Gabriel Pereyra

The through-line

Everything Pereyra says about Harvey traces back to a single dependency: what the models can actually do determines what is worth building, and everything else is the work of closing the distance between a model that impresses and a system that survives high-stakes legal work. He and Winston Weinberg started in 2022 with an interest in consumer access to justice, and he is direct that the idea did not survive contact with the technology of the moment: "when we first started the company, they just weren't there" [1]. Early access to GPT-4 through the relationship with OpenAI changed the answer, and the product went where the pull was rather than where the founders' original thesis pointed [1].

The second half of the through-line is that the model was never the product. His summary of what Harvey learned from its first large deployment is that "there's just a huge gap from like an impressive demo to like something that works in production for high stakes legal work" [1], and most of what the company has built since sits in that gap: data connections, case law coverage, saving and sharing, security, privacy, regional infrastructure. The conversation frames where this is heading as Harvey becoming an operating system for legal work, with agentic AI handling complex workflows and a shift from one-off tasks to client matter-centric work [1].

On waiting for the models to be good enough

Pereyra's read on capability was calibrated by having a practising litigator in the apartment. Because Weinberg had been doing litigation work at O'Melveny & Myers, "we had a good sense of how good the models need to be to be useful in big law" [1][2], and by that standard the early models failed. What shifted was reasoning, and he says he could see it from two directions at once: watching Weinberg spend roughly fourteen hours in his room re-running the legal work he had done at his firm, and using the same models himself to write code, so that "by analogy, you could kind of see how much better the reasoning was" [1]. The distinction that mattered was not more data in the prompt but the ability to reason over the same context and reach a correct conclusion, which earlier models could only manage on tightly rule-bound fact patterns [1].

On following market pull rather than a thesis

Once there was something to show, the founders demoed it to every kind of lawyer they could reach, including law firm lawyers, startup lawyers and in-house teams. Pereyra says the surprise was where it landed: "where we actually found the most traction was big law firms, which I thought was pretty interesting. But that market pull, I think, kind of put us in this direction" [1]. One early in-house demo, on a memo about data localization laws for Snowflake's expansion into Indonesia, worked partly because the general counsel had already done the work himself and could verify the output, which he describes as one of the early light bulb moments [1]. That relationship became structural rather than anecdotal: the general counsel joined as roughly Harvey's seventh hire, over objections from investors who expected that role at three or four hundred people [1]. The same instinct now runs downmarket, with Harvey selling to firms of around ten attorneys, and Pereyra's colleague notes the technical direction is easier that way because the architecture and security demands are lighter than at the top of the market [1].

Reaching anyone at all in the beginning was brute force. He and Weinberg messaged lawyers on LinkedIn at a volume that got them blocked repeatedly: "we get blocked basically every day of like you reached out to too many people and then we'd have make like another account" [1]. Almost nobody responded, and no large firm did, until a warm introduction produced a live demo that reached David Wakeling and led to the A&O Shearman deployment [1]. When the press announcement landed and the inbound arrived, the company was four people, one of whom had started the day before, and had just rolled out to a four-thousand-person organisation [1]. The waitlist they used to absorb the demand is something the founders now question, with the retrospective view that introducing the team and its actual size publicly first would have been better than being introduced through a customer's announcement [1].

On the gap between demo and production

Pereyra describes the first big deployment as a machine for exposing rough edges. Real use immediately surfaced everything the model did not do on its own: connecting to firm data, missing case law, saving documents, sharing with teammates, plus security, data privacy and ethical walls [1]. He puts the scale of it plainly: "We just got, I mean, just from Alan and Ovary, like, a three-year product roadmap of you want to be able to do all of these things" [1]. He treats that flood of requirements as the point rather than the cost, describing the same period as a source of enormous feedback and product intuition from a large body of real users [1]. Working with a scaled customer from day one meant the architectural problems arrived early rather than later, including multiple model calls across a thousand-document corpus in Vault, which required a different architecture than the assistant product [1].

On infrastructure as a real constraint

One of his more counterintuitive early lessons was that compute was not simply available. Against the assumption that a Google or a Microsoft has unlimited cloud capacity, Harvey ran into hard limits: "we don't have enough model capacity, we don't have enough embedding capacity" [1]. That constraint has persisted as the company globalised, and he frames the current version of it as wanting models in regions where providers have not stood up regions yet, which makes solving deployment globally a continuing engineering problem [1]. The regulatory side compounds it, with data processing rules, particularly around financial data, pushing Harvey to stand up Azure instances country by country across fifty-nine countries so client data stays in jurisdiction [1].

On who inside a firm adopts first

The adoption pattern ran opposite to the usual expectation. Enthusiasm came from individual lawyers rather than from firm leadership, and it was not sorted by practice group; leadership across the market only really pushed in 2025 [1]. Pereyra's own conviction came from a specific and unexpected signal: senior partners telling him the product was useful to them personally, "not, this is useful for my associates to use, like I'm using this technology" [1]. He read that as a ceiling worth aiming at, reasoning that if some of the best senior partners were getting value, the goal became getting the product to a state where that experience is available to everyone and not only to people already deep in language models [1].

On what has and has not changed

Asked about being named among the new billionaires of the AI boom, the founders' answer was that almost nothing about their lives has changed: the same shared San Francisco apartment they started in, a third roommate who also works at the company, the same hours, and essentially no furniture beyond boxes and a mattress on the floor [1]. The company around them is a different matter, having grown to more than a thousand customers, an $8 billion valuation and hundreds of employees, over twenty percent of whom are lawyers, built by two founders who came in without management experience [1].

Takeaways

• Capability gates strategy: the original consumer access-to-justice idea was shelved because the models "just weren't there," and early GPT-4 access is what made big law viable [1].
• The measurable jump was reasoning over supplied context, not more context; Pereyra triangulated it by watching a litigator re-run real firm work and by using the same models to code [1].
• Big law traction was a surprise the founders followed rather than a plan: "that market pull, I think, kind of put us in this direction" [1].
• The hard part is after the demo: "there's just a huge gap from like an impressive demo to like something that works in production for high stakes legal work" [1].
• A single large early customer generated "a three-year product roadmap" covering data connections, case law, sharing, security, privacy and ethical walls, and forced scaling problems to surface early [1].
• Compute is not infinite: capacity limits on models and embeddings, plus regional data processing law, drove country-level infrastructure across 59 countries [1].
• Adoption came bottom-up from individual lawyers and innovation teams, with firm leadership only pushing hard from 2025; senior partners using the tool themselves gave Pereyra his strongest conviction signal [1].
• Distribution began as brute-force LinkedIn outreach that repeatedly got the founders blocked, and the breakthrough came from a warm introduction leading to a live demo [1].