90,000 AI layoffs, 10% headcount growth, same year: the bottleneck moved

Through May 2026, companies pinned close to 90,000 layoffs on AI. The same month, a study of nearly 22,000 firms found the ones spending most on AI grew headcount 10.2%, and grew their entry-level ranks 12%.

Both numbers are real. Both are from this year. They do not contradict each other.

The story everyone is telling is “AI takes the work.” The story the data is telling is quieter and more useful: AI took the easy part of the work, and made the hard part more expensive. The bottleneck did not disappear. It moved. And almost every headline this fortnight is the same relocation, filmed from a different seat.

The tool got cheap. The judgment got dear.

Start with the Ramp and Revelio Labs numbers, because they break the doom narrative cleanly. The companies that committed to AI, roughly $30 per employee per month in the first three months, added headcount across engineering, sales, finance, marketing. The dabblers, the ones who bought a few seats and ran a pilot, added nothing.

The tool was available to both but only one group got the outcome.

That is the pattern. When a capability becomes universal, it stops being the thing that separates winners from losers. The separator moves up the stack, to the thing that is still scarce. Right now that thing is judgment: knowing what to point the tool at, and having the taste to tell a good output from a confident wrong one.

PwC put a price on it. Its 2026 Global AI Jobs Barometer, one billion job ads, found the wage premium for AI skills hit 62%, up from 57% a year earlier. Jobs that need those skills are growing roughly eight times faster than the market. The tool is free. The person who can drive it is not.

Even the biggest buyer can’t buy the outcome

If money bought the result, Meta would have it. It is spending up to $145 billion on AI infrastructure this year.

On July 2, Mark Zuckerberg told staff that agent progress “hasn’t really accelerated in the way that we expected” over the previous four months, and that some of the reorganisation bets “haven’t come to fruition yet.” This from the company that cut 8,000 roles partly on the assumption the agents would arrive on schedule.

Read that next to the enterprise hype and it lands hard. The constraint on getting value out of AI was never access to the model. Meta has all the access there is. The constraint is the messy, human work of redesigning how the thing actually operates, and that work does not speed up just because you spent more.

Watch what happens when you automate the wrong seam

Here is the trap, in one hiring statistic.

Greenhouse surveyed active job seekers: 63% have now been interviewed by an AI, 38% have walked away from a process because it used one, and 70% were never told AI would be evaluating them. Companies automated the top of the funnel to save recruiter hours. The saving was real. The outcome, filling the role with a good hire, got worse, because four in ten of the people you wanted quietly closed the tab.

That is what automating the wrong seam looks like. You optimise the cost you can see and pay the cost you cannot. The task got cheaper. The result got worse. Nobody put that trade on a slide.

Meanwhile the roles stay open. ManpowerGroup’s 2026 survey, 39,000 employers across 41 countries, has 72% unable to find the talent they need. In Europe it is sharper: Germany 83%, France 74%, the UK 73%. And for the first time, AI skills are the single hardest capability to hire for, ahead of engineering and IT. The scarce thing is not labour. It is a specific human capability the market has not grown yet, the same capability carrying that 62% premium. Two reports, one seam.

The cost of starting collapsed. The cost of finishing didn’t.

The same split runs through who is building.

Solo founding is at an all-time high. Carta’s data shows the share of new startups with a single founder rose from 23.7% in 2019 to 36.3% by mid-2025. AI made starting a company a weekend project. But only 0.2% of solo businesses ever cross $1 million in revenue, and 78% never clear $50,000. The on-ramp got wider. The finish line did not move an inch closer.

And at the top of the market, capital is concentrating into the few teams that can convert it. Sifted counted eight European startups closing rounds above $1 billion in the first half of 2026, an all-time record, overwhelmingly AI-native. Money is not the scarce input. The team with the judgment to turn it into a company is.

Even your buyers now run on this logic. Forrester’s survey of 18,000 business buyers found 94% used AI during their most recent purchase, and AI answer engines have overtaken your website and your sales team as the number one place they research you. The tool did their homework. Whether you get shortlisted still depends on something the tool cannot manufacture on your behalf: whether you were worth citing.

So what do you actually do with this

Stop buying tools and calling it a strategy. The dabblers in the Ramp data bought the same software as the winners and got nothing, because a tool laid over a broken process just runs the breakage faster.

Three moves for the operator this quarter:

1. Find the seam, not the task. Ask where the expensive-but-invisible cost is (the candidates you lose, the buyers who never call), and protect it before you automate the cheap visible one. Aim for the real “Seamless”.

2. Redesign the workflow before you buy the seat. The companies getting returns rebuilt the process around AI. The ones bolting it on are the 0.2%, not the record round.

3. Grow the judgment, in yourself and your juniors. The scarce, 62%-premium capability is the one thing you can compound in-house. AI removed the routine work that used to train your apprentices. If you do not replace that training on purpose, you are starving the exact skill the market is paying most for.

None of this is a story about machines taking over. It is a story about where the difficulty went. It went up, into the human layer, and it got more valuable on the way.

AI did not make the work disappear. It made the easy part free and sent the bill for the hard part to whoever still knows how to do it.

If this was useful, subscribe and the next field note lands in your inbox. Reply and tell me which seam AI moved in your world.