Market Commentary

Entry-Level Jobs and AI: who trains the senior people of 2032?

Companies are hiring fewer beginners, and one of the clearest signals comes from a company that builds the AI doing their work. The saving is real and immediate. The cost is deferred, and it is not obvious who will pay it: the graduates who never get the first job, the professions that need a steady supply of experienced people, or the firms that made the cut.

By Work desk × Folka

1. From 27% to 9%

One of the most telling figures on AI and early careers this autumn came from inside an AI company. In a report on its own workforce, Anthropic disclosed that juniors' share of its technical hires had fallen from 27% to 9%, and that juniors now make up 14% of its technical staff, down from 21% at the start of 2025. The same report found that meeting time per employee had risen by a fifth in a year.

It is a company describing the change from the inside, and it should be read with that in mind. The report gives no definition of "junior" and no headcount. A frontier lab hiring fast for scarce research talent has an unusual mix of staff, and Anthropic itself attributes part of the shift, technical teams falling from 61% to 56% of the workforce, to business functions growing as the company scales. It published the report in the same week it announced a public-benefit fund that names labour disruption among the causes it might support. None of that makes the figures wrong. It makes them one company's account, not a forecast for everyone else.

2. One company, or the start of a pattern

The wider evidence points the same way, with varying strength. The most careful study is from Stanford's Digital Economy Lab, which uses payroll data from ADP. Its latest version, from August, finds that employment of 22- to 25-year-olds in the occupations most exposed to AI now stands 19% below where it would be had it kept pace with less exposed jobs. Experienced workers in the same occupations show no comparable gap, and the effect works mainly through fewer hires rather than more departures. The authors are explicit that these are early, descriptive indicators, not proof that AI is the cause.

The New York Federal Reserve's tracking of recent graduates is consistent with that. Unemployment among graduates aged 22 to 27 stood at about 5.6% in March, against 3.6% in March 2019. About four in ten of those in work are in jobs that do not need a degree, though that share has hovered around the same level for two decades, so it describes a long-standing condition rather than a new one.

The rest of the evidence is softer. In a survey by the Graduate Management Admission Council, one in three employers said they were replacing entry-level roles with AI. Pantheon Macroeconomics estimates that AI is now causing more than 10,000 US job losses a month. Kristalina Georgieva, head of the International Monetary Fund, described AI's effect on work as a "slow-moving tsunami" and named young people entering the labour market as the most exposed.

3. What the routine work was for

The case for cutting junior roles rests on a simple observation: much of what juniors do is routine, and AI now does routine work well. The observation is correct. What it misses is what the routine work was for.

The anthropologists Jean Lave and Etienne Wenger studied how people learn a trade, from tailors to midwives, and found that newcomers learn by taking part in the real work at its edges, doing small, low-risk tasks while watching how the experienced people handle the hard ones. They called it "legitimate peripheral participation". The philosopher Michael Polanyi made a related point in 1966: "we can know more than we can tell". Much of what experts know, how to judge a draft, when a number looks wrong, which client to call first, cannot be written down. It is absorbed by being near the work for long enough.

Read carefully, Lave and Wenger's argument is about access more than about the tasks themselves. What made the edge of the work valuable was that it sat next to the centre: the junior could see how the hard cases were handled. The routine tasks were simply how juniors got that access. That cuts two ways. A firm that automates the routine work loses the apprenticeship only if it also loses the access. And a firm that keeps juniors on routine work, but out of sight of the people who handle the hard cases, may keep the jobs and lose the apprenticeship anyway.

4. The ladder is moving, and nobody knows how far

If AI capability stood still, the damage would be contained: a narrower bottom rung, and a longer climb from it. It is not standing still. OpenAI says its latest models make about half as many mistakes on its own tests as the last generation, at half the price. If models keep improving at the pace the labs claim, and this is a scenario, not a forecast, the line between work AI can do and work that needs a person will move up the ladder, from junior tasks towards the work of the people a few years in.

The counter-evidence matters as much. Capability is not the same as reliability. The labs themselves have pulled their agents off the live internet this month because they cannot yet control them, which suggests that for some years much AI work will need an experienced person to check it.

These two cases lead to different futures, and the argument about entry-level jobs depends on which one arrives. If capability keeps climbing, demand for mid-career people may fall along with demand for juniors, and a firm that stopped training may never feel the gap. If capability stalls near junior work, firms will need more people with judgement to supervise the machines, at the same moment they have trained fewer of them. Call the senior capability that goes missing in the second case the apprenticeship debt: experience that was never built because the first rungs were removed.

5. Whose debt it is

Economics has a long answer to who owes that debt, and it is not mainly the firm. Since Gary Becker's work on human capital in the 1960s, economists have argued that firms in competitive labour markets under-invest in general training, because a trained worker can leave and take it with them. Daron Acemoglu and Jörn-Steffen Pischke showed in 1998 that firms do pay for general training where wages are compressed and workers are harder to poach, which helps explain why firms in Germany and Japan train more than those in the US. Applied here, a single firm that buys experience later may be acting rationally. If a shortage comes, it is shared across the market rather than called in on the firm that caused it.

That makes the apprenticeship debt mostly a debt owed by a profession, or a country's skills system, as a whole. It is run up by individual firms and may be repaid by everyone, or by nobody. That is why governments step in: the UK's apprenticeship levy, charged since 2017 at 0.5% of pay bills above £3 million, exists because firms train less than the economy needs when they can hire the trained away. Not all of the debt is shared, though. Knowledge of a firm's own clients, systems and habits cannot be bought from a competitor, and Becker's own distinction holds that firms do pay for training of that kind. The part of the debt a firm keeps is the part only it could have taught.

6. Who bears it

The first people to bear the cost are the graduates. A first job has always done more than pay a wage. It is how people without family connections get into a profession: the place where someone with no network can prove they are good. If entry-level roles shrink, those with connections will still find a way in, through internships, introductions and second chances. Those without will find fewer doors. Two other changes narrow the way in further. In a Gallup survey for the Lumina Foundation, which funds higher education and so has an interest, 48% of hiring managers said degrees would become more important over the next five years. And the screening itself is being automated: HackerRank's AI interviewer has already run more than 500,000 interviews, while in a survey by TestGorilla, which sells AI recruiting tools, a third of recruiters who found candidates faster had gained no confidence in them. Fewer first jobs, more weight on credentials and more automated gatekeeping all fall hardest on the people with the least help.

The second group is the people already inside. Senior work is shifting from doing towards directing and checking, increasingly the output of AI rather than of junior colleagues. There is early evidence that this has a cost of its own. A 2025 study in The Lancet Gastroenterology & Hepatology, from four centres in Poland, found that after AI support was introduced, experienced endoscopists detected fewer precancerous growths in procedures done without it. The study is observational and covers one specialty, but it is real-world evidence of a risk the engineering psychologist Lisanne Bainbridge described in 1983: automate the routine, and the skill needed for the rare hard case erodes. The apprenticeship debt may not only be a missing generation of juniors. It may also be current seniors losing sharpness as their work turns into checking.

7. The case that this is overstated

The strongest argument against this read deserves a full hearing. The Stanford authors stress that their figures describe a pattern, not a proven cause. The decline in US junior tech hiring began in 2022 and 2023, when other forces were at work: the end of near-zero interest rates, a correction after pandemic over-hiring, and a tax change that, from 2022, forced US firms to spread the cost of software developers' wages over five years rather than deduct it at once. The Federal Reserve's chair has described a "low-hire, low-fire" labour market that is cautious about everyone, not only juniors.

That tax change has since been reversed, for tax years beginning after 2024. That offers something close to a test. If junior hiring in AI-exposed jobs kept falling through 2026 after that cost was removed, the AI explanation gets stronger, and the August Stanford figure falls in that period. It is not conclusive, but it is the cleaner window.

The other counter-argument is that AI may teach, not only replace. In a study of more than 5,000 customer-support agents, Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that an AI assistant raised productivity by 14% on average and by 34% for novice and less-skilled workers, with little effect on the most experienced. Their suggested mechanism is that the tool spread the practices of the best workers to newer ones. That is tacit knowledge passing through the machine. A call centre is not a law firm, and the study cannot say whether newcomers learned the judgement or only its performance. But it is the strongest evidence that apprenticeship may change form rather than disappear.

8. Where the ladder still has rungs

The picture is not uniform, and the differences are instructive. Outside the US, the world's largest employers of new graduates are moving in different directions. India's IT services firms have long been the main entry point into technology work for very large numbers of graduates. Tata Consultancy Services, which hired more than 44,000 freshers in its last financial year, has offered about 25,000 for this one and says more depends on demand. Infosys says it will hire at least 20,000 and is recasting them as "AI-native" engineers. Wipro has paused campus hiring for the year, reported as reflecting both uncertain demand and AI-driven efficiency. HCLTech brought in more than 6,000 freshers in its latest quarter.

In Britain, the professional-services pyramid is the apprenticeship model in its purest form. PwC UK cut its graduate and school-leaver intake from 1,500 to 1,300 in 2025; its UK head named the weak economy as the main reason, said AI was reshaping roles, and named offshoring alongside it.

Demography matters as much as technology. In China, unemployment among 16- to 24-year-olds outside education reached 18.9% in August, a one-year high, as a record cohort of about 12.7 million graduated, for reasons that have far more to do with the economy than with AI. In Japan the position is reversed: there are 1.62 job openings for every private-sector applicant graduating next March. Where young workers are scarce, AI that does junior work does not shut a door. It fills a seat nobody was available to take.

The industries differ too. Law, software, finance and consulting each depend on junior work in different ways, and the apprenticeship debt will fall hardest where expertise is built slowly, through practice close to experienced people, and is hard to buy in later.

9. Our view

The entry-level question is being treated as a cost decision. It is also a decision about where future judgement will come from, and the usual calculation struggles to see it, because the saving appears in this year's budget and the cost, if it comes, appears in seven or ten years with nothing to connect it back.

Several forces push in the same direction. The AI companies are selling, accurately, a tool that does junior work well. Boards reward headcount discipline in an uncertain economy. Psychology plays its part: a cost avoided now feels more real than a capability lost later, and the people who would have been hired are invisible precisely because they never arrived. The economics adds a harder truth. A firm that stops training may never pay for it, because the shortage, if it comes, will be shared by everyone. That is what makes this a question about professions and societies, not only about companies.

Whether it matters at all depends on how far AI climbs. If it keeps rising, the debt may be smaller than it looks; if it stalls near junior work, it may be larger than anyone has counted. Nobody yet knows which, including the companies building the tools. What can be said is who carries the uncertainty in the meantime. Graduates carry it first, and those without connections carry the most of it. Seniors carry some, as their work turns to checking. The firms carry least, at least for now.

The junior intake was never only a cost line. The firms cutting it are making a bet about the future of expertise, and the people who will find out whether the bet was right are mostly not the ones who placed it.

10. The forks, 2032, and what to watch

There are four realistic responses. Each rests on an assumption that could prove wrong, and each looks different depending on whether AI's reach keeps climbing or stalls near junior work.

Keep a junior intake, as a training cost. Hire beginners deliberately, even where AI could do their tasks. This assumes the firm will keep the people it trains. The assumption most at risk is the textbook one: competitors who stop training can hire the trained away, and the training firm pays for everyone's apprenticeship. It looks wise if AI stalls near junior work, and expensive if AI keeps climbing.

Rebuild apprenticeship around supervising AI. Give juniors a different first rung: checking, directing and correcting AI output, close to experienced people. This assumes judgement can be learned by supervising machines rather than doing the work. The evidence cuts both ways: the call-centre study suggests AI can pass on good practice, the endoscopy study that supervising machines can erode skill. If it works, the firms that design it first set the standard.

Buy experience later. Hire fewer juniors now and recruit experienced people when needed. Its strongest case is the economics: for a single firm in a competitive market, letting others bear the cost of training can be rational. This assumes the market will supply them. The assumption most at risk is the supply itself. It looks wise if AI keeps climbing, and costly if it stalls and every firm has made the same choice.

Share the cost of training. Treat apprenticeship as a responsibility of a profession or a country rather than a firm: levies, industry academies, licensing bodies that require supervised practice. This assumes collective schemes buy real apprenticeship. The assumption most at risk is that they produce box-ticking instead, as critics of levy schemes have long argued.

---

How we read this

This commentary draws on Anthropic's workforce report as covered by TechCrunch; Stanford's Digital Economy Lab study of ADP payroll data (August 2026 revision); the New York Fed's recent-graduate data; surveys by GMAC, Gallup for the Lumina Foundation and TestGorilla; estimates by Pantheon Macroeconomics; hiring disclosures from TCS, Infosys, Wipro, HCLTech and PwC UK; China's National Bureau of Statistics and Japan's Recruit Works Institute; and research by Lave and Wenger (1991), Polanyi (1966), Becker (1964), Acemoglu and Pischke (1998), Brynjolfsson, Li and Raymond (2025), Budzyń and colleagues (2025) and Bainbridge (1983). Company and vendor figures are their own claims. The two capability scenarios and the conditional 2032 projection are Folka's editorial read. The one number that would most change this read is junior hiring in AI-exposed jobs through 2026, after the US tax change was reversed.

Grounded in Folka’s corpus: The Junior Pipeline Collapses · Post-Credential Hiring · Proof Without the Degree · The Absorbed Hour · The Automation Wage