This week, we talk with Gabe Pereyra, President and co-founder at Harvey, about his path from DeepMind and Google Brain to launching Harvey with Winston Weinberg; how a roommate’s real-world legal workflows met early GPT-4 access and OpenAI backing; why legal emerged as the right domain for large models; and how personal ties to the profession plus a desire to tackle big societal problems shaped a mission to apply advanced AI where language and law intersect.
Gabe’s core thesis lands hard, “the models are the product.” Rather than narrow tools for single tasks, Harvey opted for a broad assistant approach. Lawyers live in text and email, so dialog becomes the control surface, an “AI associate” supporting partners and teams. Early demos showed useful output across many tasks, which reinforced a generalist design, then productized connections into Outlook and Word, plus a no-code Workflow Builder.
Go-to-market strategy flipped the usual script. Instead of starting small, Harvey partnered early with Allen & Overy and leaders like David Wakeling. Large firms supplied layered review, which reduced risk from model errors and increased learning velocity. From there the build list grew, security and data privacy, dedicated capacity, links to firm systems, case law, DMS, data rooms, and eDiscovery. A matter workspace sits at the center. Adoption rises with surface area, with daily activity approaching seventy percent where four or more product surfaces see regular use. ROI work now includes analysis of write-offs and specialized workflows co-built with firms and clients, for example Orrick, A&O, and PwC.
Talent, training, and experience value come next. Firms worry about job paths, and Gabe does not duck that concern. Models handle complex work, which raises anxiety, yet also shortens learning curves. Harvey collaborates on curricula using past deals, plus partnerships with law schools. Return on experience shows up in recruiting, PwC reports stronger appeal among early-career talent, and quality-of-life gains matter. On litigation use cases, chronology builders require firm expertise and guardrails, with evaluation methods that mirror how senior associates review junior output. Frequent use builds a mental model for where errors tend to appear.
Partnerships round out the strategy. Research content from LexisNexis and Wolters Kluwer, work product in iManage and NetDocuments, CLM workflows via Ironclad, with plans for data rooms, eDiscovery, and billing. Vision extends to a complete matter management service, emails, documents, prior work, evaluation, billing links, and strict ethical walls, all organized by client-matter. Global requirements drive multi-region storage and controls, including Australia’s residency rules. The forward look centers on differentiation through customization, firms encode expertise into models, workflows, and agents, then deliver outcomes faster and at software margins. “The value sits in your people,” Gabe says, and firms that convert know-how into systems will lead the pack.
Listen on mobile platforms: Apple Podcasts | Spotify | YouTube
[Special Thanks to Legal Technology Hub for their sponsoring this episode.]
Email: geekinreviewpodcast@gmail.com
Music: Jerry David DeCicca
Determining Generative AI’s effect on law firm profitability requires a bit more than a “back of the napkin” calculation with rough percentages based on keywords in time entries, as we did when roughly calculating the effect on revenue.