Two true statements, made in the same month, that look like they contradict each other. The first: there is no mass unemployment caused by artificial intelligence — 91% of companies kept or grew headcount over the last three months. The second: employment among 22-to-25-year-olds in occupations highly exposed to AI is about 19% below where it would otherwise be.
Both are measured, published and correct. They coexist because what is happening is not jobs disappearing. It is the advantage changing hands — and it is possible to know, fairly precisely, which side of the counter you land on.
The aggregate: there is no apocalypse
The World Economic Forum's Future of Jobs Report projects 170 million roles created and 92 million displaced by 2030 — a net gain of 78 million. The basis is not a hunch: more than a thousand employers representing 14 million workers across 55 economies.
On the hiring side, Principal's Well-Being Index, run between 22 June and 13 July 2026 with a thousand companies of 2 to 10,000 employees, shows 91% maintaining (44%) or increasing (47%) headcount over the prior three months, and 72% of small and mid-sized businesses not expecting to cut staff next year.
And Stanford's Digital Economy Lab, the hardest source in this piece, puts it plainly: there is no widespread, economy-wide job displacement associated with AI.
The detail: the damage is concentrated, and it is getting worse
It is the same Stanford study — by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, using ADP payroll data from November 2022 to June 2026 — that carries the uncomfortable number. Among workers aged 22 to 25 in highly AI-exposed occupations, employment sits roughly 19% below where it would be had it kept pace with similarly aged peers in less-exposed occupations.
Two qualifiers matter more than the number:
The gap is widening. It was 15% in July 2025 and reached 19% by June 2026. Eleven months, four points.
It does not reach older workers. Experienced workers in the same occupations show no comparable gap. In other words: the occupation is not being eliminated — its entry rung is.
And there is a criterion for which roles fall into that bucket: those that rely on codified knowledge — formal, standardised, documented. Where AI automates rather than complements, employment falls. Where it complements, it does not.
The other side: half-adopting does not count
Here is the finding that most changes a business owner's decision, and it is almost always reported wrong. The line going around is that "firms that adopted AI grew employment by 10%". That is not what the study says.
Stanford's SIEPR work matched spending data from more than 21,000 US firms (from Ramp) with workforce data from Revelio Labs. The result splits in two:
High-intensity adopters — averaging $33.67 per employee per month in their first three months — grew headcount 10.2% over the following two years. And, against the dominant narrative, entry-level headcount at those firms grew 12%.
Low-intensity adopters — averaging $2.78 per employee per month — showed no statistically significant change in employment. Not up, not down. Nothing.
📌 Translated for whoever signs off the budget: a ten-dollar subscription handed to the team is not AI adoption, it is a line of expense. The gap between the two groups in the study is roughly twelvefold in spend per person — and that, not the decision to "use AI", is what separates the firms that grew from the ones that stood still.
Why both sides are the same story
Put the three pieces together and the shape appears: at firms that adopted intensively, employment grew right down to the base; in codified-knowledge occupations, the entry rung shrank; and in the aggregate, one offsets the other and the total barely moves.
It is not that there are fewer jobs. The job moved — away from where the value was knowing the rule, and toward where the value is deciding with the rule in hand. The same World Economic Forum report estimates that 39% of current skills will be obsolete within five years. It is not the job that expires. It is the skill.
What to do about it, without the drama
If you decide for the company: choose between adopting properly and not adopting. The middle ground is the only scenario the study shows as having no effect — you pay the licence, train nobody, change no process, and harvest exactly what the $2.78-per-head group measured. If you are going to do it, concentrate on one process and take it all the way, instead of handing out access.
If you hire people at the start of their careers: the rung that vanished is the codified-knowledge one — summarising, transcribing, reviewing against a template, looking up precedent. It is worth rewriting the role around what remains: judgement, client contact, verifying what the machine produced. The firms in the study that grew at the base did exactly that.
If you sell professional services: separate, in what you deliver, the codified part from the judgement part. The first becomes a commodity over the next few years, with or without your participation. The second is what holds the price — and it is the one that needs to be visible in the proposal, on the website and in the conversation.
If your career is already at the point where the market seems to have no place for you: that is a different reading, and we wrote about it in career in decline. The data here helps on one specific point: the gap does not reach the experienced. The problem, in that case, is almost never age — it is how much of your work is codified knowledge.
What this piece does not claim
Three caveats, because certainty sells cheap on this subject. The World Economic Forum projection is from January 2025 and it is a projection, not a measurement — 2030 may not look like it. Both Stanford studies look at the United States; other markets have a different sector mix and a different pace of adoption, so the numbers indicate direction, not magnitude. And correlation is not causation: firms that spend more on AI per employee were probably already different in several ways, and part of their growth may come from that.
What survives all three caveats is solid enough: no apocalypse in the aggregate, real and growing damage in a specific slice, and gains conditional on intensity — not on intention.
If your company sits exactly in the middle ground the study shows to be useless — tool paid for, process unchanged — that is what our AI implementation for business is for: pick one process, measure the before, and carry it through to a result. And if the question is still whether AI replaces the team, the IKEA and Meta cases are in another piece of ours.
Frequently asked questions
Is artificial intelligence destroying jobs?
In the aggregate, no. Stanford's Digital Economy Lab states explicitly that there is no widespread, economy-wide job displacement associated with AI, and Principal's Well-Being Index found 91% of companies maintaining or growing headcount in the three months before the survey. What does exist is concentrated damage: among workers aged 22 to 25 in highly exposed occupations, employment is roughly 19% below where it would be expected.
Why are young workers the most affected?
Because the entry rung of most professions is made of codified knowledge — formal, standardised, documented: summarising documents, transcribing, reviewing against a template, looking up precedent. That is exactly where AI automates instead of complementing. More experienced workers in the same occupations show no comparable gap, because the judgement half of the work is still theirs.
Does adopting AI make a company's headcount grow?
Only with intensity. In Stanford's SIEPR study, firms spending an average of $33.67 per employee per month in their first three months grew headcount 10.2% over the following two years, with entry-level roles growing 12%. Firms spending around $2.78 per employee per month showed no statistically significant change at all. Half-adoption does not show up in the result.
How much does a company need to spend on AI to see an effect?
The study sets no universal minimum, and it would be wrong to read $33.67 per employee per month as a budget target. What it shows is the order of magnitude of the difference: roughly twelvefold between the group that grew and the group that stood still. The useful signal is about proportion, not price — and concentrating on one process beats handing access to everyone.
Which skills become obsolete with AI?
The World Economic Forum's Future of Jobs Report estimates that 39% of current skills will be obsolete within five years. The practical criterion that emerges from the Stanford data is the type of knowledge: what is codified — rules, patterns, documented procedure — loses value fastest; what depends on judgement under uncertainty, human contact and verifying what the machine produced does not.
Do these numbers apply outside the United States?
With caveats. Both Stanford studies and the Principal survey look at the United States, and other markets have a different sector mix and pace of adoption. The World Economic Forum projection is global, but it is a projection made in January 2025. The honest use of this data is as a direction — whoever sells codified knowledge loses advantage, whoever adopts intensively gains — rather than as a magnitude forecast.