AI is absorbing the work that used to train experts.
Stanford's payroll analysis finds early-career employment falling only in the occupations where AI substitutes for tasks, and only through hiring. Switzerland is the one country that has already priced what juniors produce — which makes the exposure measurable.
Gli Insights sono pubblicati in inglese.
Since August 2025, Stanford’s Digital Economy Lab has tracked whether generative AI is yet visible in American payroll data. The August 2026 version of Canaries in the Coal Mine?, by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, uses administrative payroll records from ADP covering millions of US workers through June 2026.
Its central finding has moved. Employment of workers aged 22 to 25 in the occupations most exposed to AI now stands 19% below where it would be had it kept pace with their less-exposed peers. In levels, employment for that age group fell by about 11% in the two most exposed quintiles between November 2022 and June 2026, while the three least exposed quintiles grew by about 10%. Earlier versions headlined a regression estimate adjusted for firm-level shocks — 13% on July 2025 data, 16% on September 2025 data; on the simpler descriptive measure they now emphasise, the shortfall was 15% at the July 2025 data vintage.
Two absences matter as much: the study finds no evidence of widespread, economy-wide job displacement, and experienced workers show no comparable gap between more and less exposed occupations.
The adjustment is running through hiring, not through dismissal
The paper’s fourth fact separates two very different worlds. A falling employment stock can reflect fewer hires or more departures. The authors find no evidence that separations explain the divergence: separation rates fell for both groups, and among young workers fell at least as much in the most exposed occupations as in the least — the opposite of what displacement would produce. What opened after 2022 was the gap in hiring rates.
The sixth fact points the same way: adjustment occurs through employment rather than base pay, with salary trends differing little by age or exposure.
For a management team this is the practically important part: a change of this kind produces no restructuring announcement and no wage signal. It appears as a sequence of reasonable decisions: a departing junior not replaced, a task given to an experienced colleague with a tool, an intake postponed a quarter.
What matters is whether AI substitutes for a task or supports it
The fifth fact distinguishes automation from augmentation using the Anthropic Economic Index, which estimates, for each occupational task, the share of AI conversations that substitute for the work and the share that complement it. Declines are concentrated where usage substitutes for human tasks; where it complements, employment is flat or rising, particularly for experienced workers.
The August 2026 version adds a mechanism test, scoring occupations for reliance on codified knowledge — formal, documented content taught from textbooks and procedures — and on tacit knowledge acquired through practice and mentorship. Higher codified-knowledge scores go with slower entry-level employment growth; higher tacit-knowledge scores with faster growth for mid-career and senior workers.
The authors call their results descriptive indicators rather than causal estimates. The patterns attenuate when education is controlled for, some divergent trends predate generative AI, and the divergence is more pronounced in the ADP sample than in national survey benchmarks. A note published on 9 February 2026 addressed the most common alternative explanation: AI-exposed occupations are, if anything, less sensitive to interest rates, so rate rises do not account for the pattern.
Switzerland has already priced what a junior produces
The economic argument beneath this is old. Luis Garicano and Luis Rayo set it out in CEPR Discussion Paper 20634, published in September 2025 and revised on 2 March 2026: juniors pay for their training by doing menial work, and AI now performs an increasing share of it. Their model turns on one ratio: the output of a fully trained graduate relative to that of a novice who knows just enough to outperform AI. Above a critical threshold, careers remain at least as profitable as before; below it, the senior’s saleable knowledge shrinks and training compresses.
Most economies discuss this as a theory. Switzerland measures it. Two-thirds of young people enter vocational education and training, choosing among roughly 250 occupations, and the cost and benefit of training them is surveyed periodically. The Swiss Education Report 2026, published by SKBF | CSRE on 23 March 2026, reports the latest wave (Gehret et al., 2025, covering 6,655 training companies): firms invest around CHF 5 billion a year, apprentices’ productive services were worth about CHF 5.7 billion in 2022/23, and the net benefit was around CHF 800 million — an average of CHF 4,500 per apprenticeship year and contract.
The composition of that benefit is where the exposure sits. Productive activities — unskilled work paid at the company’s lowest rate, and skilled work valued at a qualified worker’s rate — account for around four-fifths of an apprentice’s time in the firm. Costs exceed benefits in the first year and are recovered later: a net benefit averaging about CHF 17,510 over a four-year VET diploma and almost CHF 14,000 over a three-year one. The margin that keeps the Swiss system running is the productive value of routine work done by people who are still learning.
Swiss firms have already been asked what they would do
The same report describes a spring 2024 survey experiment by the University of Bern’s Centre for Research in Economics of Education, covering more than 2,000 Swiss companies. HR managers were given scenarios in which a share of work activities disappears — automation of 20%, 40% or 60%, arriving in 2026, 2028 or 2030 — and asked what this would do to their apprenticeship vacancies. On average they expected an immediate and significant decline, though smaller than the loss of tasks in the scenario, suggesting they assume new tasks will replace some of the old. Responses were stronger in occupations already highly automated.
Young people are reading the same signals. Using searches on the LENA apprenticeship platform, Goller et al. (2025) found that after ChatGPT’s launch in November 2022, interest in cognitively demanding occupations fell sharply, while manual occupations were far less affected.
None of this is a forecast. A German study cited in the same report found that firms using AI tend to train more apprentices, though its data predate generative models. And the Swiss entry market is not under strain today: youth unemployment stood at 2.3% in 2024 on SECO’s register, or 8.2% on the internationally comparable ILO definition, against 15.5% across the EU in June 2026 (Eurostat). The question is whether the ledger that finances Swiss training still balances in 2030.
The decisions available to a management team are ordinary ones
Little of this is within a single firm’s control. Five things are.
- Separate the junior year into tasks AI now does better, tasks it supports, and tasks it cannot touch. The Stanford results track that distinction, not job titles.
- Recost your own entry-level position. The national average net benefit is CHF 4,500 per apprenticeship year and the first year a net cost; knowing your own figure turns attrition into a decision.
- Move judgment work earlier, then test it. If AI produces the first draft, the junior’s task becomes checking, correcting and deciding — closer to what seniority consists of. Write down what a second-year should manage unaided.
- Track hiring of under-26s as a separate figure. The effect in the data is in hiring rates, not headcount.
- Protect the training relationship. Companies that have not trained for some time are unlikely to resume quickly, the Education Report notes, and Swiss school-leaver numbers are rising again since 2023.
The next readings arrive on fixed dates
Three markers belong in the calendar. Stanford’s AI Economic Indicators, released monthly, now track the paper’s facts, with updates on 23 September and 21 October 2026. SERI’s Nahtstellenbarometer surveys school leavers and companies each April and August, with a detailed report in December; a shift in Swiss apprenticeship supply would show there first. And the World Economic Forum’s Future of Jobs Report 2025 — over 1,000 employers, more than 14 million workers, 55 economies — gives the aggregate intention: employers expect 39% of core skills to change by 2030, down from 44% in 2023, and 85% plan to prioritise upskilling.
Stated intentions are not outcomes, and none of this evidence yet establishes causation. What it does establish is where the decision sits. The senior specialists a company expects to have in 2035 are being selected now, in hiring meetings and task allocations that no one records as strategy.