Harvey’s bet: make every law firm AI-first, don’t become one

Harvey builds AI for law firms and large in-house teams. Co-founder and president Gabe Pereyra tells No Priors hosts Sarah Guo and Elad Gil — original seed investors in the company — that Harvey is approaching 1,000 customers and 500 employees a little over three and a half years after founding. The roster now includes Walmart, AT&T, PwC, large private equity firms, and some of the world’s biggest law firms. For a technical audience that still pictures legal work as emailing a lawyer and getting a redline back, that scale needs explaining.

The product did not start as a specialized legal system. After raising from OpenAI and getting GPT-4 access, the instinct was simply to give lawyers the model and let them play. Law is so text-heavy that a raw chat interface already delivered a lot of value. Hallucinations and missing context showed up immediately, so the first two years were spent building what Pereyra calls an IDE for lawyers: connecting the model to the context an individual attorney needs in order to be productive.

Over the past year the problem changed. Harvey is no longer optimizing only for a single lawyer’s speed. The live question is how a team on one client matter, and a firm running thousands of those matters, becomes more productive and more profitable. At that scale, Pereyra says, many of the hard parts are orchestration, governance, and ordinary enterprise product work, not just model intelligence.

The buyer set changed with it. About a year and a half ago, law firms started showing Harvey to their clients. Those clients wanted to collaborate more smoothly with outside counsel and to use the same tools inside the legal department. Harvey began building a platform for in-house work that never goes out to a firm — contracting and the long tail of legal operations — plus the “collaborative tissue” of securely sharing data on a large transaction or litigation. Security and privacy across that boundary are now first-class product problems.

Most people, Pereyra says, picture consumer legal: a lease, some comments, a negotiation. That is a different world from what massive firms actually do. A cleaner picture, familiar to anyone in startups, is the work firms do for venture and private equity clients. Those clients mainly raise money and invest it. The legal work on the raise is fund formation: structuring the entity that will hold the capital.

It sounds simple until a large private equity firm has a sovereign wealth fund that needs a tax-driven structure and a pension fund with another set of requirements. The limited partnership agreement can run to 100 pages. There may be a hundred investors, each with a side letter that modifies the deal, and the lawyers have to understand what each change implies. Coordinating that project for a billion-dollar fund is as much project management as drafting. Once the fund exists, every investment repeats a version of the same complexity: a data room, contracts that have to match claimed revenue, and a hunt for lurking litigation.

Pereyra’s analogy for a technical audience is direct: understanding a company this way is like understanding a codebase, except the codebase is contracts. Legal is hard for the same reason programming was hard to tool before language models. Workflows were unstructured. Programmers had an IDE and wrote in whatever language they used; there was no separate “Python tool” or “C++ tool.” Legal was the same. That is why coding and law are seeing traction first: both are extremely text-heavy, and until these models arrived you could not structure the work.

On the coding side, people are starting to call systems agentic: decompose a logic tree, take the next action, check it, go on. Pereyra says Harvey is beginning to do the same. He had done reinforcement-learning research at DeepMind, so when the team got GPT-4 the intuition was immediate — string model calls together, and eventually train a fully differentiable agent. The first day of GPT-4 access, co-founder Winston sat in a room for 14 hours and redid a pile of associate tasks. Pereyra looked at the work and saw a hacky agent loop: look up the case law, summarize it, take the summary, use it to draft.

That session set the product direction. Associates already behave like agents. A partner hands over a high-level case strategy and asks them to find supporting case law, cite it, and write a memo. The systems Harvey is building look more and more like that delegated junior. Coding labs now build RL environments where an agent enters a codebase and tries to pass unit tests. In legal work, Pereyra says, that environment is a client matter: the full context of a fund formation, an acquisition, or a litigation. Models are learning to go into the document management system, the data room, and case-law research, then take feedback from the partner.

When Guo led Harvey’s Series B, she called large customers and found two surprises. Institutions that had been famously hard to sell legal software were adopting fast. And they did not feel threatened. The worry they did raise was longer-dated: if a firm used to hire 100 associates so that 10 might become partners, and it someday needs 50 or 20, is it still hiring enough people to know who should run a complex acquisition?

Pereyra treats that as a training problem as much as a headcount problem. Leverage ratios exist because not everyone becomes a partner; the grind is how a firm learns whom to trust. He is optimistic because of how he learned to program. A decade ago it meant Stack Overflow, one language at a time, and a reluctance to ask “dumb” questions at Google. Models made learning fun: write this in Python, translate it, why is it written this way? Lawyers already do the analogue in Harvey — generate a merger agreement, then ask why it was structured that way.

The larger conversation is how to restructure firms at all. Law firms are collections of practice areas; a litigation shop looks nothing like a large-transactions shop, and the biggest firms are bundles of both. Harvey is sitting practice area by practice area — the fund-formation group and its private equity clients — and rethinking workflows, staffing, and pricing. The goal, Pereyra says, is not a prettier product. It is to make the firms more profitable, which means fitting into the whole business.

Guo compares a partner’s job to consulting: pattern recognition and strategy on one side, sales and client relationship on the other. Pereyra does not think that half changes soon. Harvey is now a large consumer of legal services, and the best partners he has worked with are not doing work models will take over. The analogy is senior engineers. They get paid to set strategy and the right abstractions, delegate the research or the code, and face the client. Lower-level functions change. The partner’s job, he guesses, does not change much.

The technical audience’s version of a rainmaker is network and reputation. Pereyra’s version is closer to a distinguished distributed-systems engineer at Google. Gordon Moodie, a former partner at Wachtell who joined Harvey early and is now an adviser, is the example. Those engineers are valuable because they have architected systems that are not in the public internet and will not enter models for a long time. They can look at a design and say that at this scale it will collapse for a reason that is not intuitive. Moodie has the same picture of a legal entity in his head.

He was involved when Michael Dell took Dell private, restructured it, and took it public again — a multi-year financial and legal rebuild of an enormous business. At one point they had to do what Pereyra calls the largest debt offering of all time and invent a new financial instrument. Talk to Moodie about how to raise a given amount of money for one slice of the deal, and it feels like talking to a senior engineer about how to architect a huge software project. Relationships matter. The scarce asset is the unwritten technical judgment.

That judgment is also what public training data is missing. For a large merger, the public record is an SEC filing: the final result, not the process. The actual dataset would be meetings and emails about the two companies, tasks Moodie handed associates, research coming back, and then a subtle correction — if you structure it this way, this other thing happens. Guo calls it a reasoning trace. Pereyra agrees: improving the models on this work means capturing decision-making, the same way reasoning traces are needed for any hard reasoning task.

Labs are pouring RL compute into coding and math because those domains are highly verifiable. Law is not. Most legal tasks are long-form generation. Some subsets are checkable — go into this data room and find every change-of-control provision — and those can become ordinary datasets. Generating a merger agreement is not binary good or bad. Inside a firm the reward function is a senior partner who has done many of these deals saying the draft looks pretty good. Firms already have the edits and feedback. Harvey is starting to think about training reward models from that. Pereyra’s twist is that real software engineering has the same problem. Unit tests cover the short term. Eventually you deploy, a million users live with it for six months, and it does not crash. A merger is the same: the filing can be correct, and three years later the companies are still combined without surprise litigation. Decade-long systems and decade-long deals are expensive human track records. You cannot unit-test them.

While the research side pushes on agents and reward models, Harvey has stood up a forward-deployed engineering group. Pereyra is careful about the comparison: this is not Palantir going into the codebase to build fully custom software. It is closer to Sierra’s agent-engineering program. Early on Harvey invested in a horizontal platform and a workflow builder so customers could customize without one-off engineering, with some extra work for very large accounts such as PwC. That is no longer enough.

Connecting firm data, billing, and governance systems, then building agents on top, means people have to sit in the customer environment. Fortune 500 legal departments are far less standardized than law firms. A large bank will say it has no document management system for legal and ask Harvey to build one. Blue Owl, a fast-growing private equity firm, wants to sit together and map operations into generative systems without knowing the end shape in advance. For Harvey, that work is also how the roadmap gets written.

Gil calls it the standard enterprise playbook Silicon Valley forgot during the SaaS years. Oracle, Dell, and IBM sold a platform, customized around bespoke data, then folded the repeats back into the product and eventually grew a third-party implementation ecosystem. The twist Harvey is seeing is that law firms are becoming that ecosystem. They take Harvey to in-house clients and offer to build the workflows, because they have the scale and expertise smaller legal departments lack. Some firms already think of it as a new revenue line.

Adoption still surprises him. Harvey’s first customer came through an introduction Gil made to David Wakeling at Allen & Overy, in year one, and went from a small pilot to firmwide investment. Pereyra puts that pace next to Cursor and OpenEvidence: the technology is so useful in text-heavy, knowledge-based industries that lacked tools that a handful of customers were willing to bet early. The speed is still faster than he would have predicted.

A popular question from X was why Harvey does not just start a law firm and compete with its customers. The company researched that path. Early on they talked to 30 people from Atrium, the failed legal-tech firm, and to Sam and Jason — then OpenAI’s general counsel, who had been Y Combinator’s GC when it backed Atrium. People who had worked there thought the idea was good. The hole, once Pereyra dug in, was that you are building two companies: a law firm and a tech company. Product, engineering, AI, and scaling sales is already hard. You can only do one thing well, and running a firm well is a different skill from running a software company.

The larger reason is ambition. The outcome Harvey wants is every law firm becoming AI-first, not Harvey becoming one firm. A single firm gets conflicted out and cannot scale. Gil’s picture of a huge cross-border deal makes the constraint vivid: a Microsoft–Activision-scale merger can involve a hundred outside counsel firms because New Zealand tax issues live with someone in New Zealand. Harvey already has about 40 customers there. Those deals also pull in investment banks, a tax adviser such as PwC, and an HR consultancy for headcount. The platform job is secure collaboration and data sharing across all of those entities, plus AI systems that can run on the project.

Pereyra puts the market in round numbers: legal services around a trillion dollars, professional services something like $3 to $5 trillion. Harvey’s expertise, as he frames it, is the product and the AI infrastructure that lets professional-service providers and their clients work together. Give that infrastructure to every firm rather than compete with them, because he does not think you can do both.

Guo notes how many deeply technical people are now building giant companies in industries far from the lab. Pereyra’s own surprise is less a single shock than a mental-model change. For the decade before Harvey he mixed AI research with attempts to start companies, always largely as an individual contributor. Once this one started working, he had to relearn what company they were building, how to do it at scale, and how to operate — from Winston and himself in an Airbnb to 500 people in three and a half years.

They started before GPT-4 shipped. Guo remembers a side-by-side of GPT-3.5 and GPT-4: the work ran on 4 and not on 3.5. In mid-2022 almost nobody was founding “AI plus legal.” Pereyra’s conviction came from two places. He had been training large language models at Meta and had watched GPT-1, 2, and 3. For a decade the field had specialist systems; nobody had the general solution. His pattern was that once you see how to do it, you usually just scale and it keeps working. Three and 3.5 were getting interesting but were not quite there. He knew a lot of the people at OpenAI and bet they might be the ones to crack it.

The other half was Winston. He was a first-year associate who talked about firm structure the way a partner might — what each partner was doing, why the strategy looked that way — and was in the process of convincing partners to leave and start a firm with him. Pereyra had never planned to found a company with him, then decided the person plus the domain was the thing to do. When GPT-4 arrived, the timing finally matched. They also refused to shrink the product to what 3.5 could do. A Python linter you could have shipped on 3.5 would never have become Cursor. Harvey assumed models would help with any legal task, so the product stayed open-ended enough to grow into other professional services.

Gil asks why coding’s winners — Cursor, Cognition, Claude Code — showed up about two years after GitHub Copilot, while Harvey had a product almost immediately. Pereyra’s guess is form factor. Legal’s first feature was almost embarrassingly simple and missing from other tools: upload a document, do something with it, and produce accurate citations. Lawyers saw their job. Coding needed stronger base models and a tighter IDE integration. The first version of Harvey is one Pereyra built himself. He still does not know React; he used JavaScript and asked GPT-4 to help, which is also how he became convinced the same loop would work on Winston’s legal tasks. The company is now hiring strong engineers, more forward-deployed engineers, frontend and product people, AI researchers, and has just named a site lead for a New York office.

Asked for a view that is still not mainstream, Pereyra first pushes back on the overnight-success story. He spent the decade before Harvey trying to start a company like Harvey; he was early, then the timing was right. Silicon Valley, he thinks, now understands the capability curve in outline and still underestimates how much better the models will get. The more interesting shift is what you do with that. Twenty years of SaaS were about using software to make massive organizations. Law firms have already 10x’d in size relative to the pre-computer, pre-internet era. He thinks that happens again, in a different way than the last twenty years.

People still talk about copilots and individual productivity. Harvey is starting to think about organizational productivity: the infrastructure a company needs so it can develop software and product faster, and the same analogy in legal. Making someone program 20% faster does not make you build a product 20% faster. Guo compares it to Figma’s jump from an individual designer to a design team working together. Pereyra wants that layer for law, for code, and for other verticals, with AI sitting on top of the collaboration rather than replacing it.

The naive story is that models get smart and do all of this. That is not how it evolves, any more than hiring 100,000 people — Guo corrects him: Walmart is more like three million — produces a Walmart. Organization is the thing. Guo sees the same pattern in AI-driven roll-ups and implementation work: the hard part is people management and workflow, much less whether you can build the model. Harvey is starting to work with private equity firms on exactly that question. Pereyra thinks it will be a very interesting space. As for the rumor that he still sleeps on an air mattress: he has a good mattress and no bed frame. It broke on the move from Los Angeles to San Francisco, UPS would not bring a replacement without mover’s insurance, and he has not had time to deal with it. The company still has other problems to solve.