AI and the Future of Law: The 10 Year "Overnight" Success Story

After practicing at a large law firm and clerking for a federal judge, Heller saw how hard it was to find crucial information for legal cases. He contrasted hours of difficult research with how easily a phone could find nearby restaurants, even when the legal work could affect someone’s freedom or a business.

Casetext began in 2013 as a crowdsourced case law library, then developed tools to help lawyers find relevant material in documents. Early enterprise customers paid $50,000 to $150,000 each, but the company eventually reached limits with both large firms and smaller firms, where customer growth and marketing slowed.

Casetext had worked with language models for years and gained early access to GPT-4. That enabled CoCounsel, which Heller described as an AI assistant for complex legal tasks. The company began adding millions of dollars in revenue each month, and some long-time enterprise prospects made decisions in about a month instead of nine to 18 months.

Heller described demos using Enron emails to show how AI could flag possible evidence of fraud, including sarcastic messages and jokes about hiding assets. A demo that compresses four or five days of document review, research, or contract work into minutes helped lawyers see the practical value.

Turning a model into a legal product required more than access to GPT: Casetext had to handle large document collections, scale concurrent use, and reduce and test for inaccurate outputs. Heller pointed to the California Innocence Project’s four-year case-review backlog as an example of work that could be accelerated, and described opportunities across the model, supporting tools, and application layers.