AI Layoffs Barely Show Up in the Data. Entry-Level Hiring Freezes Do.

Every few weeks a company announces layoffs and points at AI as the reason. The Federal Reserve looked for that effect in the actual hiring data and could not find it, at least not in the aggregate way the headlines describe.

A Federal Reserve Board study published in March 2026 by economists Jessica Liu and Douglas Webber combined Lightcast job-postings data, drawn from more than 65,000 sources, with the Census Bureau's Business Trends and Outlook Survey of roughly 1.2 million firms, covering September 2023 through November 2025. They found no evidence that firms or industries adopting AI more heavily are posting fewer jobs. Only about 10 percent of surveyed firms report current AI use, 5.5 percent of firms post any AI-related roles at all, and the coefficients linking AI adoption to reduced postings were small and mostly not statistically significant, even a year out.

That is a real finding, not a hedge, and it should temper the instinct to treat every AI-attributed layoff as proof of a trend. But it does not mean AI is doing nothing to the labor market. It means the actual effect is narrower and more specific than "AI is taking jobs," and the businesses getting hurt by it are not the ones you would expect from the headlines.

Here is the narrower effect, and it is well documented. Stanford's Digital Economy Lab has been tracking ADP payroll data since ChatGPT launched, and its "Canaries in the Coal Mine" research, led by Erik Brynjolfsson's team, finds that workers ages 22 to 25 in the most AI-exposed occupations now have employment about 19 percent below where it would sit if it had tracked similarly aged workers in less-exposed fields. That gap was 15 percent as of July 2025 and has widened every reading since. The mechanism is not firing. It is reduced hiring of the youngest, least experienced workers into roles where AI automates rather than complements the work. Where AI instead augments experienced people, employment holds flat or grows.

Put the two studies together and a specific picture appears: AI is not shrinking headcount across the economy, it is shrinking the bottom rung of certain career ladders while leaving everything above it roughly intact. The New York Fed's own tracking backs this up from a different angle. Recent college graduate unemployment sat around 5.6 percent in the second quarter of 2026, not a crisis number, but underemployment for that group, graduates working jobs that do not require a degree, edged up to 42 percent. Graduates are not mostly unemployed. Many are employed below where their degree should place them, in exactly the routine, task-heavy work AI now handles.

The steelman for cutting anyway is fair. Margin pressure is real, boards reward fast cost action over a multi-quarter retraining plan with uncertain payoff, and not every displaced worker's skills map onto whatever higher-value work is left. Blaming AI for a cut you were already going to make is bad honesty, but the cut itself can still be the correct call for that company's finances. Not every layoff is a mistake.

What I am arguing is that the Fed and Stanford data together should change what a business owner watches. The Fed's finding means if you are attributing a broad, company-wide cut to AI, the aggregate data does not back you up, so look harder at your real reason. Stanford's finding means the actual risk for a smaller company is quieter than a layoff: you stop backfilling the one entry-level seat where a model can now do most of the task list, per the Anthropic Economic Index, which puts average current AI task coverage around 35 to 40 percent of a job and shows usage concentrated in computer, administrative, and sales work. Do that twice and you have not saved a salary, you have canceled your own future senior hire, because that senior person was supposed to come from the junior seat you froze.

The fix is to redefine the role around the residual, not eliminate it. Keep the junior hire, hand them the 60 to 65 percent of the task list a model cannot yet do well, and add the higher-value work you previously could not afford to staff. Some employers are trying this at scale: Amazon, Anthropic, Microsoft, and the OpenAI Foundation have committed roughly $500 million toward a $1 billion goal for a retraining initiative called RAISE US, an effort worth noting even though the backers are also the companies profiting from the adoption creating the problem. You do not need $500 million to apply the same logic at your own scale. You need a task list, not a job title, and the discipline to keep the person once you have one.

That task-level mapping, not a headcount number, is the actual output of a good AI readiness review. If you are deciding whether to freeze a hire or cut a role and calling it an AI decision, get the task map first, then decide. Reach out if you want a second set of eyes on that map before you make the call.