The Harvey Playbook: How AI Cracked Big Law

Two days ago I tore down SpaceX buying Cursor for $60 billion, a company that went horizontal, every developer on earth, and got exposed on one thing: it ran on someone else's models. Harvey is the opposite bet, and it is worth studying precisely because it is the mirror image.

Harvey sells AI to lawyers. Not "legal," lawyers, the most change-resistant, risk-allergic, billable-hour-defending buyer in the enterprise. And it hit an $11 billion valuation doing it. Here is how, and what transfers.

The short version

Harvey went from a 2022 experiment by two roommates to $100M ARR in three years, roughly $350M by July 2026, and an $11 billion valuation, selling into elite law firms that are famous for saying no. Winston Weinberg, a former litigator, and Gabriel Pereyra, an ex-DeepMind researcher, made a set of choices that are almost the inverse of a horizontal tool. Go narrow and deep into one vertical where a founder has real credibility. Land a flagship firm, Allen & Overy, and turn it from a logo into co-development and revenue share. Sell to the partners and the firm's economics, not the IT department. Price per seat and let land-and-expand do the compounding, median seats double inside a year. And in a domain where a wrong answer is malpractice, make trust the product, not capability. The model underneath is increasingly commoditized. The moat is the vertical, the workflows, and the distribution into Big Law, and none of that is something a frontier lab can ship in a weekend.

First, the numbers

Founded summer 2022 by Winston Weinberg and Gabriel Pereyra, roommates in LA, after they started playing with GPT-3. Backed early by the OpenAI Startup Fund. It crossed $100M ARR in August 2025, about three years in, hit $195M by year end, and roughly $350M by July 2026. In March 2026 it raised $200M at an $11 billion valuation, co-led by GIC and Sequoia. Its software is used by more than 142,000 lawyers across 1,500 customers in 60 countries, including half the Am Law 100. That is not adoption at the edges. That is the center of the profession.

Choice 1: Go narrow and deep, with founder-market fit

Harvey did not build a horizontal AI assistant and hope law firms adopted it. It built for one vertical, law, and one tier, elite firms, and it had a founder who had lived it. Weinberg was a securities and antitrust litigator. He knew the work, the fear, and the buyer.

That matters more in law than almost anywhere. The workflows are specific, the tolerance for a wrong answer is zero, and the buyer can smell an outsider. A generalist tool reads as a toy. A tool built by someone who has drafted the memo at 2am reads as credible. Steal this: in a high-trust vertical, domain credibility is not a nice-to-have, it is the entry ticket, and founder-market fit is the cheapest version of it.

Choice 2: Turn one flagship partner into co-development

Allen & Overy was first. Not a pilot, a partnership: 3,500 lawyers across 43 offices, exclusive at launch. And Harvey did something smart with it. It did not treat A&O as a logo for the deck. It turned the relationship into co-development and revenue sharing, building the product with the firm rather than at it, and later rolling out agentic agents for the hardest workflows together.

That one partner became the credibility artifact that opened everyone else. When the most respected firm in the world has bet on you and helped build the thing, the rest of the Am Law 100 stops asking "does this work" and starts asking "how fast can we get it." Steal this: one deep, public, co-built flagship is worth more than fifty shallow logos, because it converts the entire peer group behind it.

Choice 3: Sell to the economics, not the IT department

Harvey's motion targets the firm's business, the partners, the billable model, the P&L, not the technology team. Enterprise, seat-based, with 25-seat minimums, annual terms, and sales cycles that run six months or longer. Reported pricing sits around $1,200 to $2,000 or more per seat per month. That is a serious number, and it is deliberate. Harvey is priced like a tool that changes the economics of the firm, and sold to the people who own those economics.

Selling to IT gets you a proof of concept that dies in procurement. Selling to the partners who feel the leverage gets you a mandate. Steal this: sell to whoever owns the P&L your product moves, and price like you belong in that conversation.

Choice 4: Let land-and-expand do the compounding

The entry point is small: a few hundred seats for research, drafting, and diligence. Then it grows. Internal usage data show median seat count doubling within twelve months. That is the quiet engine under the ARR curve, the same shape as the best seat-based businesses, where the initial deal is a foothold and the net expansion is the actual business.

Weekly active users grew fourfold year over year, queries more than five-fold. Usage pulls seats, seats pull revenue, and because the users are lawyers billing against the tool, the expansion is self-justifying. Steal this: design the first deal as a beachhead, and put your energy into the expansion motion, because that is where the compounding lives.

Choice 5: Make trust the product

In consumer software a wrong answer is an annoyance. In law it is malpractice, a sanction, a lost client. So Harvey's real product is not capability, it is trust: accuracy, citations, transparency, human oversight, the ability to defend the work product. This is the cost of being wrong applied at the level of a whole company. The buyer is not pricing how clever the model is. They are pricing what it costs them when it is wrong.

That reframes the whole build. Every feature is downstream of "can a partner stake their name on this output." Steal this: in a high-stakes domain, your moat is the reliability discipline, not the model, and you should be able to explain why the wrong answer is expensive before you explain why the right one is impressive.

Choice 6: Ride the model, own the vertical

Harvey was, in the founders' own framing, built knowing the models would become its competitor. It started on OpenAI, went multi-model, and never bet the company on owning the model. It bet on owning the vertical, the legal-specific workflows, the firm relationships, the trust, the data exhaust from 142,000 lawyers using it daily.

This is the exact opposite answer to Cursor's problem. Cursor was so horizontal that the model supply was the whole game, and the fix was a $60 billion acquisition by a company that owns a lab. Harvey went so deep into one vertical that the model commoditizing underneath it is fine, because the moat was never the model. Two ways to survive "the model becomes your competitor." Own the lab, or own a domain so specific the lab will never go there. Steal this: pick which one you are, on purpose, early.

Where the playbook could break

The rich part first: $11 billion on roughly $350M ARR is about thirty times revenue, priced for a category win that is not finished. Competition is heating up, Harvey acquired Hexus in January 2026, and the incumbents, Thomson Reuters with CoCounsel, LexisNexis, are not asleep, and they own decades of legal content and distribution. The frontier labs could also decide law is worth a vertical push of their own. And the whole edifice still rests on outside models plus a sales motion that is expensive and slow. Depth is a moat and a ceiling at once: it protects the Am Law 100 and makes every adjacent market a fresh six-month sale.

What I am taking into Heidi

Three things. Founder-market fit is not a pitch-deck line, it is the difference between reading as credible and reading as a toy, so lead with the domain, not the model. A single co-built flagship beats a wall of logos, so I would rather have one design partner building the product with me than ten watching from the sidelines. And in any high-stakes domain, and healthcare is the highest, trust is the product: I should be able to explain the cost of a wrong answer before I explain the magic of a right one. Harvey and Cursor together are the clearest lesson of the year on the same threat and two opposite answers to it.

Try this week

Ask one question about your own company. When the frontier model absorbs the capability you are selling, and it will, what is left that is yours. If the honest answer is "the workflows, the trust, and the distribution into a specific buyer," you are playing Harvey's game and you should go deeper. If the honest answer is "nothing but the model," you are playing Cursor's game, and you had better own the lab or start building the vertical moat now, while you still have time to choose.

Sources: Legal AI startup Harvey raises $200M at $11B valuation, CNBC; From roommates to decacorn founders, Forbes; A&O announces exclusive launch partnership with Harvey, A&O Shearman; OpenAI-backed Harvey raises $100M, TechCrunch; Harvey, Sacra.