Jake suggests targeting jobs people already pay others to do, such as customer support, insurance adjusting, or legal research. He groups AI opportunities into assisting professionals, replacing parts of their work, and doing tasks that were previously too costly to attempt. He argues that this can expand the market from software subscriptions to a share of the salaries or service fees companies already pay.
From Idea to $650M Exit: Lessons in Building AI Startups
Speaker
Jake, co-founder of Casetext, explains at Y Combinator how his team built the legal AI assistant CoCounsel and grew it into a $650 million acquisition by Thomson Reuters.
Jake recommends learning what professionals actually do, then mapping how the best person would complete a task with ample time and AI support. For legal research, that means clarifying the request, planning searches, reviewing results, taking notes, writing an answer, and checking citations. Each step can become code or a prompt; deterministic tasks can use simple workflows, while variable tasks may need an agent.
Jake says demos can work around 60 to 70 percent of the time yet fail with real customers. He recommends defining what good work looks like with domain experts, creating objective evaluations, and growing a test set from a dozen examples toward 100 or more. Teams should use customer failures as new tests and keep improving prompts and models; he says CoCounsel's team sometimes spent weeks refining a prompt.
Jake says a strong product can drive referrals and make sales easier, but founders still need to choose pricing and help customers adopt it. He advises pricing around the value delivered, asking customers how they prefer to pay, and using side-by-side comparisons or pilots to build trust. He warns that paid pilots may not convert to recurring revenue, so teams need to train customers and ensure the product is used successfully.
In the Q&A, Jake says he should have focused on product-market fit at every company stage, from seed through Series C. He describes recruiting, culture, sales, and marketing as work that should support building a product customers love. For choosing a market, he recommends looking for large, shared pain points, including work companies already outsource.
Takeaways
- Start with work customers already pay people to do.
- Learn the real workflow from practitioners before designing prompts.
- Test each step and the full task against examples of good work.
- Turn customer failures into evaluations and iterate continuously.
- Price for delivered value, then support adoption beyond the pilot.
Mentioned in this talk
Search inside this talk
What was claimed 7
- foundedJake Heller founded CaseText
“He's the co-founder of Casetext”
0:00inAI and the Future of Law: The 10 Year "Overnight" Success StoryYouTube Verified3 sources· Talk
- featureCoCounsel: AI assistant for lawyers
“the first ever—and I think still the best—AI assistant for lawyers”
YouTube Unverified1 source· Talk
- metricCaseText: $650 million in cash acquisition price
“We were acquired by Thomson Reuters about two years ago for $650 million in cash.”
YouTube Unverified1 source· Talk
- acquiredCaseText acquired
Thomson Reuters“We were acquired by Thomson Reuters about two years ago for $650 million in cash.”
YouTube Unverified1 source· Talk
- uses
“build something totally new based on this new technology”
YouTube Verified2 sources· Talk
- built
“That became a product called CoCounsel”
YouTube Verified3 sources· Talk
- feature
Promptfoo: Open-source evaluation framework that runs on the command line“It's open source and runs on the command line.”
YouTube Unverified1 source· Talk









