Jake Heller is the co-founder & CEO of Casetext, the AI legal startup behind CoCounsel, which was acquired by Thomson Reuters for $650 million.
In his talk at AI Startup School on June 17th, 2025, he shared how his team did it—from picking the right idea to building AI products that actually work—and how founders can turn a cool demo into a reliable tool used by real customers.
Chapters:
00:00 — How We Built a $650M AI Company
01:00 — Picking the Right Idea in the AI Era
04:45 — Three Types of AI Startups: Assist, Replace, or Do the Unthinkable
09:25 — How to Build Reliable AI Products (Not Just Demos)
16:30 — The Importance of Evals and Testing
24:20 — Why Product Quality Beats Marketing and Hype
26:00 — How to Price and Sell AI Products
27:45 — Building Trust with Customers
29:30 — Product Isn’t Just Pixels, It’s Everything Around It
33:00 — What Founders Should Really Focus On
36:00 — Q&A: Picking Markets, Focus, and Defensibility
TL;DR 🚀
A successful AI startup founder shares how to pick impactful ideas, build reliable AI products with domain expertise and rigorous evaluation, and effectively market them by focusing on product quality and customer trust.
🚀 Key Points
Pick ideas by targeting tasks people currently pay others to do: assist, replace, or enable unthinkable work with AI 🤖
Build AI by deeply understanding professional workflows, breaking tasks into steps, and using precise prompts combined with traditional code ⚙
Focus heavily on rigorous evaluation with many tests and continuous prompt refinement to achieve high reliability ✅
Marketing success depends on an outstanding product that generates word-of-mouth and trust; price according to value, not just software seats 💡
Post-sale support and onboarding are crucial; invest in customer success and human interaction to ensure product adoption and retention 🤝
Summary generated by:
Cutsio Youtube Summarizer Chrome Extension
Amazing thanks! I have a product that has 100+ chained prompts + the code where needed, that I was starting to doubt, as it’s good but not great.
Now I can see the potential market it has, if it can do 90% of what a human would do compared to the 60% it currently does.
Thanks
So much gold. What a great time to be a builder/founder!
very good talk.. this one resonatted alot with me
I have learned nothing from just talking and I just wasted 40 minutes of my life I can never get back. Typical talk about how to build something: general ideas, sprinkled ambiguous technique, and presentations that don't add any value to the subject. You could have easily picked a concrete example, some draft idea or app and walk through the process from a to z.
if you meant Cursor ai when you said $20/mo then im here to let you know sir that for $20 you can simply hit the limit in 1 day.
What did that 14yr old build for Deloitte?
every single video i see on youtube has thousands of comments, and this gem has 18. and we except humanity to keep advancing fast.
It's not safe you cannot use that info online!!!
I like the approach to neural networks primarily from the standpoint of usefulness. Naturally, many people also need someone to talk to. But in medicine, gerontology, and other research sciences, neural networks can help analyze large volumes of data quickly. That’s the first practical aspect computers have long been developing. Mechanization and automation will live forever. I wouldn’t call neural networks “artificial intelligence,” even though that term is popular and partly true. I just prefer not to use it to keep the boundary clear. Still, those boundaries are becoming increasingly blurred. And yet, we are moving toward artificial intelligence and can already see a kind of map of our path – not a very clear one, but we know it exists.