AI Layoffs: the cuts came before the proof

In March 2026, 750 CFOs told researchers at NBER, Duke and the Federal Reserve that AI would drive roughly 502,000 job cuts in 2026, nine times the ~55,000 blamed on it in 2025. Then the same survey asked a second question: has AI changed your own firm’s employment or output over the last three years. Nine in ten said no.

That is not a contradiction. It is the whole story of AI in the middle of 2026. The decisions are running well ahead of the evidence, and almost everyone is making them anyway.

3 different data sets, same pattern. The story is being told before the receipts exist.

The layoffs are a forecast, not a ledger

Most of the 502,000 cuts NBER’s researchers describe are anticipatory: driven by what CFOs expect AI to do to their headcount, not by a system already doing the job. The number is a bet, dressed up as a line item.

Gartner is making the same kind of bet from the other side: over 40% of agentic AI projects will be cancelled by the end of 2027, on cost, unclear value, or risk controls nobody built in time. Of the thousands of vendors calling themselves “agentic AI”, Gartner reckons about 130 are real.

Even the labour market’s own data does not agree with the labour market’s own headlines. LinkedIn’s economic research puts hiring down 20% since 2022, and attributes the slowdown to interest rates, not AI displacement, even as recruiters plan to lean on AI harder than ever this year.

Text within this block will maintain its original spacing when published

3 different data sets, same pattern. The story is being told before the receipts exist.

Capital cannot decide which story to believe either

If executives are placing a bet, investors are placing the same one at a larger denomination. European VC hit $17.6 billion in the first quarter of 2026, up almost 30% year on year, with AI claiming more than half the region’s funding for the first time.

Then June happened. Robotics briefly out-raised AI for the month, taking 15.6% of Europe’s €8.3 billion in total funding, largely on the back of a single €1.3 billion round into Germany’s NEURA Robotics. AI won the quarter (and likely the year). Robotics won the month. Nobody actually changed their mind about which technology matters. The money is just as unsettled as the CFOs are.

The real gap is not access, it is readiness

Here is the part that should worry operators more than either of those. The hardest skill to hire for in Europe is no longer engineering. It is AI. ManpowerGroup’s 2026 survey of 39,000 employers across 41 countries found AI model and application development (20%) and AI literacy (19%) now sit above every traditional technical skill on the hardest to fill list. Slovakia, Greece and Germany feel it hardest.

Microsoft’s own Work Trend Index found the same seam from the inside: culture, manager support and talent practice explain roughly twice as much of an organisation’s AI payoff as individual tool use does, a 67%/32% split. The tool was never the bottleneck. The organisation wrapped around it was.

Marketing shows the same shape from a different angle. Adoption is close to universal, and the share of marketers who can actually prove AI’s return on investment fell from 49% to 41% this year, even as budgets kept climbing. More people are using it. Fewer people can show what it did.

I see this seam constantly through JobMentis. Companies will buy the screening tool, the sourcing agent, the interview copilot, in the same quarter they cannot describe who owns the process those tools now sit inside. The purchase order clears in a day. The readiness behind it takes a year nobody budgeted for.

One person closed the gap without waiting for the org chart

Then there is Matthew Gallagher, who built a tele-health company called Medvi alone, on $20,000, using ChatGPT, Claude and Grok to write the code, the copy, and handle a lot of the customer service. No round of hiring. No change management deck. First full year: $401 million in sales, 250,000 customers, a 16.2% net margin. He is on pace for $1.8 billion this year.

Gallagher did not wait for his organisation to become AI-ready, because he did not have one to wait on. He was the org chart, and the org chart moved at the speed of one person’s judgement instead of a committee’s.

That is the actual difference between the two halves of this story. Everywhere else, AI is a decision being made about an organisation from the outside: a CFO forecasting cuts, a fund choosing a sector, a maturity model measuring a gap. In the one place it visibly worked, someone stopped treating it as a decision to route through the org chart and started treating it as their own job to redesign.

What this means before your next planning cycle

You do not need Gallagher’s market or his risk tolerance to take his actual lesson. Before your team signs off on another AI tool purchase, ask who in the room is prepared to own the redesigned workflow personally, the way Gallagher owned his. If the honest answer is nobody, you have found your real gap, and it was never the model.

That question is uncomfortable on purpose. Committees are good at approving tools and bad at owning outcomes, and AI punishes that split faster than any previous technology did, because the tool now makes decisions the committee used to get days to review. Somebody has to be willing to answer for the workflow, not just sign off on the subscription.

AI did not cut those jobs. The forecast did. The people already capturing the upside did not wait for either one to be proven right.