Scaling Legal AI and Building Next-Generation Law Firms with Harvey Co-Founder and President Gabe Pereyra

Gabe explains Harvey's growth to nearly 1,000 customers and 500 employees in 3.5 years. The product evolved from a simple GPT-4 interface for lawyers to a comprehensive platform solving team collaboration and law firm profitability, not just individual productivity. They're expanding from law firms into Fortune 500 in-house legal teams like Walmart and AT&T.

• •Started with GPT-4 early access from OpenAI, building an IDE for lawyers around LLMs
• •Past year focused on making entire law firms more profitable across thousands of client matters, not just individual lawyer productivity
• •Expanding to Fortune 500 in-house teams who want to collaborate with external counsel and handle internal legal operations
• •Core problems shifted from model intelligence to orchestration, governance, and enterprise product challenges at scale

Detailed breakdown of what large law firms actually do, using venture capital fund formation and M&A as examples. These workflows involve 100+ page documents, complex side letters, data room analysis, and contract review - analogous to understanding a massive code base. The unstructured nature of legal work made it difficult to build tools until LLMs arrived.

• •Fund formation involves structuring entities with complex investor requirements (sovereign wealth funds, pension funds) requiring 100+ page LPAs and numerous side letters
• •M&A due diligence requires analyzing entire 'code bases' of contracts, revenue structures, and litigation risks
• •Legal workflows are unstructured like programming - you can't build specialized tools for each task until you have general models
• •The text-heavy, knowledge-based nature of legal work makes it ideal for LLM transformation, similar to programming

Harvey is deploying agentic workflows where AI breaks down complex legal tasks into subtasks, similar to how associates work under partners. They're building RL training environments using actual client matters (fund formations, acquisitions, litigations) where models learn to search documents, research case law, and get partner feedback - but verification remains challenging.

• •Winston manually created agentic workflows on day one with GPT-4, chaining research → summarization → drafting tasks
• •Associates function as agents: partners give high-level strategy, associates research case law and write memos
• •Client matters serve as RL environments where models interact with document management systems and data rooms
• •Major challenge: legal tasks are long-form text generation without binary verification - reward functions must come from senior partner feedback, not unit tests

Discussion of how AI will transform law firm structure, training, and profitability. While partner roles won't change much (they delegate and interface with clients), the challenge is training future partners with fewer associates. Harvey is helping firms think practice-area by practice-area about workflows, staffing, and pricing to make firms more profitable.

• •Programming became easier to learn with LLMs - same happening in legal where lawyers can ask 'why structure it this way?'
• •Big opportunity: using internal partner feedback data to train next generation, not just improve models
• •Law firm transformation varies by practice area, region, size, and specialty - requires customized approach
• •Harvey's goal is making law firms more profitable holistically, not just selling product - requires understanding their entire business

Gabe draws a powerful analogy between top legal partners like Gordon Moody (ex-Wachtell, now Harvey advisor) and distinguished engineers at Google. Both have decade-long experience architecting complex systems with non-public knowledge that won't be in models. The value is predicting what will break at scale and understanding subtle implications - this reasoning trace data is what's missing from public filings.

• •Gordon Moody worked on Michael Dell taking Dell private and public again - multi-year complex restructuring requiring largest debt offering ever
• •Like senior engineers who know distributed systems will collapse in non-intuitive ways, top partners understand legal entity architecture
• •Public data only shows final SEC filings - missing the decision-making process, partner feedback, and reasoning traces
• •The valuable training data is in meetings, emails, and associate tasks showing why certain structures work or fail

Harvey launched a deployed engineering force (not full Palantir-style FTE) to help customers implement AI systems, connect data sources, and build custom workflows. This follows the standard enterprise playbook of Oracle, IBM, Dell - platform plus customization that eventually becomes productized. Law firms are even offering this as a service to their clients.

• •Not building custom software per customer, but helping connect business systems, billing, governance, and document management
• •Fortune 500 companies like banks have no document management systems and want technical people to map operations to GenAI
• •Working with fast-growing PE firms like BlueOwl to explore what can be mapped to GenAI
• •Law firms starting to take Harvey to clients and offer implementation services as a new revenue line - creating an ecosystem

Gabe explains the strategic decision not to build a law firm despite investor questions. They researched Atrium extensively and concluded you can't build two companies well simultaneously. The bigger opportunity is making every law firm AI-first rather than building one firm that gets conflicted out. Large M&As involve 100+ counsel firms globally - better to be the platform.

• •Talked to 30 Atrium employees - learned building a law firm and tech company simultaneously is too hard
• •Best outcome is helping every law firm become AI-first, not competing with customers
• •Large transactions (e.g., Microsoft-Activision) involve 100+ outside counsel firms across jurisdictions
• •Bigger opportunity: building the platform for professional services ($3-5T market) and client collaboration, not competing in legal services ($1T)

Gabe shares how Harvey started before GPT-4 launched, driven by conviction from seeing OpenAI's trajectory and Winston's deep legal industry knowledge. The key was finding the right initial form factor: upload a document and do something with it, plus accurate citations. This was immediately valuable for lawyers, unlike coding which took longer to find the right IDE integration.

• •Spent a decade trying to start an AI company before Harvey - was early, then recognized the right moment
• •Conviction came from watching GPT-1, 2, 3 progression and knowing OpenAI team members from Meta/DeepMind
• •Winston had unusual depth for first-year associate - understood firm strategy, was convincing partners to start new firm with him
• •Initial killer feature: document upload + accurate citations - immediately solved core legal tasks, unlike coding which needed more capability development

Harvey is hiring across engineering (especially frontend, AI, and FTEs) and just hired a New York site lead. Gabe's contrarian view: people underestimate continued model improvement and focus too much on individual copilots versus organizational productivity. The future is about how humans and AI work together at scale, similar to how SaaS enabled 10x larger organizations.

• •Hiring: strong engineers generally, frontend/product, AI researchers, and deployed engineers; scaling NY office
• •Prediction: people still don't appreciate how much better models will continue getting despite recent progress
• •Individual productivity (20% faster coding) doesn't equal 20% faster product development - need organizational infrastructure
• •Future is collaborative AI at scale: how do you organize thousands of humans and AIs effectively, not just replace humans with smart AI