Sixty years ago, an economist named Kenneth Arrow sat down and worked out something that seemed almost too obvious to say: workers get better at their jobs by doing them. The insight was simple, but Arrow, who would go on to win the Nobel Prize, formalized it into a theory with sweeping implications. Learning, he wrote, “can only take place through the attempt to solve a problem and therefore only takes place during activity.” Experience wasn’t just good for workers, he argued—it was the engine of productivity growth for firms and, ultimately, the entire economy.

Now, as artificial intelligence chips away at the entry-level jobs that once served as the on-ramp to white-collar careers, researchers at the Federal Reserve Bank of Atlanta are dusting off Arrow’s 1962 paper and warning that companies racing to automate their way to lower payroll costs may be sawing off the branch they’re sitting on.

The unemployment rate for young degree-holders is now consistently higher than overall unemployment, a reversal from recent labor trends that many blame on AI replacing entry-level knowledge work. Some segments of college graduates are now grappling with unemployment at a similar rate as peers without a degree, suggesting a college education might become harder to justify, and the appeal of a secured posting in an office job could be losing its shine.

But take away enough entry-level jobs, and those white-collar employers will start hurting too. That’s the conclusion of a paper published last week by researchers at the Federal Reserve Bank of Atlanta, analyzing the tradeoffs on both sides of the managerial aisle of automating junior office jobs.

Arrow argued that innovation and productivity growth were byproducts of experience and practice. The Fed researchers applied this framework to the drudgeries of entry-level work, arguing that the experience is foundational to building up expertise required for senior roles. Crucially, the type of repetitive activity and skill-building that happens early in a young person’s career cannot be replicated in college or grad school, with entry-level roles effectively becoming a specialized crash course to prepare workers and ensure a firm’s institutional knowledge remains intact.

“The tasks that fill entry-level positions are not merely low-value work—they are the curriculum through which workers accumulate the human capital that makes them productive later in their careers,” the researchers wrote.

By automating more of these job roles, firms risk eviscerating the pipeline of competent senior workers they might need in future, trading short-term cost savings in the present for long-term stability. And because Arrow’s theory argues that experience-based learning and productivity growth spills over and ripples throughout the economy rather than staying confined to one firm, a single company’s choice to automate an entry-level task or role will eventually impact the rest of the industry as well.

There are likely multiple reasons for the difficult job market for entry-level roles in 2026, and not all have to do with AI. Businesses in general have slowed hiring in response to global uncertainty, the war in Iran, tariffs, and yes, in some cases to experiment with AI. Many white-collar industries overhired after the pandemic and are now trimming staff. The reality of too few white-collar jobs available and many graduates competing for spots means the market has become saturated, part of the reason why a growing number of Gen Z Americans are considering careers in skilled trades instead.

But even if young Americans’ woes cannot be exclusively pinned on AI, the fact remains that many young graduates are either unemployed or underemployed in 2026, missing the crucial learning by doing experiences Arrow argued were central to their professional development and the economy’s productivity.

The Fed researchers proposed two policies incentivizing firms to keep employing young workers while making the most of AI: a tax on automation-derived profits, accompanied by subsidies rewarding companies that expand the amount of tasks entry-level workers are needed to accomplish. This mix would discourage full automation and support the creation of new work that allows young workers to learn their trade.

The long run alternative would be a smaller cohort of “low-quality managers” who will be less capable at driving innovation. In the shorter term, however, company profits will likely go untouched, given the cost savings of using AI. If employers choose to automate more entry-level tasks, the authors write, “the welfare cost of coordinating on low learning falls almost entirely on workers.”