First Block: Interview with Gabe Pereyra, Co-Founder and President of Harvey

Harvey’s growth story hinges on a shift from early, high-value pilots to building the “broader machine” enterprises need: not just AI outputs, but the infrastructure, security, and collaboration workflows that let lawyers safely use models on real client matters. Gabe Pereira, co-founder and president of Harvey, describes landing unusually large seven-figure deals early for an enterprise startup—then learning that winning contracts wasn’t the same as having the operational muscle to scale. The key lesson was moving from a product that drafts or assists with legal work to an enterprise system that can orchestrate tools, manage data scope, and support long-running tasks across a firm.

Harvey began with co-pilots for lawyers and evolved into a platform for large law firms and their clients to collaborate on complex matters like transactions and litigation. The company’s origin traces to Pereira’s work on Meta’s large language model team during the release of GPT-3.5 and GPT-4-era capabilities, alongside Winston’s background as a litigation associate at a top firm. Their early traction came from identifying “customer zero” and “customer one” inside the legal ecosystem: Winston’s network and a senior technology partner at a major firm who quickly rolled the system out firmwide to thousands of lawyers after a pilot.

A recurring theme is that Harvey didn’t start as a narrow automation tool. Early demos that could draft NDAs sounded simple to lawyers, but users needed a mental model closer to “talk with this system and figure out what to do,” especially once they hit limits like missing access to case law or internal firm data. As model reasoning improved—particularly with step-change unlocks from stronger reasoning models—Harvey shifted toward agents that can chain tool calls for tasks such as legal research, reviewing emails, and synthesizing findings for litigation. Accuracy expectations also differed from consumer legal software: large law firms operate with layered review processes, similar to software engineering, where outputs are checked by multiple associates before reaching production.

Instead of relying primarily on custom model training, Pereira emphasizes that most value comes from infrastructure: an “IDE for lawyers” that connects the relevant tools and data for a specific matter (acquisition, litigation, fund formation) and constrains the agent to the correct scope. Domain improvements—post-training, retrieval-augmented generation, and other techniques—still matter, but the company’s emphasis is on orchestration and secure workflow design. He also argues that the next frontier is collaborative training between enterprises and their law firms, since client data can’t be reused broadly; the likely future is “inverted” training where legal models learn from the combined work between a specific enterprise and its counsel.

On go-to-market, the legal industry’s buyer-user split is compounded by an additional stakeholder: the law firm’s clients. Harvey’s early sales leaned on founder-led vision work with CIOs, partners, and clients to align on how teams would collaborate. Pereira’s advice to founders is to use models immediately—because the obvious applications are only the first wave—and to build domain expertise by pairing technical founders with real industry mental models. If rebuilding from scratch, he would invest earlier in foundational security and retention architecture, which became necessary as enterprise deals scaled, but could now be designed with clearer abstractions and collaboration workflows in mind.