Building AI Unicorns: Lessons from Casetext's $650M Exit

“I cannot believe that they are doing it this way.” This sentiment, articulated by Jake Heller, co-founder and CEO of Casetext, encapsulates the entrepreneurial spark that ignited his $650 million AI legal startup, CoCounsel, recently acquired by Thomson Reuters. His candid talk at the AI Startup School on June 17th, 2025, offered a masterclass in navigating the nascent, yet rapidly expanding, AI landscape, providing invaluable insights for founders, investors, and industry leaders alike.

Heller, a self-described lifelong coder, embarked on a brief but conventional legal career before returning to his roots in technology. The inefficiency he witnessed in the legal profession was a glaring revelation. This realization, coupled with a deep-seated conviction that AI could revolutionize law, led him to found Casetext in 2013, long before "AI" became a mainstream buzzword. His team, focused on natural language processing and machine learning, gained early access to GPT-4 in the summer of 2022. This pivotal moment prompted a radical pivot, halting all existing projects to build something entirely new: CoCounsel, the first AI assistant for lawyers, which subsequently led to their substantial acquisition.

A core tenet of Heller’s philosophy for identifying viable AI startup ideas revolves around a simple yet profound observation: "What do people want? Well, people want, for example, things they're paying for right now." This insight transforms the daunting task of market discovery into a clear directive. Instead of guessing at future needs, founders should examine existing human-powered tasks that people or businesses are actively paying for. This approach, Heller argues, makes the "problem of choosing what people want just got a lot easier."

He categorizes these opportunities into three types: AI that assists professionals (like CoCounsel for lawyers), AI that replaces human tasks entirely (automating customer support or accounting), and AI that enables previously unthinkable capabilities (analyzing millions of documents instantly). The sheer scale of these existing markets, coupled with AI's ability to democratize access and drastically reduce costs, creates a Total Addressable Market (TAM) that is "a thousand-X bigger" than traditional software models. This isn't a dystopian vision of job displacement, Heller insists, but a "beautiful future" where essential services become accessible and affordable to a far broader population.

Building reliable AI products, not just impressive demos, is paramount. Heller stresses that many early AI startups falter because they focus on flashy prototypes rather than robust, accurate tools. The key, he explains, lies in deep domain expertise, truly understanding the specific tasks professionals perform. Founders should immerse themselves, becoming "undercover agents" to meticulously document the step-by-step processes of human experts. Most of these steps, he notes, translate directly into prompts for large language models or discrete pieces of traditional software engineering.

The real grind, however, begins with rigorous evaluation. Heller emphasizes the need to define "what does GREAT look like" for each micro-task and the overall workflow. This involves creating extensive test cases, starting with a dozen and iterating until the AI consistently achieves near-perfect accuracy (e.g., 99%). This process of continuous prompting, testing, and refining is arduous, with many founders giving up when initial accuracy hovers around 60-70%. Yet, it is this relentless iteration, often involving sleepless nights spent tweaking a single prompt, that separates reliable products from fleeting demos. He advises creating holdout sets of tests, ensuring that improvements aren't merely overfitting to known examples.

Related Reading

• From Discord's AI Growing Pains to Promptfoo's Red Teaming Triumph
• AI's Everyday Revolution: 24 New Ways Artificial Intelligence Reshapes Our World
• You will see a 30 to 50% correction in many AI-related names next year, says Dan Niles

Heller challenges the prevailing startup wisdom that marketing and sales are the most important drivers of success. "The most important thing you could do for marketing and sales is to build a fucking amazing product." He notes that Casetext’s earlier product, while "okay," required significant marketing effort. With CoCounsel, an "awesome product," news spread organically, and sales teams became "order-takers." This underscores that a truly exceptional product generates its own momentum, creating word-of-mouth and attracting media attention far more effectively than any paid campaign.

Furthermore, building trust is critical, especially in sensitive domains. Customers are accustomed to human imperfections but are wary of unpredictable AI. Companies must demonstrate reliability through head-to-head comparisons, pilot programs, and transparent studies. Beyond the pixels on the screen, the entire customer experience, support, training, thoughtful onboarding, and continuous engagement, is part of the "product." Founders must invest heavily in these aspects, ensuring users truly understand and adopt the technology. This commitment to customer success and adoption, rather than just initial sales, is what will differentiate winning AI companies.