Kenneth Arrow won the Nobel Prize in Economics in 1972 for a body of work that included a 1962 paper on how firms actually learn. His insight was simple and important: learning cannot happen in the abstract. It "can only take place through the attempt to solve a problem and therefore only takes place during activity."
You learn to write contracts by writing contracts. You learn to close clients by closing them. You learn to manage a P&L by managing one, with someone more experienced watching over your shoulder.
Arrow called this "learning by doing," and he argued it was the fundamental engine of productivity growth, not just for individual workers, but for entire firms and industries.
Sixty years later, Federal Reserve researchers at the Atlanta Fed are pulling that 1962 paper off the shelf. Their finding: companies racing to automate entry-level work may be trading a short-term cost savings for a long-term leadership deficit that won't show up on any balance sheet until 2031.
If you run talent at a company that has been cutting junior roles, this deserves your attention.
What's Actually Happening to Entry-Level Work
The numbers are clear and getting worse. Job postings on Handshake, the leading platform for early-career roles, are down 2% year-over-year and 12% below pre-pandemic levels. The unemployment rate for college graduates aged 22 to 27 is 5.6%, higher than the overall unemployment rate of 4.3%. Nearly 9 in 10 members of the Class of 2026 say they're worried AI or automation will replace entry-level roles, up sharply from 64% the year before.
The industries hit hardest are the ones most amenable to AI-assisted output: marketing, legal, accounting, human resources, and IT. These are precisely the fields where entry-level work has historically been high-volume and lower-complexity. That's exactly the problem. That "lower-complexity" work is not just cheap labor. It's curriculum.
The Apprenticeship Nobody Called an Apprenticeship
MIT researcher Andrew McAfee, who co-leads the school's Initiative on the Digital Economy, put it plainly in a May 1 Fortune interview: "How else are people going to learn to do the job except via on-the-job learning and training apprenticeship? That's how you learn to do difficult knowledge work: by helping somebody who's good at that with the routine stuff."
This is how every functional domain produces its senior talent. Every CFO spent time doing reconciliations and basic financial modeling. Every general counsel wrote the memos that got marked up by a partner. Every VP of Marketing ran campaigns that a director tore apart and made her redo. Every senior HR business partner handled benefits queries before anyone let them near an executive coaching conversation.
The "routine stuff" is not just work to be done. It is the mechanism by which junior employees develop judgment. Remove the routine work, hand it to AI, and you do not get faster, smarter junior employees. You get junior employees who never do the repetitions that build expertise.
The Spillover Problem the Fed Is Warning About
The Atlanta Fed researchers' most important warning is not about individual companies. It's about the industry-wide effect.
Arrow's theory holds that experience-based learning does not stay confined to one firm. When an entire industry reduces the entry-level work available to junior workers, because every firm is making similar AI adoption decisions at roughly the same time, the talent deficit becomes systemic. In five years, there are fewer experienced mid-level managers in the market. In ten years, fewer senior leaders.
And you can't simply poach your way out of it. The pool is smaller for everyone.
The Fed researchers proposed two policy responses: a tax on automation-derived profits, paired with subsidies for firms that preserve entry-level task work. Neither is likely to move quickly. Government is slow. But the talent shortage they're predicting will arrive on schedule regardless.
The Industries Most Exposed
Look at where AI is absorbing entry-level work fastest, and you see the long-term risk most clearly.
Accounting and finance. Automated reconciliation, document processing, and basic financial modeling have allowed Big Four firms to quietly reduce junior associate headcounts over the last two years. The pipeline that produces today's CFOs ran through those roles.
Legal. Contract review, discovery work, and legal research are being absorbed by AI. Junior associates build judgment by doing these tasks: learning what matters in a contract, what language signals risk, what fact patterns change outcomes. If AI does the reading, the associate does not develop the reading skill.
Marketing. Junior marketers who used to run the analysis (A/B tests, campaign attribution, audience segmentation) are increasingly watching dashboards built by AI. The cognitive work of "what does this data mean for our strategy" is moving upstream faster than the people assigned to answer it can keep up.
IT and software. AI-assisted code generation reduces the volume of greenfield work available to junior developers. For developers who needed repetition to build foundational skills, the repetition is being automated away.
What Smart Talent Leaders Are Doing
The correct response is not to resist AI in workflows. That is neither realistic nor competitively viable. The correct response is to redesign entry-level roles so that AI is a tool junior employees use, not a substitute for the problems they need to solve themselves.
Several firms are already doing this deliberately.
Instead of routing client communications through AI entirely, keep the junior account manager as the primary contact with AI as a drafting tool. They write the first email. AI makes it better. They send it. They still learn the relationship.
Instead of having AI do the full contract review, build workflows where junior associates audit the AI's output and escalate items for senior review. The cognitive work of "what matters here" still happens in the junior employee's brain, which is where it needs to happen if that employee is going to become a senior one.
Some firms are formalizing what used to be informal: explicit apprenticeship structures, defined learning objectives tied to specific experiences, and deliberate preservation of work that builds judgment even where AI could handle it.
This is not charity toward young workers. It is talent infrastructure investment. The companies that build it now will have materially stronger managers in 2031. The companies that don't will be trying to hire those managers from each other, at prices that make today's AI/ML talent market look reasonable.
Two Things Worth Doing This Quarter
If you run talent at any organization that has been reducing entry-level headcount, two actions are worth taking now.
First, audit which entry-level roles have been eliminated or reduced in the past two years. Ask what skills those employees were developing, and what happens to your senior talent pipeline if those skills are no longer being built anywhere in your organization.
Second, if you are hiring junior talent right now, understand that recent cohorts have had fewer on-the-job repetitions than their predecessors. They need more structured development, not less. The firms that build explicit learning frameworks for junior hires (not just onboarding, but ongoing skill development tied to increasingly complex work) will come out ahead.
The pipeline problem is not visible yet. But the economics are not complicated. If you stop training people at the bottom, you run out of experienced people at the top. Arrow knew it in 1962. The Fed is reminding us of it now.
If you're building early-career pipelines or restructuring your junior hiring strategy, BlueLine can help you find and match candidates efficiently.