The setup
In the late summer of 2024, four AmLaw 10 firms (Cravath, Skadden, Wachtell, and Sullivan & Cromwell) quietly began the most consequential retraining program the profession has seen in a generation. They did not announce it. They did not market it. They did not even use the same vocabulary internally.
But the through-line was identical: the entry-level associate role, as practiced for the better part of forty years, had to be rebuilt from the ground up. Not because AI was coming for the work. Because AI had already eaten 60% of it.
What changed in 18 months
Three things broke at roughly the same time. First, GPT-4-class models proved capable of producing first-pass drafts of routine memoranda that were materially indistinguishable from junior-associate output, at roughly 1/200th the cost. Second, the procurement teams at top GCs noticed. Third, the firms billing realization on first-year work collapsed.
By Q2 2026, internal data we obtained from one of the firms showed first-year billable hours down 19% year-over-year, with realization on those hours down a further 11%. The math no longer worked.
Inside Cravath rebuild
Cravath response, which insiders refer to internally as "the Pivot," reframed the first two years of practice around what one partner called supervised judgment. First-years no longer draft from blank pages. They review AI-generated first drafts against a curated playbook, identify gaps, and own the explanation to the client.
The training emphasizes three competencies the firm believes will define the next decade of practice: hallucination detection, structured prompting, and what the firm calls issue spotting at speed: the ability to read fifty AI-generated paragraphs and find the three load-bearing sentences in under ninety seconds.
The economics, restated
The firms are not, contrary to widespread reporting, cutting headcount. They are restating the economics. A first-year now costs the firm roughly the same in salary and overhead but produces, on average, 2.4x the throughput. The leverage ratio that made BigLaw work for forty years is being rebuilt around a different denominator.
Whether clients eventually demand to share that productivity gain through lower rates is the open question that hangs over every partnership meeting we attended.
Where this lands
The firms that move first are not necessarily the ones that win. The firms that move correctly are. The risk profile of getting the AI integration wrong (sanctioned filings, privilege breaches, blown diligence) is more existential than the cost of moving slowly.
But the firms that have not started restructuring the first-year role by the end of 2026 will, in our view, be running a model that no longer exists.