Summary

Harvey’s wager has expanded from making one lawyer faster to making an entire firm more profitable. In just over three and a half years, it has reached almost 1,000 customers and 500 employees. The product evolved from an “IDE for lawyers” into infrastructure for “orchestration, governance” and thousands of client matters.
The enterprise opportunity extends beyond law firms into the largest buyers of legal services. Harvey recently announced Walmart as a customer and is working with AT&T, Fortune 500 companies, private-equity firms and Global 2000 companies for internal contracting and legal operations, as well as “collaborative tissue” for securely sharing data and work with outside counsel.
Legal agents have a natural operating environment—a client matter—but an open reward-function problem. An associate already behaves like an agent: research case law, summarize it, draft a memo, cite it and incorporate partner feedback. Yet while finding change-of-control provisions can be verified, judging a merger agreement ultimately depends on an experienced partner saying, “Yeah, this looks pretty good.”
AI may eventually compress associate leverage ratios without changing the senior partner’s role much anytime soon. Elad Gil’s challenge: if a firm needs only 20 or 50 associates instead of 100, it may no longer train enough people to identify the next generation of partners. Gabe Pereyra expects lower-level functions to change, while strategy, delegation, expertise and client interaction remain largely durable for now.
Forward-deployed engineering is becoming part of Harvey’s enterprise offering and product-discovery loop. Large banks may lack even a legal document-management system, while clients such as Blue Owl see many processes that could map into generative AI but need help defining them. Repeated bespoke work can enter the platform—and law firms themselves may earn implementation revenue by deploying Harvey for clients.
Harvey explicitly rejects building a law firm, favoring a platform that makes every firm “AI-first.” Pereyra says Harvey spoke with roughly 30 people from Atrium, and also with Sam and Jason; the central challenge was that a law firm and a technology company are two different businesses to build. A single firm would create conflicts and would not scale across a legal market described as roughly $1 trillion and professional services at “something like $3 to $5 trillion.”
Pereyra’s non-consensus call is that AI’s largest gains will come from redesigning organizations, not merely accelerating individuals. “Making someone program 20% faster doesn’t make you build a product 20% faster”; the next layer is coordinating specialized humans and AIs across whole firms—especially as models keep becoming much better than organizations can readily internalize.

Deep dive

1. Harvey is becoming an operating layer for legal organizations

Harvey now serves almost 1,000 customers with 500 employees, just over three and a half years after starting. Its initial pitch was straightforward—AI for large law firms and in-house teams—but the unit of value has steadily expanded from an individual lawyer to the firm and its clients.
The hosts’ baseline challenge—why not simply use Copilot, ChatGPT or Claude?—captures Harvey’s evolution. Early access to GPT-4 made direct model interaction unusually valuable in a text-heavy industry, but lawyers immediately encountered hallucinations and missing context; Harvey therefore spent roughly two years building an “IDE for lawyers” around the models.
Pereyra’s current framing is broader: “The big problem we’re solving is not how do you make individual lawyers more productive.” It is how teams handle a client matter, and how a firm handling thousands of matters becomes “more productive and more profitable”—a problem increasingly defined by orchestration, governance and enterprise-product requirements rather than raw model intelligence.
Distribution is widening through the law firms themselves. Firms began showing Harvey to clients about a year and a half ago, leading to a recently announced Walmart signing and work with AT&T, large private-equity firms and other Fortune 500 and Global 2000 companies that want both internal legal operations and secure collaboration with outside counsel.

2. Legal workflows offer a natural setting for agents

Pereyra’s fund-formation example shows why legal workflow is more than emailing a lawyer. A $1 billion fund may require a 100-page limited partnership agreement, roughly 100 investors and side letters reflecting different investor requirements, including tax implications; every modification can create downstream implications that lawyers must coordinate.
Investment work creates another dense context. Lawyers inspect a company’s data room, understand its contracts to test whether claimed revenue is structured as represented, identify litigation and trace obligations across documents. Pereyra likens this to “understanding a codebase,” except the codebase consists of contracts and legal work whose workflows were difficult to structure before language models.
The agent intuition appeared on Harvey’s first day with GPT-4, when Winston spent 14 hours recreating associate tasks: find case law, summarize it, feed the summary into drafting and cite the result. “You can kind of think of associates as agents”—they receive a partner’s strategy, gather evidence and return a memo.
In Pereyra’s RL analogy, the client matter is the environment. An agent working on a fund formation, acquisition or litigation can navigate a document-management system, inspect a data room, research case law and obtain partner feedback, much as coding agents interact with repositories and attempt to pass tests.

3. Expert judgment is central—and a training bottleneck

Legal RL becomes difficult when outputs move from retrieval to long-form drafting. Finding all the change-of-control provisions can support a conventional benchmark; generating a merger agreement cannot be reduced easily to “good” or “bad.” Pereyra calls the construction of that reward function “one of the really big problems.”
His answer is that “the reward function is the partners.” Firms possess drafts, edits and partner feedback that could provide training material unavailable in public filings. The published SEC document shows the outcome, not the decision process that generated it.
Pereyra argues that mature software engineering eventually has the same verification problem. Unit tests help in the short term, but production success may mean a system served a million users for six months without crashing; a merger’s real test may arrive three years later, when the combined company has avoided unexpected litigation.
Gordon Moody, a partner at Wachtell who joined Harvey early and is now an adviser, embodies the missing expertise. While part of the process in which Michael Dell took Dell private, restructured the business and took it public again, he had to reason across a multiyear restructuring, the largest debt offering of all time and a newly invented financial instrument—“technical understanding of how you architect these things,” not merely relationships.

4. AI may change the associate pyramid before it changes partners

Gil’s pushback—worth keeping: firms traditionally hire perhaps 100 associates knowing only around 10 might become partners. If AI eventually reduces the required cohort to 50 or 20, firms may lose the experiential funnel that reveals who can be trusted with a complex acquisition, even if today’s effect remains augmentation and business expansion.
Pereyra is optimistic that models can accelerate training. Just as programming models let a learner translate Python or ask why code was written a certain way, lawyers can request a merger agreement and then ask, “Why did we structure it that way?” Firms could also turn accumulated partner feedback into training data.
Restructuring cannot be prescribed once for the entire industry. Litigation, large transactions and midsize deals have different workflows, staffing and pricing, so Harvey is working practice area by practice area—for example, sitting with fund-formation teams and their private-equity clients to redesign the operating model.
Senior partners may change less than junior functions. Pereyra compares them with distinguished engineers: both define strategy and abstractions, delegate execution, detect subtle failure modes and interface with clients. “I don’t think the models are doing what they do anytime soon,” though his guess is explicitly that lower-level work will change.

5. Deployment work is turning organizational friction into product

Harvey initially emphasized a horizontal platform and tools such as Workflow Builder, with limited customer-specific development outside very large accounts such as PwC. Training firm-specific models and agents now requires connecting document repositories, billing systems, governance systems and other private data inside customer environments.
Enterprise heterogeneity increases that burden. A large bank may tell Harvey, “We don’t have any document-management system for our legal department. Can you just build us one?” Blue Owl similarly sees numerous processes that might map into generative AI, but wants technical teams alongside it to discover what those systems should be.
The hosts frame forward deployment as a classic Oracle, Dell or IBM enterprise playbook: start with a platform, perform customization around bespoke data and repeatedly absorb common implementations into the core product. Pereyra compares Harvey’s version more closely with Sierra’s agent-engineering program.
An implementation ecosystem is already emerging. Law firms can recommend Harvey to in-house clients, help build their workflows and implement deployments that smaller legal departments cannot staff themselves—potentially creating a new revenue line while deepening Harvey’s distribution.

6. Harvey wants to enable every law firm, not own one

Pereyra says Harvey spoke with roughly 30 people from Atrium, as well as Sam and Jason. The Atrium participants viewed the idea positively, but the challenge was that Harvey would be simultaneously building a law firm and a technology company: “I think you can only do one thing well.”
The larger strategic objective is to help “every law firm become an AI-first law firm,” improving profitability while delivering faster, cheaper service to clients. Owning one firm would create conflicts and prevent the model from scaling.
A major global acquisition illustrates the platform opportunity: around 100 outside-counsel firms might participate because local issues require specialists. Pereyra notes that Harvey has “what, 40 customers in New Zealand,” and says such transactions can also involve investment banks, PwC or a tax adviser and an HR consultancy. Harvey’s target is the secure data-sharing and AI infrastructure connecting those participants across a legal market of roughly $1 trillion and professional services of “something like $3 to $5 trillion.”

7. The next productivity frontier is organizational, not individual

Harvey began before GPT-4 came out, when a side-by-side comparison showed its product working on GPT-4 but not GPT-3.5. Pereyra says he had spent the preceding decade trying to start something like Harvey and was early rather than an overnight success. His AI work made the scaling trajectory legible: once researchers found a general approach, “you can usually just scale and this stuff keeps working.”
Winston supplied the legal insight, while Pereyra supplied capability conviction. Their crucial product choice was to remain open-ended enough to support “any type of legal work,” rather than optimize a narrow GPT-3.5-era task—the legal equivalent of building an assistant for any programming language instead of only checking Python bugs.
Pereyra says legal found its form factor early: upload a document, do something with it and return highly accurate citations. He thinks coding needed stronger base models and better IDE integration, helping explain why coding products emerged later even though GitHub Copilot had already demonstrated demand.
Pereyra’s forward call moves beyond copilots: “Making someone program 20% faster doesn’t make you build a product 20% faster.” He says law firms have 10×ed in size compared with before computers and the internet, and thinks that could happen again through systems that organize specialized humans and models rather than an assumption that one sufficiently smart AI simply does everything.

Vol.78 Talking Venture Capital with Lao Qian—A Cross-Podcast Meetup Our Updated AGI Timeline, 57% Job Automation Risk & US Debt Crisis