Why Vertical LLM Agents Are The New $1 Billion SaaS Opportunities

Casetext spent years trying to improve legal research, first by asking lawyers to annotate case law. That approach failed because lawyers had little time to contribute, so the company shifted toward machine learning and tools such as case recommendations. Heller says those early products improved workflows incrementally, but did not create the urgency that came with large language models.

After getting early access to GPT-4, Heller and his co-founder decided within 48 hours to build CoCounsel, a legal assistant that could review documents and draft research memos. Heller built the first prototype himself before showing it to the company and customers. Their reactions helped persuade employees and board members to shift focus, even though Casetext already had $15 million to $20 million in annual recurring revenue.

For legal research, the desired result was a memo answering a lawyer’s question with citations to source material. Casetext worked backward from that result, breaking the work into steps such as generating search queries, reviewing cases, gathering insights, and drafting. The team wrote tests for each step, growing from dozens to hundreds and then thousands of examples.

Heller argues that a useful legal AI product includes much more than prompts: legal data, customer system connections, OCR, and handling unusual document layouts. Casetext used tests to find error patterns and refine instructions, aiming for answers lawyers could trust. He says a first bad experience can deter busy lawyers, so the initial use needs to work well.

Heller describes o1 catching subtle changes in a 40-page legal brief that earlier models missed, such as an altered word that reversed a quotation’s meaning. He says the team was exploring whether expert guidance about how lawyers think through a problem could improve results, but had no conclusive evidence yet. He sees room for AI products in fields where organizations spend millions on labor-intensive work.