What the AI layoff numbers measure
The number you have seen quoted is a count of announcements, not a count of work that moved. Somewhere between your last board meeting and this one, someone put a slide up with the layoff figures attributed to AI and used them as proof that the shift has already happened. It has not happened in the way that slide suggests.
I keep meeting directors who treat those figures as a market signal and plan their own headcount against them. That is planning against other people's press releases. The evidence underneath tells a slower and more awkward story, and it changes what you should do this quarter.
The number everyone quotes
Challenger counted layoffs attributed to AI and the curve is steep. The series runs as follows:
- roughly 18,000 in 2024
- more than 100,000 in 2025
- more than 150,000 in the first six months of 2026
AI came up as the most cited reason for four months running, and in one of those months it accounted for around 40 percent of all American layoffs. That is what companies said. It is a record of stated reasons, collected from public statements, and it travels fast because it is easy to repeat.
What the service leaders reported
Gartner asked 321 customer service leaders whether AI had actually shrunk their workforce, and only about 20 percent said yes. These are the leaders closest to the use case with the clearest business case. Chat deflection, call summarisation, draft responses. The function where automation has the longest track record is the function where four out of five leaders report their headcount intact.
So the announcements and the operations are describing two different years.
A stated reason is a communication decision. A reduced team is an operational fact.
What executives say when the microphone is off
The National Bureau of Economic Research found that roughly nine in ten executives said privately that AI had done nothing to employment in their organisation. Nine in ten. The same population that signs off on the public language about efficiency gains.
That gap is the paradox in the boardroom. You are asked to invest aggressively in AI and to cut costs now, and the only honest way to hold both is to build a capability you do not yet have. Attributing a reduction to AI before the capability exists borrows credibility you will have to repay.
Klarna went in both directions
Klarna's assistant genuinely did the work of 700 agents and handled two thirds of conversations within weeks. The workforce went from around 5,500 to 3,400. Then the company moved back to a hybrid model with people in the loop.
Read that sequence twice. The technology performed. The operating model around it had to be rebuilt afterwards, with service quality as the reason. Klarna is the strongest case available, and even there the end state is people and machines together.
What the spending tells you
Block attributed around four thousand roles to intelligence tools and the share price rose 24 percent that day. Markets reward the story quickly, which is precisely why the story is told.
The same set of companies is heading towards seven hundred billion dollars in AI infrastructure spending in 2026. Organisations spending at that scale are buying future capacity. Capacity you buy today produces results after you have redesigned the work, trained the people and set the decision rights.
What you do on Monday
- Ask your service lead one question: how many conversations ended without a person this month, and how many people are on the team today.
- Put the Klarna sequence in front of your executive team and discuss the hybrid model it landed on.
- Pick one journey, measure its current handling time, and set a target for the quarter after this one.
- Name the person who owns AI decision rights for that journey, with a budget and a date.
- Say out loud in the next board meeting which part of your cost plan depends on capability you have already built.
Your next board conversation will go better with an honest picture of where your organisation actually stands. The AI Maturity scan gives you that in one session, and chapter 3 of CX+AI works through the paradox in full. See the materials.