the output is worth. That is the right first question and it isn't enough. The second question is what doing the task taught the person who did it. Put the two together, value of the output up the side and judgement taught along the bottom, and every entry-level task lands in one of four places.

High value, taught nothing. Automate it, and bank the gain in the same quarter. There is no loss here and there is no reason to wait.

High value, and it was the curriculum. Keep a person deciding while the software drafts, because the points where they disagree are what the junior learns from. The drafting gets faster. The learning survives, as long as the person is deciding and not just approving.

Low value, taught nothing. Delete it. No AI required, and the piece on value explains why this is the most valuable box on the grid.

Low value, but it was how juniors learned. Automate the volume, keep a set share of it for people, and say so out loud. Ten per cent of the routine checks done by hand is a training budget, and it should be defended like one.

The two right-hand boxes, where the task taught something, are where the learning shock happens. Most firms sort only by value, so they automate the right-hand column without noticing that it was a curriculum. The grid makes that trade-off visible at the moment it is being made.

Two middles are going in opposite directions

The picture everyone draws is a pyramid becoming a diamond: fewer juniors, a fat middle of people supervising machines. The picture isn't mine, and it hides the two things that matter.

The first is that there are two middles and they are moving in opposite directions. One middle brokered information up and down the chain: it collected, summarised, chased and reported. That middle is being cut, and rightly, because a system that holds the full context can do much of that job with less lost on the way. The other middle is new. It verifies what the software produced, owns the exceptions, and decides when an agent can be trusted with more. That middle is being created, and it shows up in no dataset, because a person supervising a team of agents looks to every HR system like an individual contributor with no reports. Firms that decide what to cut by its layer on the chart will cut both middles together. Decide by what the role does.

A person can supervise fewer agents than the plan assumes

The second thing the diamond hides is that the new middle is smaller than the plan needs it to be.

Every plan I have seen for how many agents one person can run assumes that one supervisor can run more agents next year than this year, and that the ratio keeps improving as the agents get better. Two well-established results say otherwise. The first is queueing arithmetic that every operations manager knows: once a person is past roughly seventy per cent busy, the queue of exceptions waiting for them grows sharply, because exceptions arrive in clumps. A plan built to maximise agents per person loads the supervisor towards a hundred per cent by design. The second is older. Human-factors research on automated systems, going back to Bainbridge's "Ironies of Automation" in 1983, keeps finding the same thing: the better the automation, the worse the human becomes at stepping in when it fails, because they no longer do the work often enough to stay good at it. A supervisor who only reviews loses the skill that made them worth having as a supervisor.

So the fat middle is a transitional shape that degrades unless the supervisor keeps doing some of the work rather than only checking it. Put those two results into the leverage plan and the ratio has a ceiling, and the ceiling is a good deal lower than the spreadsheet says. The gap between the middle you have assumed and the middle you can staff is risk you are carrying without knowing it, and Earned autonomy is about how to measure it.

When this can be shown wrong: 2029 to 2031

Two facts look like they contradict each other. The number of associates for every partner at US law firms (the industry calls it associate leverage) reached a record at the end of 2025, in Thomson Reuters' index, while firms kept hiring. In the same period, several of the largest UK law and accounting firms were reported to be cutting their trainee and graduate intakes for 2026 and 2027, and two of the Big Four were reported to have roughly halved their UK partner promotions.

They are the same story read at two points in time. Leverage is a stock: it reflects hiring decisions made two to five years ago. Intake is a flow: it reflects the decision being made now. The stock peaks while the flow is already being cut.

So here is a claim with a date on it. If the learning-shock reading is right, associate leverage in Thomson Reuters' index peaks and turns down between 2029 and 2031, as the reduced 2026 and 2027 intakes fail to arrive at mid-level. If it's still climbing in 2031, I was wrong about the mechanism, and this page will say so. The about page keeps score.

What to do with this

If you run the business. Before this year's intake decision, run the grid on the junior roles in one department. It takes an afternoon. Then decide, in writing, which judgement the firm will keep producing in people and how. That decision can't be delegated to the software.

If you sit on the board. Ask for the supervision ratio the AI plan assumes, and ask what happens to exception handling at that ratio. If nobody has the number, the plan has a hole in it where the middle used to be.

If you're building a start-up. You have no pyramid yet, so you get to design one. If agents do the junior work from day one, decide now who your first ten hires learn from, and what work you keep for them because it teaches.

If you invest. For the first few years a firm building a durable position and a firm eating into its own future look the same in the accounts. The tell is the intake and the training budget. A firm that has cut both and has no written answer to which judgement it is keeping is cheaper now and worth less later.

What this rests on

The seventy per cent utilisation point is standard queueing theory, and the automation result is Bainbridge (1983) and the literature that followed it. The productivity-versus-learning distinction is from Friebel, Huang, Li, Shukla and Zhang, "Pyramids, Diamonds, and Oscillations" (April 2026). How automation changes the way careers develop is studied in Afrouzi, Blanco, Drenik and Hurst, "Automation, Learning, and Career Dynamics" (2026). The trial in which junior engineers using AI help scored lower on understanding than those without it is Anthropic's. The law-firm leverage figure is from Thomson Reuters' Law Firm Financial Index for the fourth quarter of 2025; the UK intake and promotion cuts were reported in the trade press in 2025 and 2026 and I haven't verified them independently. The curriculum grid and the two-middles reading are mine.

The strongest objection

"AI is an apprenticeship accelerator. Juniors learn faster with it."

Partly right, and the right-hand boxes of the grid are where it is right. A junior who drafts with the software and then argues with a senior about the draft can learn faster than one who spent a week producing the draft alone. But the acceleration only happens if the junior is still in the loop and still deciding. The firms cutting intake are removing the junior, and a curriculum with no students accelerates nobody. There is also a question of what gets measured. The strongest study on this side, Brynjolfsson, Li and Raymond's work with customer-support agents, found novices improved most while using the tool. It didn't test whether they could do the job without it. When that has been tested, the answer has been less comfortable: in Anthropic's trial, junior engineers who leaned on AI understood their own code less well. The objection describes what a firm could do. The intake numbers describe what firms are doing.