What AI spend does to headcount

Two studies came out this summer about young workers and AI, and they reached opposite answers.

Stanford, using payroll records for millions of Americans, found that employment of 22-to-25-year-olds in the more AI-exposed occupations fell 11% between November 2022 and June 2026. Ramp, the corporate card company, found that the firms spending the most on AI per employee averaged 12% more entry-level staff over the two years after they started paying for it than similar firms that adopted later.

So how can both be true? If you're planning your team or trying to understand what AI should do to headcount, these two studies point you in very different directions.

One way to estimate AI’s impact on a job is to look at the tasks inside it. A junior salesperson might spend the week writing personalized outreach and updating the CRM when leads respond. AI can already do both. If enough of those tasks can be automated, the job itself starts to look like it should shrink.

The benchmarks make that argument tempting. On OpenAI’s GDPval, built from tasks supplied by professionals with an average of 14 years of experience, GPT-5.2 Thinking matched or beat expert work in 70.9% of blind comparisons. (But these were well-defined assignments, not complete jobs.)

AI is taking on longer work too. METR estimates that the length of a task AI can complete on its own has doubled about every four months since 2023, from four minutes with GPT-4 to at least 16 hours of skilled work this spring.

Yet U.S. unemployment is 4.1%, and AI's measured effect on it is only 0.1 to 0.2 percentage points. How can AI beat people at so many tasks without causing much more unemployment? Because a job is more than the sum of its tasks. What's more intangible is the part between them: deciding which task matters, in what order, and when to stop; coordinating with people on other teams; and being the one accountable when it goes wrong.

You can put a number on it. Handing a task to AI always costs you the time to explain it and check the result, and when the AI fails, you do the task yourself anyway. So delegating pays only when the AI's success rate beats the share of the task you spend explaining and checking it. On a 30-minute task that takes 5 minutes to explain and check, AI needs to succeed 1 out of 6 times to break even. If explaining and checking take 20 minutes, it needs to succeed 2 out of 3 times. The more work it takes to hand something off, the better the AI has to be.

Junior-level work is often easier to specify. A junior is usually given a defined task and told what good output looks like, which makes the work easier to hand off to AI. Senior-level work is harder to spell out. An expert is given a goal and has to judge which tasks matter, which to skip, and sometimes whether the goal itself is wrong. That may explain why Stanford's decline is concentrated among 22-to-25-year-olds, while experienced workers in the same occupations haven't lost ground.

Why is that judgment so hard to hand over? In 1966, the philosopher Michael Polanyi wrote that "we can know more than we can tell." Ask a writer for their routine, and they'll tell you. Ask how they find the right words, and they can't. The part between tasks is mostly this kind of knowledge, which is why it's so hard to hand over. The irony is that the person you're trying to replace holds the knowledge to build the replacement.

Which brings us back to the two studies. For anyone running or funding a company, that gap raises three questions:

What happens to headcount when a company buys AI? Ramp's heaviest adopters grew faster after adoption. But a 30-person startup adopting coding agents is a different company from a Fortune 500 firm adopting AI, and so far only one of them shows up clearly in the data.
Should you stop hiring juniors? IBM is tripling U.S. entry-level hiring this year, betting that cutting juniors now means a shortage of senior staff later. Software job postings have recovered, but mostly for senior roles.
Where is the growth coming from? If AI makes people more productive, adopting is how you grow. One of the Stanford authors suspects adopters are instead taking customers from firms that don't adopt. If he's right, non-adopters risk losing market share, and their workers are exposed only through that.

Stanford and Ramp are only two pieces of the picture. Our team at Social Capital examined six datasets, company hiring behavior, wages, layoffs, and AI adoption to see where the AI pressure is actually showing up in the labor market. Read the full Deep Dive below for the complete analysis.