From Idea to $650M Exit: Lessons in Building AI Startups

Jake Heller explains how Casetext built CoCounsel and sold to Thomson Reuters for $650M by obsessing over evals, domain expertise, and product quality over marketing.

• Casetext got early GPT-4 access in summer 2022 at $20M ARR and 100 employees, then scrapped everything to build CoCounsel.
• TAM reframe: AI replaces salaries, not SaaS seats — a $20/month tool competes with $5–20K/month professionals, a 1000x expansion.
• Most AI demos run at 60–70% accuracy; Heller’s standard was 99/100 evals passing before beta, with daily prompt PRs after launch.
• Casetext charged $6,000/seat/year after asking customers — who preferred predictable budgets over cheaper per-usage pricing.
• Heller warns of a coming mass extinction of pilot revenue: companies reporting high ARR on pilots that are not converting to real contracts.
• Defensibility comes from accumulated complexity — data integrations, fine-tuned prompts, model selection — not from model exclusivity.
• Founders should focus exclusively on product-market fit at seed, Series A, B, and C; everything else (HR, marketing, culture) is a means to that end, not an end itself.
• For the previously-unthinkable category (e.g., reading millions of documents), price at 10–20% of customer value delivered, then expect competition to compress margins over time.

2025-10-28 · Watch on YouTube