No PriorsNo Priors Ep. 142 | With Harvey Co-Founder and President Gabe Pereyra
CHAPTERS
- 0:00 – 0:36
Harvey at a glance: AI platform for law firms and in-house legal teams
Gabe Pereyra explains what Harvey builds and who it serves today, emphasizing rapid scaling in customers and headcount. He frames Harvey as enterprise AI infrastructure purpose-built for legal work rather than a generic chatbot.
- •Harvey builds AI for large law firms and in-house legal departments
- •Scale snapshot: approaching ~1,000 customers and ~500 employees
- •Started ~3.5 years ago and scaled quickly
- •Focus on serving complex, high-stakes legal work at enterprise scale
- 0:36 – 2:03
Why Harvey isn’t just ChatGPT: from lawyer IDE to firm-wide orchestration
Gabe contrasts early “give GPT-4 to lawyers” experimentation with the need to solve sharp edges like hallucinations and missing context. He describes the strategic shift from individual productivity to team and firm-level workflows, governance, and profitability.
- •Early value came from simple interaction with frontier models in a text-heavy domain
- •Key limitations: hallucinations and lack of matter-specific context
- •Initial product direction: an ‘IDE for lawyers’ around models
- •Current direction: orchestration, governance, and enterprise controls
- •Goal expands to firm profitability across thousands of matters
- 2:03 – 3:21
Expanding beyond law firms: enabling client–counsel collaboration for enterprises
Harvey’s growth into Fortune 500 and Global 2000 customers is driven by law firms introducing the product to their clients. The platform evolves to support in-house work (like contracting) and secure collaboration with outside counsel.
- •Law firms began showing Harvey to their clients, pulling demand into enterprises
- •Examples include Walmart, AT&T, private equity, and large banks
- •Building tools for internal legal ops (e.g., contracting and long-tail workflows)
- •Secure data sharing across organizations becomes a core technical problem
- •Collaboration layer between in-house teams and external counsel
- 3:21 – 6:20
What legal workflows really look like: fund formation, diligence, and unstructured work
Elad prompts a concrete explanation of “workflow” in biglaw, beyond emailing a lawyer. Gabe walks through complex VC/PE legal work—fund formation, side letters, diligence—framing it as understanding a ‘contract codebase’ with heavy project management and weak standardization.
- •Consumer legal differs sharply from large-firm, high-complexity matters
- •Fund formation complexity: LPAs, side letters, investor constraints (tax/structure)
- •Deal work mirrors codebase comprehension: many interdependent documents
- •Diligence and data rooms: validating contracts, litigation exposure, revenue structure
- •Core challenge: workflows are often unstructured and hard to formalize
- 6:20 – 8:28
Agentic AI in law: associates-as-agents and matter-centric environments
Sarah introduces agentic workflows; Gabe connects this to how associates execute partner directives through multi-step research and drafting. He describes Harvey’s trajectory toward agents that can navigate document systems, data rooms, and research sources within the context of a client matter.
- •Associates naturally operate like agents executing multi-step tasks
- •Early internal behavior: chaining research → summarization → drafting
- •Agents interact with matter context: DMS, data rooms, and case law research
- •Feedback loops with partners become part of the agent training setup
- •Legal ‘RL environment’ analogy: the client matter as the interactive world
- 8:28 – 13:34
How AI changes law firms: training partners, restructuring leverage, and pricing
The discussion turns to the long-term evolution of the law firm pyramid and partner training when fewer associates may be needed. Gabe suggests AI can accelerate learning (like coding tutors) and highlights the opportunity to use internal firm feedback and edits as training signal for better systems and business transformation.
- •Concern: shrinking associate classes may affect partner pipeline development
- •AI can make learning faster via explanations and ‘why structured this way’ queries
- •Internal partner feedback and edits are valuable training data
- •Transformation varies by practice area, region, firm type, and client mix
- •Value shifts from tooling alone to enabling holistic firm restructuring
- 13:34 – 17:19
Expert reasoning as the missing dataset: Gordon Moody analogy and decision traces
Elad asks about the analogy between elite partners and distinguished engineers. Gabe explains that the crucial missing piece in public data is the expert decision-making process—meetings, emails, delegation, and subtle corrections—not just the final filings. This motivates capturing reasoning traces to improve systems.
- •Elite partner value mirrors senior systems design expertise: anticipating non-obvious failure modes
- •Example: complex Dell take-private/restructure and bespoke financial instruments
- •Public artifacts (e.g., SEC filings) omit the real decision process
- •Training needs the intermediate reasoning, delegation, and feedback loops
- •Reasoning traces and expert edits become core learning signals
- 17:19 – 19:46
Reinforcement learning for legal work: reward functions and “verifiability” challenges
The conversation dives into why RL is harder in law than in coding or math: outputs are long-form and correctness isn’t binary. Gabe notes some tasks are verifiable, but high-stakes drafting isn’t, so partner judgment and internal edit histories may serve as proxy rewards—similar to real-world software engineering where unit tests aren’t enough.
- •Legal drafting quality is hard to score with binary correctness
- •Some subsets are measurable (e.g., extracting specific clauses)
- •Key open problem: constructing reward functions for complex documents
- •Partners are the ultimate evaluators; firm edit trails can approximate rewards
- •Analogy: mature software engineering success is only proven over time in production
- 19:46 – 23:45
Forward-deployed engineering at Harvey: connectors, customization, and enterprise reality
Elad questions why Harvey is building an FDE motion; Gabe explains it’s not full bespoke software but necessary to integrate messy enterprise systems and data. As Harvey expands into less standardized enterprises, the team helps map operations into AI workflows and feeds insights back into the product roadmap.
- •FDE model is closer to ‘agent engineering’ than full Palantir-style custom builds
- •Integration needs: connecting data sources, billing, governance, and document systems
- •Enterprises are less standardized than law firms, increasing implementation work
- •Customers often want help redesigning ops to leverage gen AI
- •FDE work informs roadmap across verticals and repeatable platform features
- 23:45 – 25:28
Adoption flywheel: customer success, fast rollouts, and firms implementing for clients
The hosts highlight how quickly conservative legal institutions adopted Harvey and how customer success enabled broader transformation. Gabe and Sarah note a new pattern: law firms themselves begin implementing Harvey-based workflows for their clients, creating a services-led distribution and new revenue lines for firms.
- •Rapid adoption surprised even the team; pilots moved to firm-wide deployments
- •Customer success is central to business transformation outcomes
- •Legal and other knowledge-heavy verticals adopt faster when value is obvious
- •Law firms begin selling/implementing Harvey for in-house clients
- •Implementation by firms becomes a new revenue stream and accelerates platform spread
- 25:28 – 27:24
Why Harvey won’t build a law firm: scaling the platform beats competing with customers
Responding to a common question, Gabe explains why operating a law firm and a tech company simultaneously is strategically and operationally difficult. Harvey’s goal is to make many firms AI-first, avoid conflicts, and enable ecosystem-wide collaboration rather than becoming a single provider.
- •Atrium lessons: building a law firm + tech company creates dual-execution burden
- •Harvey focuses on being infrastructure for the ecosystem, not a competitor
- •Platform approach scales better and avoids conflict-of-interest constraints
- •Aim: improve profitability and service outcomes across many firms and clients
- •Bigger opportunity is enabling AI-first operations across the market
- 27:24 – 29:25
The scope of legal (and professional services): multi-firm matters and secure collaboration
Elad and Gabe discuss how large deals involve dozens to hundreds of specialized firms and advisors globally. This expands Harvey’s ambition toward a broader professional-services collaboration platform with secure data sharing and cross-entity AI deployment as core capabilities.
- •Large M&A can involve ~100+ outside counsel firms across jurisdictions
- •Matters also involve banks, tax advisors, consultants, and HR specialists
- •Core platform need: secure collaboration and data sharing across entities
- •Legal is ~trillion-dollar; professional services even larger (multi-trillion)
- •AI systems must operate across complex, multi-party projects
- 29:25 – 37:23
Building Harvey in the gen-AI rise: capability conviction, product form factor, and timing
Sarah asks about founder surprises and early conviction; Gabe credits prior lab exposure to LLM scaling and the insight from co-founder Winston’s deep legal intuition. They discuss why legal’s initial form factor was clearer—documents and citations—and how a ‘capability-first’ belief shaped an open-ended platform from early days.
- •Founder shift: from IC research mindset to operating/scaling a 500-person company
- •Conviction came from watching LLM progress (GPT-1→4, LaMDA) and scaling laws
- •Winston’s understanding of firm structure and workflows grounded product direction
- •Early killer features: upload docs + do work; accurate citations
- •Belief: models would keep improving, so keep the product broad and extensible
- 37:23 – 44:17
Hiring and closing: roles, offices, and a lighthearted outro
Gabe shares what Harvey is hiring for—especially strong engineers, FDE, AI, product, and office scaling. The episode ends with playful banter (pull-ups, TikTok, bed frame story) and the hosts’ sign-off.
- •Hiring focus: strong engineers, FDE, front-end/product scaling, and AI roles
- •New York office site lead and broader geographic scaling
- •Cultural moment: pull-up jokes and a ‘24-hour challenge’ aside
- •Personal anecdote: bed frame logistics and founder time constraints
- •No Priors outro with where to follow and find transcripts
