At Allen & Overy, one of the first large firms to deploy Harvey — the legal AI startup built on top of OpenAI's models — a first-year associate can now generate a comparable first-pass contract review in the time it used to take to open the file. The memo still gets written. The associate who used to spend a week drafting it, only to have a partner mark it up in red and hand it back, is no longer required to produce that draft from scratch to get one.
The mechanism is economic before it is educational. Entry-level knowledge work was always overpriced relative to its output and underpriced relative to its training value. Clients paid junior billing rates for junior drafts because the alternative — a partner doing everything — cost more, even accounting for the rework. That arrangement quietly subsidized the profession's own reproduction. Once a tool like Harvey, or McKinsey's internally built assistant Lilli, produces the same rough draft for a fraction of the cost, the subsidy disappears, and with it the logic for hiring as many juniors to do the work at all.
Law, consulting, and journalism share a structure even though they look different from outside: a pyramid where a wide base does high-volume, low-judgment work under supervision, and a narrow top exercises judgment earned by having done that work for years. Compress the base and the pyramid does not become a smaller pyramid. It becomes a shape with fewer people positioned to eventually reach the top.
The professions most exposed are the ones where entry-level work looked most like what a model does well: summarize, draft, format, extract. Litigation document review, first-pass consulting decks, wire-service rewrites. The professions least exposed involve physical presence, live judgment under pressure, or relationships that cannot be delegated — a narrower set than most people assume, and one that does not include most white-collar entry jobs built in the last fifty years.
Firms are responding unevenly. Some have simply shrunk incoming associate and analyst classes, treating the change as a cost-saving opportunity. Others, including parts of the Big Four, are experimenting with compressed timelines that push new hires toward judgment-level review work faster, using the AI-drafted output as a teaching artifact rather than a deliverable. Whether that produces the same depth of skill as years of repetition is untested, because no cohort trained this way has reached the career stage where the gap would show.
Expertise itself will not disappear — partners trained under the old system will keep working for decades. What is arriving on a schedule instead is a demographic gap: a shortage of people who logged the reps needed to become the next generation of partners, timed almost exactly to when the current generation retires.
Nobody designed the apprenticeship model on purpose, which is part of why nobody at Allen & Overy or McKinsey owns the job of redesigning it. It emerged from an economics that Harvey and Lilli just broke. Whoever rebuilds it will have to fund talent formation as a line item, on purpose, instead of collecting it as a side effect of overpriced junior work — because that side effect is not coming back.