173,568 jobs cut in AI's name. Not one audited pound of proof.

Since 2023, employers in the United States have announced 173,568 job cuts and named artificial intelligence as the reason. Challenger, Gray and Christmas have counted them every month since firms started saying it out loud.

Here is what nobody has produced. A single audited pound of proof that it worked.

What actually happened

The American numbers are steep. AI was cited in 54,836 announced cuts across 2025, then 101,743 in the first half of 2026 alone. In May 2026 it accounted for 40% of every cut announced that month, the first time it topped the list, and it has led every month since.

The largest single reduction was Oracle's, at roughly 30,000 roles. Amazon cut 14,000 corporate jobs in October 2025 and another 16,000 in January 2026, about 9% of its corporate workforce in three months. Meta announced 8,000. Salesforce took its support team from 9,000 to 5,000.

Now the correction that almost nobody prints. Total US announced cuts through July 2026 were 477,033, down 41% on the same period in 2025. The wave is not bigger. It is more AI-attributed. Anything you read implying otherwise is wrong.

Britain did something different, and this is the part that matters if you work here. UK job postings fell 11% in the first half of 2026 and sit 32% below their February 2020 level, a considerably worse recovery than the euro area or the US. Unemployment reached 5.2%. BT has said it will shed up to 55,000 roles by 2030, roughly 10,000 of them attributed to AI. HSBC has up to 20,000 under review. Standard Chartered around 7,800.

One honest gap: there is no British equivalent of the Challenger series. Nobody counts the stated reasons for UK redundancies. We cannot answer this question for Britain because the data has never been collected.

What they said

They were not coy. Marc Benioff, explaining the Salesforce cuts: "I need less heads." Oracle put it in a regulatory filing, stating that AI adoption had resulted, and might continue to result, in workforce reductions. A filing is a stronger signal than a podcast, and should be weighted as such.

The British version arrived in May 2026, when Standard Chartered's chief executive described replacing what he called lower value human capital with machines. He later apologised for the phrase. He did not withdraw the plan.

All of it traces back to one announcement. In February 2024 Klarna said an AI assistant was doing the work of 700 customer service agents, with a projected $40m benefit. That claim was quoted in hundreds of board presentations. Both of its headline numbers were softer than they sounded: the $40m was a forecast, and the 700 described a volume of work, not people dismissed.

Why they really did it

Three things were happening, and only one of them was artificial intelligence.

First, the capital expenditure had to be paid for. Microsoft's 2026 capex runs around $190bn. Amazon around $200bn. Meta between $125bn and $145bn. Payroll is the only cost line that can be cut fast enough to partly offset a build-out on that scale. JPMorgan's 2026 outlook names the resulting decoupling of corporate spending from hiring, and treats the fall in labour demand as a warning signal.

Second, the market paid for the announcement. Cisco announced 4,000 job cuts and its shares rose about 13%. The reward arrives on the day. It does not wait to see whether the AI works, and there is no mechanism to take it back if it does not.

Third, a lot of this was ordinary over-hiring, relabelled. Oxford Economics, in January 2026, found that firms "don't appear to be replacing workers with AI on a significant scale" and suspected some were dressing up layoffs as good news rather than admitting to past over-hiring. Yale Budget Lab found no significant change in the occupational mix for AI-exposed jobs. Sam Altman has conceded there is some AI washing.

Payroll was converted into compute, and AI was the word used to describe the conversion, because that word moved the share price.Oxford Economics, Jan 2026 · Cisco share move via Fortune

The British version is plainer still, because no UK firm is building $200bn of data centres. There was no capex to offset. What there was: rising employment costs, weak demand, and a labour market where the cheapest lever is simply not to hire. AI supplied a more flattering vocabulary for that.

And the consultancies?

The honest charge is not fraud, and I am not going to imply it. The charge is structural, and worse, because nobody will be sanctioned for it.

They were paid on bookings, not outcomes. Accenture reported around $5.9bn in new generative AI bookings in FY2025 against $80.6bn in total new bookings, and IDC projects AI consulting and services past $500bn by 2027. Meanwhile MIT's NANDA study found about 5% of enterprise pilots reached measurable profit-and-loss impact. Those two facts coexist comfortably, because the fee was never contingent on the second one.

They also sold certainty on a question the evidence had already answered differently. Every large firm published agentic AI predictions promising virtual workers and digital workforces. Every one included a human-in-the-loop caveat, in the small print. The caveat was the finding. It was printed as a footnote.

And they ran the same play on their own staff. McKinsey has been reported pulling headcount back toward 40,000. Accenture's chief executive said staff who could not be reskilled on a compressed timeline were being exited. The junior analyst pyramid, the thing generative AI genuinely does compress, is what they cut.

So: an industry was paid several billion pounds to supply confidence, at a moment when the honest deliverable was doubt. Boards did not buy analysis. They bought cover.

What the research already said

In April 2023, Brynjolfsson, Li and Raymond published a field study of 5,179 customer support agents. Average productivity gain from AI access: 14%. For novice agents: 34%. For experienced agents: close to nothing.

That is twenty-two months before the first big year of AI-attributed cuts. The finding that predicts exactly which staff AI helps, and which it does not, was in the public domain before any of these decisions were taken.

Nobody had to guess. They chose not to look.Brynjolfsson, Li & Raymond, April 2023, n=5,179

Everything published since has agreed. Cui and colleagues measured a 26% gain across 4,867 developers, concentrated in the less experienced. A trial in Argentina found an AI assistant closed roughly three quarters of the performance gap between higher and lower education groups. The Swedish MASAI trial randomised 105,934 women to AI-supported mammography: 44% less reading work and 29% more cancers found, with the human read kept in the design.

And the counter-case, which fits the same pattern. METR ran a randomised trial with 16 experienced developers on codebases they had maintained for years. With AI they were 19% slower.

What separates these is not the model. It is how expensive it is to check the answer. A novice agent's reply is checked by a supervisor who already exists. A radiologist's second read was built into the trial. An expert developer has to reconstruct the reasoning before they can trust the code, and reconstruction costs roughly what writing it would have cost.

Which brings us to the single most damning number in this piece, and it is British. The ONS reports that UK AI adoption has risen from about 12% of businesses to about 35% since late 2023. Over the same period, the number of AI tools used by the average adopting business went from 1.4 to 1.6.

British firms cut graduate intake by 42% on the strength of an average of 1.6 tools per company.

Did it work?

No. And unusually, the evidence on this is clean.

Gartner surveyed 350 executives at organisations with $1bn or more in revenue, all of them already running AI agents or automation. About 80% had reduced headcount. Those reductions did not translate into return on investment, and the reduction rate was near identical between the firms reporting strong returns and those reporting weak or negative ones. Helen Poitevin, who led it: "Workforce reductions may create budget room, but they do not create return."

The National Bureau of Economic Research went wider. Nearly 6,000 chief executives and finance directors across the US, UK, Germany and Australia, surveyed with the Bank of England, the Bundesbank and the Atlanta Fed. 89% reported no impact on productivity. More than 90% reported no impact on employment. PwC asked 4,454 chief executives across 95 countries: 56% had seen no revenue or cost benefit, and only 12% had seen both.

Bank of England staff found early signs of AI feeding into UK productivity, but almost entirely in the industries that build AI rather than those using it. Financial services, Britain's heaviest adopter, contributed negatively.

Then the reversals. Klarna rebuilt human support after its chief executive conceded the company had gone too far. Robert Half surveyed more than 2,000 hiring managers: 32% who cut a role for AI later rehired for the same or a similar one, highest in finance at 44%. Gartner expects half of all AI-driven cuts to be reversed by 2027.

Three years and several hundred billion pounds in, there is no published, audited case of a large firm proving that AI-driven headcount reduction improved the bottom line.Gartner, n=350 · NBER WP 34836, n≈6,000 · PwC, n=4,454

Who got hired instead

Not the same people, and not at the same price. Postings for AI engineers grew 654% between the first half of 2024 and the second half of 2025. Robert Half counted 49,200 AI, machine learning and data science postings in 2025, up 163%. AI staff engineers earn 18.7% more than their non-AI equivalents.

In Britain, AI or related tools now appear in a record 9.4% of all job postings, and in data and analytics nearly half of every advert mentions them.

Now put that next to the other British number. Adzuna counted 8,398 graduate job adverts across the entire United Kingdom in May 2026, down 42.1% in a year, the steepest fall it has recorded and lower than at any point during the pandemic. The Institute of Student Employers puts the drop in tech graduate roles at 46% since 2024, with a further 53% expected.

So the market now demands AI skills that entry-level candidates have no way to acquire, because the entry-level jobs where they would have acquired them are the ones that were cut. Stanford's AI Index records employment among software developers aged 22 to 25 falling nearly 20% since 2024, while headcount among their older colleagues grew.

That is not a labour market correcting itself. That is a ladder with the bottom rungs sawn off.

How to do this properly

If you run a smaller business, you are in a better position than any FTSE 100 board was in 2024. Not because the tools improved. Because somebody else has already paid for the experiment, and published the results.

You do not copy what the big players did. You take what they proved. Cutting headcount to demonstrate AI value produced no measurable improvement in returns, across 350 large firms. Nine in ten executives measured no effect at all, across 6,000 firms. The organisations that did see returns were the ones investing in the people around the technology rather than removing them.

The rule that follows is short enough to remember. Do the opposite of the visible thing. The visible thing was announce, cut, then discover. The order that worked is measure, deploy to the least experienced, then, much later and only on evidence, consider structure. Figure 05 sets out the four questions, one automation at a time: is a mistake cheap to spot, is it really saving hours, who runs it day to day, and only then, does anything change. None of them is answered on a forecast.

One last test, borrowed from the wreckage. Klarna's full loop, from announcement to correction to a working hybrid model, took twelve to eighteen months. Gartner expects half of AI-driven cuts to be reversed by 2027. So push your payback model out two years and add the cost of hiring the people back. If it still works, you have a business case. If it does not, you have a press release.