The founding scene is a San Francisco apartment, and the division of labor is almost suspiciously neat. One roommate, Gabe Pereyra, worked on large language models at Meta. The other, Winston Weinberg, was a young litigator at O’Melveny & Myers. Pereyra showed Weinberg what the models could do. Weinberg showed Pereyra the documents, searches and drafts that filled a lawyer’s day. Between those demonstrations sat the idea that became Harvey.
It had not been a partnership drawn up in advance. The two met through mutual friends in San Diego, became close, then roommates. Pereyra had spent years trying to find a consequential use for AI and had considered personalized education. Weinberg knew a profession whose most valuable work often arrived as an intimidating pile of language. When they compared notes, legal work began to look less like an odd niche and more like a demanding test.
There is a delightful moment in Pereyra’s account of the early pitch. He brought the idea to people at OpenAI. Asked whether he had a co-founder, he looked at Weinberg and said yes. Weinberg, Pereyra recalled, answered that he would leave his job the following week. Pereyra was surprised by the speed. The rest of the story, naturally, took longer.
Before the apartment, a decade of questions
Pereyra’s route into law did not run through a law school. He studied computer science at the University of Southern California, then pursued doctoral study in neuroscience at Oxford with DeepMind funding. He later worked as a research scientist at DeepMind and a machine learning engineer at Meta. The biography helps explain both sides of Harvey’s origin: he knew the machines, and he needed someone who knew the job.
In an interview, he said the aim behind his AI research had always been to start a company. The application was unsettled. He left DeepMind to try a personalized education startup and built other experiments. At Meta, where he worked on large language models, the technology was moving quickly enough that old ideas could be reconsidered. His roommate supplied a new one, with an unusually convenient first test user.
“The goal of doing AI research was always to start an AI company.”Gabe Pereyra
The pair began with a general assistant for lawyers. That decision can sound broad to the point of recklessness, but Pereyra’s reasoning was specific. The early models could help with many pieces of a lawyer’s work without completing any one task perfectly. He thought a flexible tool would let lawyers find useful applications for themselves. In demonstrations, they would show a draft of one document, and the viewer would assume the product was a specialist in that document alone. Then the founders would show another task.
The company’s name carried a private joke into a public market. Pereyra has said that he and Weinberg were fans of Suits; Harvey was a nod to the television lawyer Harvey Specter. The name had no affiliation with the show. A fictional closer gave the software a memorable introduction, while real lawyers soon supplied more demanding questions than a scriptwriter could.
The customer who changed the scale
At first, the founders showed Harvey to many kinds of lawyers: general counsel, startup advisers and large firms. The decisive early relationship came through David Wakeling, a senior partner and technology leader at Allen & Overy. Pereyra has described a rapid firmwide rollout after a pilot, reaching roughly 3,500 lawyers. This gave Harvey a customer large enough to expose problems that no apartment demonstration could have revealed.
A solo user can paste a passage and judge the response. A major law firm has thousands of users, many clients and strict boundaries around client information. The assistant needs access to case law, then to the firm’s own documents. Its output must travel from an associate to a partner, sometimes across a deal team and eventually to a client. Each step has permissions, confidentiality and accountability attached. Pereyra says users pulled much of the product roadmap out of those rough edges.
The first breakthrough also created a new kind of workload for the founders. Pereyra has spoken about early seven-figure contracts arriving before the company had much of a sales organization. In another interview, the pair remembered sending thousands of LinkedIn messages in search of anyone willing to see the product. Those two scenes belong together. One is the scramble to open a door; the other is the work that starts once a large customer walks through it.
A lawyer’s skepticism is part of the product
Pereyra talks about legal AI as someone who learned a profession by watching its objections closely. In a 2026 conversation about trust, he emphasized the unglamorous architecture of a law firm deployment: security, privacy, governance and ethical walls. The work of an AI system, he said, should be tracked and reviewed in much the same way that senior lawyers supervise junior work. Law firms already have habits for citations, decomposing a problem and checking a result. Those habits are design material.
His view of billing is similarly resistant to the easy prediction. AI may reduce the time some tasks take, but Pereyra is skeptical that the billable hour simply vanishes. Hourly billing, he argues, is a standardized way to price work whose final size and value are difficult to know in advance. A fund formation matter, a merger and a lawsuit are different businesses inside the same profession. He sees hybrid fees emerging where a defined component, such as diligence, can be priced separately while the rest remains hourly.
This is a subtle admission from the president of a company selling speed. The model can alter the labor involved in a task. It cannot, on its own, settle how a firm and its client divide risk, judge quality or agree on a fee. For Harvey, those surrounding decisions are part of the product problem. Pereyra’s story has moved from making a response appear on a screen to making that response useful inside institutions with long memories and exacting clients.
“I’m actually somewhat skeptical that this is the death of the billable hour.”Gabe Pereyra
The question now is how to show the work
By 2026, Pereyra’s public writing had turned toward agents: systems asked to carry out linked steps across much larger collections of material. In September, he co-authored research with Harvey and Baseten colleagues on AI agents handling merger and acquisition diligence. One test environment involved data rooms with as many as 5,000 documents and 80 million tokens of text. The agent’s assignment was a diligence memo with findings, citations and next steps. This is a world away from asking a chatbot to draft an NDA.
The team reported that a system dividing the review into bounded tasks improved the share of rubric criteria passed across seven models, from an average of 23.3% to 62.4% in its test. Those figures describe a research benchmark, not a guarantee about a client matter. Their interest lies in the shape of the work. A human deal team also splits a giant data room into categories, checks findings against evidence and assembles a final memo. Pereyra’s recent work asks how software can follow that discipline at scale.
Harvey has become a large enterprise while these questions grow more technical. In September 2026 the company announced a $550 million funding round at a $15.5 billion valuation and said 80% of Am Law 100 firms use its products. Those are company figures, and they describe adoption rather than the quality of any individual legal answer. The distinction matters. Pereyra’s own interviews return repeatedly to the need for review, governance and firm knowledge as the software spreads.
He has also said the company eventually wants what it builds to help consumers, while its present focus remains on law firms and enterprises. The legal boundary is real: who may give advice, who bears responsibility for it, and who can check the work. That makes the professional customer a practical place to build. A lawyer can ask a model to help and still remain the lawyer.
The apartment origin has a tidy cast, but its lesson is less tidy. Pereyra knew AI. Weinberg knew litigation. The first version showed possibility; early lawyers supplied the missing requirements. Today’s research asks whether agents can examine enormous records and show why they reached a conclusion. The distance from the roommates’ first demonstration to that question measures how much of legal work happens around the sentence on the page. Pereyra’s career has become an effort to build for that surrounding work.