The AI labs are automating the work that trains the experts they need

Microsoft is hiring a "principal legal engineer" — a practicing lawyer with at least seven years of experience, paid between $147,000 and $279,000, to build AI agents that automate the manual legal work of attorneys and paralegals ^[1]. The posting is specific about what it wants.

This is a role for a builder. — Microsoft

A senior professional, hired to build the tool that removes the junior role that produces the next senior professional. The want ad is the whole arrangement on one page. The labs are not doing this from a position of strength. OpenAI is projected to miss its five-year advertising revenue forecast by roughly 90% — eMarketer puts the entire U.S. chatbot ad market at $5.41 billion by 2030 against OpenAI's $100 billion projection ^[2]. The general-purpose chatbots are not paying for themselves. Meanwhile, the labs have turned to specialized enterprise tools, and this week Google launched its own legal suite arguing that general-purpose AI falls far short of what lawyers need ^[3]. The tools are aimed squarely at the work that used to train people. Document review, due diligence, credit risk assessment, market briefings — the tasks a junior associate or analyst learns the profession through ^[4][5][6][7]. IBM's head of HR put a name to what gets lost.

I think there’s a small, short term, realistic thing that’s happening as people are saying, we don’t know what these entry-level hires will do, because with our old programs, we don’t need them to do those things anymore. — Nickle LaMoreaux

The lab CEOs are not pretending otherwise. At Davos, Anthropic's Dario Amodei said he can foresee needing fewer people, not more, at the junior and intermediate levels, with the first signs already visible in software ^[8].

I can see it within Anthropic, where I can look forward to a time where on the more junior end and then on the more intermediate end we actually need less and not more people. — Dario Amodei

The high-margin revenue has not arrived. Legal tech's growth is still largely pilot programs booked like recurring revenue, and 62% of lawyers use ChatGPT — more than twice the share using specialized tools like LexisNexis or Copilot ^[9]. Gartner forecasts more than 40% of agentic AI projects will be canceled by the end of 2027 ^[10]. The general-purpose chatbot the labs are trying to leave behind is the thing customers actually use. And the biggest customers are not buying at all. Kirkland & Ellis is spending $500 million to build its own platform, drawing on 250 of its own lawyers ^[11]. The firm with the deepest bench of legal expertise has decided it does not need to license the labs' tools — it needs the labs' problem, solved in-house. Then there is the accuracy gap, where the trap closes. Sullivan & Cromwell, one of the most prestigious firms on Wall Street, filed a court document with roughly 40 errors — fabricated quotations, citations to cases that do not exist ^[12]. The Oregon Court of Appeals now tracks staff time spent cleaning up AI-fabricated legal authority, and its chief judge says the fake filings are draining resources from the court's actual work of deciding cases ^[13]. LexisNexis's CEO warns that attorneys will eventually lose their licenses over this ^[14]. Every one of these errors is caught by an expert — the kind of expert the tools are busily automating out of existence. The endgame is already visible. Superlegal, licensed by the Utah Supreme Court, sells commercial contract review for $117 with 24-hour turnaround ^[15].

We built our AI from the ground up to specialize in legal practice, with an attorney in the loop. — Noory Bechor

The attorney in the loop is a senior one. The junior role is the part that got automated. The seniors came up through a pipeline this model no longer produces. The labs' own research is rosier. Goldman Sachs found AI's labor-market impact has stayed limited so far, and Sam Altman has admitted his early predictions of rapid entry-level elimination were wrong ^[16][17]. But the institutions publishing the rosy findings are the same ones building and buying the displacement tools, and their product roadmap points one direction while their research points another. The lab and the firm are converging on the same shape. Microsoft needs a seven-year lawyer to build the tool; Superlegal needs a senior attorney to check the output. Both are drawing on a stock of expertise the tools themselves are no longer replenishing. The question is not whether the lab replaces the firm. It is what happens when both are running on the same dwindling supply of people who know enough to catch the errors.