What is the shift from a pyramid to a diamond org structure?

The traditional org chart was a pyramid: a narrow leadership tip, a working middle, and a wide base of juniors and graduates doing high-volume routine work. As AI absorbs that routine work, companies are hiring fewer entry-level people while the middle band stays large or grows. The resulting shape — compressed bottom, bulging middle, unchanged top — is a diamond.

Is entry-level hiring really down 80% at companies adopting AI?

No. That figure is a widely repeated misquote. The Harvard working paper by Hosseini Maasoum and Lichtinger, covering 65 million workers across more than 280,000 US firms, found junior employment fell by roughly 9% after six quarters at firms adopting generative AI, while senior employment continued to rise. Separately, Stanford's Digital Economy Lab found employment for 22-25 year olds in highly AI-exposed occupations sits about 19% below where it would be had it kept pace with less-exposed occupations. Both are serious structural signals, but neither is an 80% collapse.

What happens if every company stops hiring juniors at the same time?

Cutting graduate intake is rational for a single firm but self-defeating across an industry, because the mid-level talent market is a shared pool. If nobody grows mid-level people for several years, everyone ends up bidding for a cohort that was never produced. The predictable consequences are mid-level pay inflating faster than budgets, shorter tenure as those people become permanently poachable, and firms that kept a pipeline effectively subsidising the training of those that did not.

Why has years of experience stopped working as a proxy for capability?

The pyramid made tenure a reasonable proxy because everyone entered at the base and progressed through the same work, so time served correlated with judgment. In a diamond, two people with the same three years can differ substantially depending on how much work they did themselves versus supervised through AI. The proxy no longer tracks what it stood in for — and with fewer junior seats, each selection decision carries more weight, so capability needs to be measured directly rather than inferred from a CV.

What should HR and CXOs do about the diamond shift?

Treat it as a deliberate trade rather than an accidental one. Boards should require a multi-year mid-level supply forecast alongside any entry-level reduction, since the saving and the cost fall in different budget years. CFOs should price the replacement cost of the middle no longer being grown. CHROs should keep a smaller but deliberately designed junior intake with an explicit capability route and measurement attached. Function heads should name which judgments used to be learned from the work AI now performs, and design how those get learned instead.

Pyramid to Diamond: Why Companies Are Hiring Fewer Juniors in the AI Era

Q: Why does cutting junior hiring matter if AI does the work?

Because the junior base did two jobs, not one. It delivered throughput on routine work — which AI now handles — and it built judgment through repetition, which nothing has replaced. A junior who reconciles a thousand records develops an instinct for what a wrong record looks like, and that instinct is what makes them a useful mid-level hire later. Remove the work and you remove the curriculum, so the pipeline into the middle disappears with it.

In short: Organisations are not getting flatter. They are getting diamond-shaped — a compressed junior band, a bulging middle, an unchanged top. AI absorbed the routine work that entry-level roles were built around, so companies stopped buying that layer. The cost is not the salaries saved. It is that the wide base was the training system: the place where judgment got built by doing the boring work. Cut it, and you are no longer growing mid-level people — you are bidding for them against everyone else who also stopped growing them.

Every org chart most of us learned was a pyramid. A narrow leadership tip, a working middle, and a wide base of juniors and graduates doing the volume work — the first-pass research, the first-draft deck, the reconciliation, the ticket triage, the document review. The base was the biggest cost line and the biggest headcount. It was also, though almost nobody wrote it down this way, the company's apprenticeship programme.

That shape is changing. The base is compressing while the middle widens. The result is a diamond.

The pyramid did two jobs: it got work done cheaply, and it turned inexperienced people into experienced ones. The diamond only does the first.

What the evidence actually says

This is a topic where the numbers circulating are considerably worse than the numbers in the research. It is worth being precise, because the strategic response to a 9% shift is different from the response to a collapse.

Study What it measured Finding

Stanford Digital Economy Lab
“Canaries in the Coal Mine?”
ADP payroll data, updated Aug 2026 Employment of 22–25 year olds in highly AI-exposed occupations, against same-age workers in less-exposed occupations. Roughly 19% below where it would be had it kept pace. The gap was 15% in July 2025. Experienced workers show no comparable gap.

Hosseini Maasoum & Lichtinger
“Generative AI as Seniority-Biased Technological Change”
65m workers, 280,000+ US firms, 2015–2025 Junior vs senior employment at firms that adopted generative AI, against firms that did not. Junior employment fell about 9% after six quarters at adopting firms. Senior employment kept rising. Driven by reduced hiring, not redundancies.

A number to stop repeating. You will see “entry-level hiring is down 80% at companies adopting AI” attributed to the Harvard study. That is not what the paper found. The finding is a ~9% decline in junior employment after six quarters at adopting firms. The real number is still a serious signal — a 9% structural shift in the shape of the workforce, driven by hiring rather than layoffs, is a large effect. But a board briefed on 80% will authorise panic; a board briefed on 9% with the mechanism explained will authorise strategy.

Two independent datasets, two different methods, and the same shape: the bottom of the pyramid compresses while the middle and top do not. That is the diamond, and it is showing up in payroll data, not in predictions.

Why the diamond forms

Entry-level roles were never defined by the difficulty of the work. They were defined by its volume and routineness. Summarise these fifty documents. Pull this data into a model. Draft the first version. Check these records against those records. That work was the reason to hire a junior, and it is precisely the work generative AI does acceptably well on the first pass.

So the economics invert. A manager who needed four analysts to get through the volume now needs one who can direct AI through it and check the output. That manager does not need a graduate — they need someone who already knows what wrong looks like. Hiring shifts to the middle. Nobody decided to stop developing talent; a thousand local, sensible hiring decisions added up to it.

The part that costs you later

Here is the trap. The wide base was not just cheap labour. It was the mechanism by which people acquired judgment. A junior analyst who reconciles a thousand records develops an instinct for what a wrong record looks like. That instinct is what makes them a useful mid-level hire three years later. The boring work was the curriculum.

What the wide base delivered In the pyramid In the diamond

Throughput on routine work Juniors, in volume. Solved — AI does it faster and cheaper. This is the win, and it is real.

Judgment, built by repetition Acquired as a by-product of doing the volume work. Unsolved. The work that taught it is gone; nothing replaced the teaching.

A pipeline into the middle Internal, predictable, and cheap relative to market rates. Unsolved. The middle is now bought on the open market.

Succession and key-person cover Bench depth formed naturally underneath every role. Unsolved. A thin junior band becomes a thin senior band, on a delay.

Cost of the middle Largely internal — you grew it. Market-priced, and rising as every competitor bids for the same people.

Three of those five rows are unsolved, and all three are invisible this year. That is what makes this a governance problem rather than a hiring problem: the saving lands in the current budget and the cost lands in someone else's, two to five years out. For a related mechanism inside the work itself, see cognitive debt and distributed deskilling.

The market maths nobody is doing

Cutting graduate intake is rational for one firm in isolation. The problem is that it is rational for every firm simultaneously, and the middle of the market is a shared pool. If the whole industry stops producing mid-level people for four years, then in four years the whole industry is competing for a cohort that was never grown.

The predictable results: mid-level compensation inflates faster than budgets, tenure shortens as those people become permanently poachable, and the firms that kept a junior pipeline end up as involuntary training providers to the ones that did not. If you are modelling only the headcount saved this year, you are modelling one side of the trade.

What breaks in the HR operating model

Most HR machinery assumes a pyramid. When the shape changes, processes that were sound start producing wrong answers without anyone getting a warning.

Process Pyramid assumption What to change

Campus hiring Hire a large cohort, let attrition and performance sort it out. Fewer seats means selection accuracy matters far more per seat. Volume filtering has to become evidence of capability. See campus hiring in the AI era.

Experience as a proxy “Three years in role” reliably implied a level of judgment. The link is broken: three years supervising AI is not three years doing the work. Measure the capability directly rather than inferring it from tenure.

Career architecture Progression by time served through evenly spaced rungs. The first rung is missing. Either build an explicit route across the gap, or accept you are a permanent net importer of mid-level talent.

Workforce planning Headcount by level, forecast on growth. Model the pipeline, not the pyramid: who is on track to replace the middle, and by when.

Learning & development Courses supplement learning that happens on the job. The on-the-job half has been removed. L&D now has to deliberately manufacture what repetition used to supply for free.

What to actually do

This does not argue for hiring juniors you have no work for. It argues for being deliberate about a trade you are currently making by accident.

Owner Action Why now

CEO / Board Ask for the four-year mid-level supply forecast alongside any entry-level reduction. The saving and the cost sit in different budget years. Only the board sees both.

CFO Price the replacement cost of the middle you are no longer growing. A cut that shifts cost from payroll to acquisition is not a cut; it is a transfer, usually at a worse rate.

CHRO Keep a smaller, deliberately designed junior intake with an explicit capability route — and measure it. A thin pipeline still works if the people in it are the right ones and they advance faster. That depends on selection and measurement, not volume.

Function heads Name which judgments used to be learned from the work AI now does, and design how they get learned instead. Only the function knows what the repetition was actually teaching. Nobody else can reconstruct it.

The measurement problem underneath all of it

The pyramid quietly supplied HR with its most useful signal: time served. Because everyone entered at the base and moved up through the same work, years of experience was a decent proxy for capability. It was never exact, but it correlated well enough to run hiring, pay and promotion on.

The diamond removes that. Two people with the same three years now have genuinely different capability depending on how much of the work they did themselves and how much they supervised. The proxy has stopped tracking the thing it was standing in for — and for a smaller junior intake, where each seat carries more weight, guessing is more expensive than it used to be.

So the response is not to hire on the old signal harder. It is to measure the capability directly: what someone can actually do with AI, where their judgment holds and where it does not, and what evidence supports the claim. That is the work behind our AI-era capability framework, the levels of working with AI, and the difference between inferred, verified and proven skills. If you are cutting the base, the people you keep have to be measured properly rather than assumed.

What to take away

The shift is real and visible in payroll data, but it is a structural tilt — roughly 9% at adopting firms, and a 19% gap for 22–25 year olds in AI-exposed occupations — not the collapse the “80%” headlines describe.
It is driven by hiring freezes, not redundancies. That makes it quiet, cumulative and easy to miss until the middle is empty.
The wide base was the training system. Removing it solves throughput and leaves judgment, pipeline and succession unsolved.
Rational for one firm, irrational for all firms at once: everyone buying mid-level from a pool nobody is filling.
Years of experience has stopped being a safe proxy for capability. With fewer junior seats, each selection decision carries more weight — so measure capability directly rather than inferring it from tenure.

The pyramid is becoming a diamond — and the training floor went with it