This piece follows one change downstream. Agents now absorb the discrete tasks that first-year staff once learned on, and firms are responding by hiring fewer juniors.
Key points
- The best agent now succeeds half the time on tasks that take a skilled person about five hours1.
- Employment of workers aged 22 to 25 in AI-exposed occupations sits 19 percent below trend, and reduced hiring drives the gap2.
- Fixed fees and outcome pricing remove the client payment that once funded junior training34.
- Firms that stop training juniors will likely compete for a shrinking pool of seniors by 20305.
- New disclosure laws track AI layoffs and miss the quieter cut in entry hiring6.
The change
METR puts the best agent's 50 percent time horizon at 320 minutes1. That horizon has doubled every 131 days since 20231. Agents still fail most end-to-end professional projects. The top Remote Labor Index score is about 16 percent7. We weighed that tension in our contradictions piece. Employers are acting on the trend line anyway. In 2025, 17 percent of CEOs planned to shift away from junior roles. In 2026 the share is 43 percent5. A Dallas Fed paper finds that AI substitutes for entry-level workers and complements experienced ones8.
What follows
First-order
- Tool budgets beat training budgets. At one forum of law firm COOs and CFOs, nearly every table protected the AI drafting platform and trimmed associate pay9.
- Elite intakes shrink first. Linklaters cut trainee places from 100 to 60. Freshfields cut from 85 to 7010.
- Clients start pricing out junior hours. Citigroup asks bidding firms to show AI savings. Morgan Stanley plans more fixed fees3.
- Software drops the seat. Zendesk charges 1.50 dollars per automated resolution4. Each resolution billed this way replaces work a junior support seat once did.
- Shadow side. A half-reliable agent still needs a human checker. Firms that cut juniors lose the people who learn by doing that checking1.
Second-order
- Junior training loses its payer. Under a fixed fee, first-year hours become a cost paid from partner profit3. We judge it likely that most firms cut that cost in a down year9.
- Experience earns a wider premium. Wages in computer systems design rose 16.7 percent since fall 20228. National wages rose 7.5 percent over the same period8.
- The degree is worth less. The US college wage premium fell from about 87 percent in 2022 to about 78 percent in 202611. A working paper that has not been peer reviewed assigns AI about 30 percent of the drop11.
- Training splits in two. Ropes and Gray credits 20 percent of first-year billable hours to AI learning12. Weil runs a ten-week AI program for 1L summer associates12. Prompting skill teaches a first-year little about reading a contract.
Third-order
- A senior gap opens. Finance leaders already name senior accountants as their hardest hire5. Firms that buy seniors externally assume someone else trained them. If junior hiring stays below trend, we put about 60 percent odds on visible senior shortages in law, finance, and software by 2030.
- Tacit skill loses its source. Senior judgment grows from years of codified work8. Automating the first rung removes the path to the second.
- Cheap capability spreads the effect. Open-weight models trail the closed frontier by about four months13. Small firms and offshore rivals will plausibly gain junior-grade agents at low cost within a year of each frontier release.
- Policy watches the wrong door. California, Connecticut, and New York now require AI disclosure in layoff notices6. A firm that quietly stops hiring juniors files nothing6. The pending federal SKILL Act would pay employers credits for hiring graduates6.
Who gains, who loses
- Experienced specialists gain pricing power as the supply behind them thins85.
- Vendors paid per outcome capture budget that once paid junior salaries4.
- Graduates from 2026 to 2028 lose entry points and carry debt priced on the old premium11.
- Firms that fund training like pilot simulator hours plausibly build the next senior bench9.
- Firms that rely on hiring seniors away from rivals lose once every competitor tries the same move5.
A glimpse ahead
Spring 2030. A mid-size firm posts a role for a fifth-year associate. The salary band sits far above its 2026 level. Twelve applications arrive. Nine come from lawyers at rival firms who will draw counteroffers. The hiring partner checks the class of 2025 and finds it was half the usual size. She opens a second posting for six first-years. It promises 400 credited hours of simulated depositions. Finance flags the line as unbillable. She approves it anyway.
Sources
- METR Time Horizon 1.1 puts the best agent at a five hour task horizon with a 131 day doubling time · grey-lit ↩
- Stanford Canaries update finds young workers in AI exposed jobs 19 percent below trend with the gap driven by hiring · grey-lit ↩
- Legal departments expect AI to cut reliance on the billable hour while most firms report no client price pressure yet · grey-lit ↩
- SaaS chiefs plan to retire seat pricing within two years as vendors bill per automated resolution · journalism ↩
- Oliver Wyman survey finds 43 percent of CEOs plan to shift away from junior roles as senior accountants become the hardest finance hire · journalism ↩
- California and Connecticut join New York in requiring AI layoff disclosure as Congress weighs a credit for hiring graduates · expert ↩
- Remote Labor Index shows top agents finish 16 percent of real freelance projects up from 2.5 percent in eight months · peer-reviewed ↩
- Dallas Fed finds AI substitutes for entry-level workers and complements experienced ones across 205 occupations · institutional ↩
- Law firm COOs and CFOs protect AI budgets and cut associate pay while few rebuild the training AI displaces · grey-lit ↩
- UK trainee solicitor places fall 1.7 percent across 109 firms as solicitor apprenticeships absorb the decline · journalism ↩
- Azar, Giné, and Sanz-Espín, The College Wage Premium in the Generative AI Era, SSRN working paper, September 2026. https://marginalrevolution.com/marginalrevolution/2026/09/the-college-wage-premium-in-the-generative-ai-era.html · peer-reviewed ↩
- Ropes and Gray credits a fifth of first-year billable hours to AI learning and Weil opens a 1L AI track · journalism ↩
- Best open-weight models trail the closed frontier by about four months, and the gap has widened slightly · grey-lit ↩