How Casetext built a $650M AI legal product over ten years

Executive overview

Legal research at 2 a.m. was as hard as ordering takeout was easy — that gap was the founding insight. Casetext spent a decade building toward that fix, surviving false peaks and slow growth before GPT-4 unlocked a product that compressed days of legal work into minutes.

Domain expertise plus the right model at the right moment is the repeatable formula for AI product-market fit.

The ten-year overnight success

• Founded 2013 as a crowdsourced case law library — Wikipedia meets Reddit for legal annotation
• Early AI tools used rudimentary NLP; useful, but not transformative
• First enterprise clients paid $50K–$150K each, creating false confidence in scale
• Exhausted early-adopter law firms; pivoted to smaller firms, saw thousands of sign-ups, then that channel also plateaued
• Maintained product velocity and customer feedback loops throughout every slow period
• Accessed GPT-4 roughly six months before public release — the inflection point

What product-market fit actually feels like

• Revenue added in millions per month, not thousands
• Enterprise clients who previously needed 9–18 months to decide signed in under a month
• Customers visibly lit up in ways no previous product had produced
• On track to triple annual revenue within a year of launching Co-Counsel

The killer demo pattern

• Co-Counsel uploads thousands or millions of documents and answers complex legal questions in minutes
• Enron demo: AI flagged emails with sarcasm ("Kenneth Lay, an honest man") as potential fraud evidence
• End of demo: 4–5 days of legal work done in 10–15 minutes
• This "golden demo" pattern — immediate visible value, immediate large-dollar commitment — is repeating across top LLM startups

Hard engineering problems between model and product

• Serving thousands of simultaneous users reviewing millions of pages requires significant infrastructure work
• Hallucination prevention: ensuring AI does not misstate document contents
• Regression testing: verifying that prompt or code changes don't introduce wrong outputs
• Most of this tooling had to be built internally; none of it came free with the model

The LLM opportunity stack

• Base layer: foundation models (GPT-4 and successors) — analogous to cloud computing
• Middleware layer: tools like LangChain that help developers use models effectively
• Application layer: domain-specific products built on top
• Each layer has genuine, large business opportunity
• Accurate, high-scale deployment is still the differentiating hard problem at the application layer

Real-world impact

• California Innocence Project faces a four-year backlog reading case files for wrongly imprisoned people
• LLM-based review could cut that backlog from four years to one month
• Legal work sits at the intersection of scale, complexity, and life-or-death stakes — exactly where AI leverage is highest