Harvey built a legal AI that lawyers prefer 97% of the time

Harvey, the AI startup targeting legal, tax, and finance professionals, just gave a rare look under the hood of its custom-built case law model — and the numbers are hard to ignore. After a year that saw the company grow to over 100 employees, 10x its revenue, and lock down $80 million in Series B funding at a $715 million valuation, Harvey partnered with OpenAI to train a model specifically on U.S. case law.

Early experiments with GPT-3 hinted at the potential. Co-founder Winston Weinberg, a former antitrust litigator, recalls feeding 100 landlord/tenant questions from Reddit into the model. Attorneys said they’d send 86 of those answers straight to a client. But real legal work demands more than plausible-sounding text. Standard techniques like retrieval-augmented generation could handle simple lookups, but fell flat when lawyers needed to build actual arguments. “With case law research, you’re finding ammo for your argument, and that’s much more difficult to do,” Weinberg said.

So Harvey went deeper. Working directly with OpenAI’s team, they injected the equivalent of 10 billion tokens of case law data — starting with Delaware and expanding nationwide — into a custom-trained model. The result isn’t just accurate; it’s thorough. When 10 major law firms compared the custom model’s output against GPT-4 on identical prompts, the lawyers chose Harvey’s answers 97% of the time. The preference came down to completeness. The custom model produced longer responses that addressed the nuance of the question and surfaced more relevant precedent. More critically, the hallucination problem that plagues legal AI seems genuinely mitigated here. Weinberg claims every sentence is supported by the case it cites, and the model doesn’t invent rulings.

Co-founder Gabe Pereyra, who previously worked on LLMs at Google Brain and Meta, has a sharp piece of advice for other AI founders: don’t build for today’s models. He warns that incremental improvements from foundation models can swallow entire startups as a side effect. Harvey’s bet is on tackling problems so complex — like drafting briefs that account for jurisdictional differences or orchestrating multi-step agent workflows — that a better base model isn’t enough on its own. That’s the moat they’re digging right now, while the rest of the field is still figuring out where to put the shovel.