This post summarizes the Y Combinator Lightcone episode Why Vertical LLM Agents Are The New $1 Billion SaaS Opportunities (October 2024). Hosts Gary, Jared, and Diana talk with Jake Heller, co-founder and CEO of Casetext, about building a vertical AI company. The episode description says Casetext sold to Thomson Reuters for $650 million in cash in 2023. Everything attributed to Heller below is from that episode.
Key Takeaways #
- YC’s pitch in the episode: vertical AI agents are key to the next generation of billion-dollar SaaS companies. That is the hosts’ view, not a measurement.
- Heller’s case is legal, where lawyers were skeptical until ChatGPT showed them what AI could do.
- His advice for an AI product: write tests first, solve the customer’s whole job, and aim for 100% accuracy.
The Rise of Vertical AI Agents #
Heller says he spent about a decade on Casetext, and that it started investing deeply in AI and natural language processing during that time. When GPT-4 became available to them, the company moved fast. He describes the moment he saw it as a turning point. Within 48 hours the whole company decided to work on a new product, Co-Counsel, built on GPT-4.
The Challenges of Law and AI #
Heller says lawyers bill by the hour and were wary of anything that might change their work. Then ChatGPT came out, and in his telling every lawyer saw that AI would change their work substantially.
Finding Product-Market Fit #
Heller says Casetext spent a couple of years trying to get lawyers to annotate case law, a user-generated-content approach inspired by sites like Stack Overflow and Wikipedia. It failed, because lawyers had no time to contribute. The company then pivoted toward natural language processing and AI.
Entering Deep Founder Mode #
Heller says that with a company of about 120 people, everyone switched to the new product, and he led by example by building the first version of Co-Counsel himself. He adds that a handful of other people helped, and that bringing in customers early helped convince the team.
The Importance of Prompt Engineering #
Heller describes writing the tests first and then writing English-language prompts until the AI passed them, for example getting the right answer 1,199 times out of 1,200. He compares this to test-driven development in software.
Going Beyond GPT Wrappers #
Some people say many companies are just building GPT wrappers. Heller’s answer is that a product that does the whole job of, say, a young associate adds many layers. In Casetext’s case that included proprietary data sets (the law itself and automatically added annotations), connections to law firms’ document management systems, and the quality of OCR.
Aiming for 100% Accuracy #
Heller says lawyers are conservative and a wrong answer costs a lot, so Casetext aimed for 100% accuracy. He says going from “almost works” to working every time is a very different task, and that test-driven prompting helped.
Thoughts on OpenAI’s Capabilities #
In the episode, Heller discusses OpenAI’s o1 model. He says it passes detailed-reasoning tests that earlier models could not, and that Casetext was investigating whether it can be prompted to think the way the best lawyers would.
Conclusion #
Jake Heller describes how Casetext rebuilt around GPT-4 and how that company ended up in a $650 million sale, per the episode description. The episode argues that vertical AI agents can become large SaaS companies. Whether it holds outside legal is not shown by one case.
