Winston Weinberg and Gabriel Pereyra approached generative AI from opposite sides of the same workflow. Weinberg understood the friction inside legal practice; Pereyra understood the capabilities and limitations of language models. Their early experiments with GPT-3 on legal questions produced answers that practicing lawyers judged surprisingly usable.
The pair sent a cold email to OpenAI and quickly secured backing from its startup fund. That origin became a template for one of the first successful vertical-AI companies: take a general model, surround it with domain data, workflow controls, security and evaluation, and sell the result into a profession where time is expensive.
Narrow focus created credibility
Harvey initially concentrated on large law firms rather than trying to serve every professional-services workflow. That decision forced the company to solve security, confidentiality and reliability problems early, because elite law firms have little tolerance for consumer-grade controls.
The result was distribution into institutions that are both conservative and highly valuable. Once Harvey proved useful inside firms, the same platform could expand into corporate legal departments, tax, due diligence and broader professional-services work.
AI agents changed the pitch
By March 2026 Harvey was talking less about a legal chatbot and more about agents that execute complete workflows. The company said customers were already running tens of thousands of custom agents and raised $200 million at an $11 billion valuation to expand that model.
That shift matters because agentic software competes for a different budget. A tool that drafts faster is productivity software. A system that runs portions of a legal process from start to finish begins to compete with labor, outsourcing and entire service lines.
The founder advantage is domain translation
Harvey's edge is not that its founders invented a frontier model. It is their ability to translate model capability into a regulated professional context where accuracy, provenance and trust matter more than novelty.
That makes the company representative of a broader American AI opportunity: large value pools may accrue not only to model labs, but to founders who understand one expensive industry deeply enough to rebuild its workflows around AI.