The first job is also a classroom.

The requisition is one page. It names the role, Analyst I, the manager it reports to and a start date in June. Then comes the box that justifies the headcount: reconcile the monthly reports, draft first-pass memos, summarize research and filings, check the figures in client presentations. The requisition in this dispatch is invented, but its shape is familiar to anyone who has approved one. Every duty on it is real work, and every one of them is now work that an AI system can do quickly and, with review, reasonably well. What the form has no box for is the other thing the job does: it turns someone who has read about the work into someone who knows how to do it.

Two weeks ago, on February 24, Scott Davis, an economist at the Federal Reserve Bank of Dallas, published wage evidence that puts this in economic terms. His reading of the data is that AI is substituting for entry-level workers in jobs built on knowledge that can be written down, while complementing experienced workers whose value lies in what they learned by doing. He ends with a warning about the usual way white-collar careers begin, with a junior doing codifiable tasks while slowly absorbing the rest. Firms, he writes, will find that AI makes this method of developing employees "cost-ineffective, at least in the short run."¹

My argument is that an entry-level job has always delivered two things at once. One is production: the reconciliations, drafts and summaries the role is paid to deliver. The other is apprenticeship: the experience that the work builds in the person doing it. Organizations budget for the first and receive the second free. When AI makes the production cheap, a role judged on its output alone looks like a cost to cut, and cutting it removes the apprenticeship too, without any line on any report recording the loss. I am not claiming that AI is the main reason graduates are finding it harder to get hired; the evidence is early and its authors say it is not causal. Nor am I arguing that firms should pay people to do by hand what a machine does better. The claim is narrower: whoever decides about the first job is also deciding where the next experienced professional will come from, and should make that decision on purpose.

Where the doors are closing

The most cited evidence comes from Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at Stanford, who used payroll records from ADP covering millions of American workers. In the version of their paper dated November 13, 2025, workers aged 22 to 25 in the occupations most exposed to AI, such as software developers and customer service representatives, saw a 16 percent relative decline in employment after controlling for shocks to individual firms. Older workers in the same occupations held steady or grew. In the most exposed jobs, employment of the youngest group fell 6 percent from late 2022 to September 2025, while employment of older workers rose 6 to 9 percent. The declines were concentrated where AI is used to automate work rather than to augment it.²

A January analysis by Tyler Atkinson and Shane Yamco, also at the Dallas Fed, used a different source, the government's monthly household survey, and found the same pattern in a revealing form. The share of employment held by young workers in the most AI-exposed occupations fell from 16.4 percent in November 2022 to 15.5 percent in September 2025. The decline did not come from layoffs. It came mainly from fewer young people moving into those jobs from outside the workforce, which is to say, from fewer first jobs.³ The authors are careful about the size of the effect: even if every lost job had become unemployment, it would account for only a 0.1 percentage point rise in the national rate. They are careful about cause, too, noting that the pattern may reflect something else correlated with AI exposure, such as education.³

That shape matters for the argument. People already inside these occupations are mostly keeping their jobs. What is narrowing is the entrance. A labor market can absorb that for a while without any alarming headline, because the people who would have been trained are not yet anyone's employees. The cost shows up later, as a shortage of people with five years of experience, and by then it is hard to trace back to a hiring freeze.

Two kinds of knowing

Why would AI hit the youngest workers in a profession and spare the experienced ones? The answer turns on a distinction the philosopher Michael Polanyi drew in 1966, when he observed that "we can know more than we can tell."⁴ Some knowledge can be written down: accounting rules, the structure of a memo, the definitions in a filing. Some cannot: the sense that a number is off before you can say why, knowing which client's quiet email means trouble, judging when a draft is good enough to send. The first kind is codified. The second is tacit, and it is acquired mostly by practice under the eye of someone who already has it.

Davis applies the distinction to jobs. For a new graduate, the codifiable tasks are the expert part of the job: they are what the degree prepared the graduate to do. For an experienced worker, the same tasks are the routine part, and the value lies in the tacit judgment around them. A tool that can replicate codified knowledge therefore substitutes for the junior and assists the senior.¹ To test this, Davis used the Bureau of Labor Statistics' separate wage estimates for entry-level and experienced workers in more than 200 occupations, and computed each occupation's experience premium, how much more its experienced workers earn. The median premium is 40 percent. It is under 10 percent for fast-food cooks, ticket agents and dry cleaners, and over 100 percent for lawyers, insurance underwriters, credit analysts and marketing specialists.¹

He then asked how an occupation's exposure to AI related to its wage growth since late 2022, allowing the relationship to depend on that premium. The result is in Figure 2. Where experience earns almost nothing extra, more exposure to AI went with slower wage growth, as you would expect if AI can do the whole job. Where experience earns a great deal, more exposure went with faster wage growth. Davis's summary is that the payoff to experience is rising in the occupations most exposed to AI.¹

Who was paying for the classroom

If juniors were learning all along, who paid for it? Gary Becker's answer, from his 1962 paper on investment in human capital, is that mostly they did. Becker distinguished training that is useful only to the firm providing it from general training that raises a worker's value to many employers. A firm in a competitive market will not pay for general training, because the trained worker can take the skill to a competitor. So the worker pays instead, by accepting a lower wage while learning.⁵

The entry-level job was a quiet version of that bargain. The junior earned less than an experienced colleague, produced less, and took up some of the senior's time. In exchange the firm received a stream of codifiable work worth enough to cover the arrangement, and the junior received an education. Nobody had to write down the tuition, because the production paid it. That is the mechanism behind Davis's warning. When a machine can produce the codifiable work for a fraction of the cost, the junior's output no longer covers the junior's salary and the senior's attention. The bargain stops paying for itself, and a manager looking at the requisition sees a cost without a matching benefit. The benefit is still there. It was just never on the form.

Learning from the edge of the work

A second line of research explains why the codifiable tasks were not merely the price of admission but the route in. In 1991 the anthropologist Jean Lave and the learning theorist Etienne Wenger described how newcomers become competent members of a trade by what they called legitimate peripheral participation: real work, with real consequences, that starts at the low-risk edge of the practice and moves toward the center.⁶ In Lave's study of tailors' apprentices in Liberia, newcomers began with finishing work such as hems and buttons, and learned to cut cloth only later.⁶ The edge tasks were not busywork. Doing them, alongside people doing the harder work, was how the harder work became visible.

In a finance department, the reconciliation is the hem. A junior who reconciles the monthly reports learns, on the occasions when they do not tie, where the systems disagree, which accounts are fragile and who knows why. If the reconciliation is done by software and only exceptions reach a person, those exceptions go to whoever can already resolve them, usually the senior. The junior is left with no edge to start from. The work has not become harder for newcomers; it has stopped having a place for them at all.

That creates a problem that no single firm feels at first. The experienced workers whose wages rise in Davis's data were juniors once, trained on exactly the tasks now being automated. A firm that stops hiring juniors can keep buying experience from the market for a while, which is to say, from firms that kept training. If enough firms make the same choice, the supply of five-year professionals shrinks in about five years, and every firm discovers it at the same moment.

Redesigning the first job

The way out is not to keep juniors on work the machine does better. It is to redesign the first job so that it is still a classroom when AI does much of the production. There is evidence that this can work. In a study of 5,179 customer support agents, Brynjolfsson, Danielle Li and Lindsey Raymond found that an AI assistant raised productivity by 14 percent on average and by 34 percent for novice and low-skilled agents, with little effect on the most experienced. Their evidence suggests the tool spread the practices of the best agents to the newer ones and helped them move down the experience curve faster.⁷ Used that way, AI is not the reason the junior has nothing to learn from. It is one of the teachers.

What does that look like in practice? I would start with three changes. First, keep juniors on real work, done with AI rather than instead of it, so that they still own a reconciliation or a memo, and are accountable for whether it is right. Second, route some exceptions to juniors on purpose, with a senior reviewing the resolution, because the exceptions are where the tacit knowledge lives. Third, make the review itself the curriculum: when a senior edits an AI-assisted draft, the junior should see what changed and why. None of this is free. It costs senior time, which is the scarcest input in most professional teams. That cost is the tuition, now visible, and it belongs on the budget as an investment with a five-year return rather than hidden inside a headcount line.

Objections, considered

The cause is not AI. Graduate hiring has been weak for reasons that have nothing to do with software: higher interest rates, the end of pandemic overhiring, a cooler market for technology firms in particular. That is a serious objection, and the Dallas Fed authors themselves warn that the pattern may not be causal.³ The Stanford paper tries to separate AI from firm-wide shocks and still finds the decline, but its authors present their results as early evidence.² The argument here does not depend on settling the question. Whatever the reason a firm stops hiring juniors, it stops training them, and the codifiable work that once justified the hire is now cheaper to get elsewhere, which makes the decision harder to reverse when conditions improve.

Firms are not schools, and trained people leave. True, and Becker would agree: firms have never paid much for general training, for exactly that reason.⁵ But a great deal of what a junior learns is specific to the firm: its systems, its clients, its exceptions, its people. That part does not walk out the door with the same ease, and it is the part a firm cannot buy from the market at any price. The question is not whether to run a charity. It is whether the firm wants experienced people who know its business, and if so, where it expects them to come from.

Schools and AI tutors can do the training. They can do more of it, particularly the codifiable part. But that is precisely the part the machines now perform, so it is not the scarce part. What Polanyi called the knowledge we cannot tell is learned in practice, on real cases, near people who have it.⁴ A classroom can describe an exception. Only a job can hand you one.

Automating the junior's work does not automate the junior's education.

Before the next requisition is frozen

• Write down what the role teaches. Next to the duties, list the judgment a person in the role acquires by the end of the second year, and who acquires it if the role goes.
• Name the pipeline. Say where the team's experienced people will come from in five years, and whether that source is a plan or an assumption about other firms.
• Keep juniors on real, reviewed work. Give them AI-assisted work they own, a share of the exceptions, and a senior who explains the edits.
• Budget the tuition. Count the senior time that training costs as an investment line, so that it can be weighed honestly instead of discovered after it is gone.

Back to the requisition. Its duties box is accurate, and it may be right to shrink it: much of that list is work a machine can now draft in seconds. What the form needs is the second section, written in a different hand: why a reconciliation does not tie, which exceptions matter, how a senior reviews a draft, whom to call when a number looks wrong. That section was always the larger part of what the role produced. It just never had a box.