Entry-level hiring is collapsing, and the pyramid with it

Because hiring beginners stopped without anyone having to make redundancies, and that is what makes the phenomenon so hard to see. The Stanford work run on ADP payroll data, updated in August 2026, puts employment among 22-25 year olds in highly AI-exposed occupations about 19% below its expected path, against 16% in the previous version.

The nature of that decline matters as much as its size. These are not redundancy plans, which would make headlines and trigger consultations, but roles that are never posted. A hire that does not happen produces no event, appears on no dashboard, and leaves nobody with a decision to justify. That is precisely the property that let the phenomenon run for three years before it was named.

What does the study measure, and what does it not?

It compares, within the same firms, employment of 22-25 year olds in highly AI-exposed occupations against the same ages in less exposed ones. That construction is what gives it its strength: it neutralises shocks specific to each company, since both populations live through the same economy, the same order book and the same finance director.

The second fact is more interesting than the first. The decline concentrates in applications where AI automates the work, and stays muted where it augments it, meaning where a person remains at the centre and gains reach. The same technology therefore produces two opposite effects on employment depending on how it is deployed, which moves the question from the occupation to the task.

What it does not measure, and this has to be said just as plainly, is full causation. High interest rates, the end of training budgets whose payback is counted in years, and normalisation after the 2021 over-hiring are all acting at once. The study establishes the shape of the decline, not the whole of its origin, and an article selling you AI as the sole culprit would also be selling you the remedy.

Why an IT services firm is the most exposed structure

Because such a firm does not sell a product, it sells a pyramid, and the base of that pyramid is recruited every year.

The model holds through a simple balance: junior profiles hired out of school, trained on supervised assignments, billed at a rate that leaves a margin because the entry cost is low, then promoted into roles where they supervise in turn. Each floor assumes the one below. When the base narrows, nothing happens in the first year, and the shortage of mid-level profiles arrives three years later, at precisely the moment those profiles are most in demand.

The narrowing here comes from both sides at once, which is new. Clients more often refuse a beginner on an assignment, explaining that a well-equipped senior does the work. And the market trains fewer beginners, so the pool from which you will recruit your mid-level people in three years is currently failing to form. You can only act on the first movement, and that is already a lot.

Add that the cost of a consultant waiting for an assignment does not depend on their seniority, and you get the temptation to stop recruiting at the bottom. It is rational quarter by quarter, and it is the decision that costs most over five years. We have costed what a consultant on the bench actually costs elsewhere, and the trade-off sits exactly there.

The criterion that tells you which junior roles hold

The distinction between automating and augmenting is not an economist’s refinement, it is a decision tool you can apply role by role this week.

Take a junior job description and look at what the person would actually do during their first twelve months. If the answer is a list of tasks whose steps are known in advance, repetitive, with no judgement to exercise, then that role was already an automation role and it will disappear, whatever your view on the matter. If the answer contains trade-offs, relationships, cases where you have to decide on incomplete information, the role holds and the tool extends its reach.

That grid overlaps the one we use to decide what belongs to an agent and what belongs to an automation, and that is no coincidence: it is the same criterion applied to a human rather than to software. A role that does not survive automation was a role whose content fitted in a procedure.

The practical consequence is uncomfortable for companies that hired juniors to absorb volume. It is rather good for those who hired them to grow future supervisors, provided they accept that the content of the first twelve months has to change.

What the firms still hiring juniors are doing

They are not doing it out of philanthropy, which is what makes their example useful. Some large technology companies increased their intake of beginners in 2026 after cutting it, by changing what a junior does rather than how many juniors there are.

The shift is always the same: less first-draft production, more client contact, and above all systematic review of what the AI produced. That last task happens to be an excellent apprenticeship, because it forces you to understand why an answer is good before having had to produce one yourself, which is the reverse of the usual order and probably faster.

The bet they are making deserves stating: they hold that the coming scarcity will be of people able to judge work, not of people able to produce it. It is the same conviction we defend in writing about what remains for humans when agents work, and it has an awkward consequence: training somebody to judge means showing them a great deal of work, including bad work.

What an IT services firm can decide this year

Three decisions can be taken without waiting for the debate about causes to be settled, and none of them requires a bet on what the technology will do.

Change the content of the first twelve months before changing the volume. A junior whose first year consists of producing first drafts is a junior whose role is on borrowed time. The same junior assigned to reviewing what gets produced, to the client relationship and to tracking files learns faster and holds a place that does not get cut. That means rewriting job descriptions, which is work rather than budget.

Sell the beginner differently. A client refusing a junior is really refusing a quality risk, not a profile. Presenting them paired with a senior who reviews their work, and saying so explicitly in the proposal, answers the actual objection. It is more honest than presenting them as experienced, which the market has done a great deal and clients have learned to detect.

Cost what not hiring will cost. The margin gain is immediate and visible, the cost arrives in three years as a mid-level profile you will have to buy on the market at scarcity prices. Nobody carries that line in a budget, which explains why the trade-off is almost always resolved the same way. Writing it down, even roughly, is often enough to move the discussion.

None of the three depends on a tool, and that is deliberate. An agent helps you hold deadlines and workload, it does not decide the composition of your pyramid, and confusing the two would be exactly the kind of shortcut this blog refuses.

What we owe you, since we sell the tool

We sell an agent to recruiters, and the technology that makes it work is the one compressing entry-level hiring. Writing this article without saying so would amount to selling a remedy while concealing that we also hold the disease, and that is the rule that makes everything else on this blog credible.

What I honestly believe is that responsibility does not sit at the level of the tool but of the role somebody decides not to open. An agent that prepares dossiers does not remove a junior job; a management team observing that dossiers now get prepared faster and concluding that one fewer beginner is needed does. The difference is not a rhetorical detail, it is where the decision is taken, and it is human.

What I believe less readily is that the sector will spontaneously make the right call. An economy that gains fifteen points of margin by not recruiting at the bottom will do it, unless it counts the cost of lacking those people in three years, which nobody is in the habit of putting in a budget.

Which leaves the question that follows immediately, when the base narrows and delays lengthen instead of shortening: where does the time in a recruitment actually go, and how much of that delay has nothing to do with the market? That is the subject of the time-to-fill gap between firms.