The Lawyer Who Torched a $20 Million Company to Bet on a Chatbot — and Won $650 Million

The pitch lands like a dare. "What we're going to talk about today is how my company built an AI app that was so good we're able to bring it to an exit for $650 million — and how you can do that too." That is the opening salvo of Jake Heller, co-founder and chief executive of Casetext, standing in front of a room full of would-be founders at Y Combinator's Startup School. It is a startup-world flex, sure. But Heller spends the next half hour making the case that the number is not the point. The number is the exhaust from a process anyone in the room could repeat — and, he insists, will soon beat.

Watch on YouTube

Heller's biography is the tell. He grew up a coder — "building stuff since as long as I can remember," the same origin story as most of the audience. Then he took a detour that would define the rest of his career: he fell in love with law and policy, went to law school, did a clerkship, landed at a big firm, and lived the conventional legal life for a brief, disillusioning stretch. "I think like anybody who builds stuff and then goes to one of these old professions like law or accounting or finance," he says, "the first thing you find out is: I cannot believe that they were doing it this way." In 2013 he left to found Casetext — a company, he notes wryly, launched when much of his audience was "about turning eight." The subtext is a warning as much as a boast: startups are "one of the most amazing adventures of your life," but they take time.

For most of Casetext's existence, the conviction was simple and, for years, ahead of schedule: AI applied to law could make a huge difference. When they started, it wasn't even called AI. It was "natural language processing, maybe machine learning." One of the company's researchers, Heller recalls, saw the potential the moment the foundational transformer papers landed — "attention is all you need, etc., this like seven years ago." That deep focus on large language models bought Casetext something priceless: a front-row seat.

The Bet: Burn the Boats at $20 Million

Here is the moment the whole talk pivots on. In the summer of 2022, because they were so deep in the LLM weeds, Casetext got early access to GPT-4 — months before the public. And they were, by any normal measure, thriving. "We were like $20 million in revenue, we were doing great, I had like 100 people," Heller says. Then comes the sentence that makes the room go quiet: "We stopped everything that we were doing and said we're going to build something totally new based on this new technology."

We were like $20 million in revenue, we were doing great, I had like 100 people — and we stopped everything that we were doing.Jake Heller, on the GPT-4 pivot

That "something totally new" became CoCounsel — what Heller calls "the first ever, and I think still the best, AI assistant for lawyers." Roughly two years later, Casetext was acquired by Thomson Reuters for $650 million in cash. Heller is quick to deflate the figure even as he cites it. "That feels like a big number," he tells the students, "but I think for a lot of folks in this room, you're going to look back at this talk and be like, I can't believe that was a big number back then." His thesis for the entire session: the tools now in these founders' hands make far larger outcomes not just possible but likely.

Picking an Idea: Follow the Payroll

Y Combinator's most famous commandment is "make something people want." Heller's twist is that AI has quietly made that commandment easier to obey. In the old world of building software, he explains, knowing what people wanted required building, shipping, failing, and iterating in the dark. But there's now a shortcut hiding in plain sight: "What are people paying other people to do?"

Demand, in other words, is already proven wherever a salary is being paid. Customer support reps, insurance adjusters, paralegals, personal trainers, executive assistants — "that is what people want," Heller says, "because they're paying people to do it." For a large share of that work, LLMs can already do the job; for the physical world, robotics is coming. The problem of guessing what people want "just got a lot easier."

From there, he sorts every AI opportunity into three buckets. The first is assistance — helping a professional accomplish a task, which is exactly what CoCounsel did for lawyers drowning in documents, research, and contract markups. The second is replacement — skipping the professional entirely. "People currently hire lawyers. What if we just became a law firm powered by AI?" The third, and the one that clearly excites him most, is the previously unthinkable: work no human would ever be assigned because it was absurd at human scale. Law firms sitting on hundreds of millions of documents would never dream of having people read, categorize, summarize, and index every single one. "It just would be insane," he says. "But now that AI is here, you can have thousands of instances of Gemini 2.0 Flash, or whatever, read over every document. The previously unthinkable is now thinkable."

Heller anticipates the obvious objection — that mining salaries as your addressable market sounds dystopian — and rejects it head-on. "I think it's kind of the opposite. I think it's beautiful." He offers two reasons. The first is a Sam Altman analogy about lamplighters, the people who once roamed cities lighting and extinguishing gas lamps by hand. Automating roles doesn't erase human potential; it "unlocks a future that we can't even imagine today," the way we can't imagine wanting the lamplighter's job back.

The second reason is the one he clearly feels in his bones, drawn from his own field: access. "Over 85% of people who are low income don't get access to legal services," he says. Human lawyers are too slow and too expensive. But make lawyers 100x faster and 10x cheaper — or simply stand up a new AI-powered firm — and the calculus flips. "All of a sudden, saying where lawyers have to turn away clients because they did not have enough money, you can now say yes." Everyone, he argues, should get the world's best financial assistant, executive assistant, and — pointing to tools like Cursor and Windsurf — coding assistant. "You're going to do something really amazing for the vast majority of consumers and enterprises by unlocking a better future and by democratizing access to things that used to be only for the very wealthy."

Building It: Boring, Obvious, and Almost Nobody Does It

Then Heller says the most counterintuitive thing in the talk. Everything he's about to describe "may sound very simple and common-sensical and maybe even obvious, but the craziest thing is nobody's doing it." He underscores one word repeatedly — reliable — because reliability is the chasm between a demo that dazzles a VC and a product that survives contact with real customers.

Step one: figure out what a professional actually does, in granular, un-made-up detail. It helped enormously, Heller admits, that he and his co-founders were lawyers, and that 30 to 40 percent of the company — "even the coders" — were lawyers who had lived the work. If you don't have that, get it. "Just go be like an undercover agent somewhere," or find a co-founder with deep domain expertise. His refrain: "Don't fly blind."

Step two: ask how the best person in that field would do the task if they had unlimited time and a thousand AIs working in parallel — then work backward into the actual steps. He walks through Casetext's version of "deep research," built two and a half years before this talk, right after GPT-4 access. The best lawyer wouldn't just "research." They'd clarify the question, make a research plan, run dozens of searches returning hundreds of results, read each one, discard the irrelevant, take notes on why the rest matters, synthesize an essay, and then check that essay for accuracy against its sources. "These are the kinds of steps that a real professional might do. So write them down."

Step three: turn to code — and here Heller draws a sharp line. Most steps become prompts, because they require what used to demand human-level intelligence: "Read the legal opinion and decide on a scale of zero to seven how relevant it is." But, crucially, "if you can get away with it not being a prompt — if it's deterministic or a math calculation — that's better. Prompts are slow and expensive. Tokens are still expensive." Sometimes the answer is just good old software engineering, one function feeding the next.

The hard part frankly isn't building it. The hard part is getting it right.Jake Heller

Step four is the decision between a workflow and an agent. If the expert always follows the same five or six steps, make it a rigid workflow — no LangChain required, "just Python code," one function's output feeding the next. That, Heller confesses, describes much of what Casetext actually shipped. Only when the right approach genuinely depends on circumstances do you reach for something more agentic — which is harder to keep reliable, and should be a last resort, not a fashion statement.

The Real Moat: Evaluations

"The hard part frankly isn't building it," Heller says. "The hard part is getting it right." How do you know the research was done well? That the document was read correctly? That the insurance adjustment was right? This is where evaluations — evals — do the heavy lifting, and it's the single thing Heller sees almost everyone skip. Most teams ship demos that are "60 to 70% accurate," raise a round on the strength of a slick demo, sign a pilot or two, and then watch the whole edifice collapse when it doesn't work in practice. "LLMs, like people — you don't have your coffee that morning, you wake up on the wrong side of the bed — it might just output the wrong stuff."

Good evals start, again, from domain expertise: what does "good" look like for this exact task, and for each micro-task inside it? Heller's favorite trick is to design tasks with objectively gradable answers — true/false, or a number from zero to seven — so grading is trivial. Feed those into a framework (he name-checks the open-source promptfoo), start with a dozen realistic test cases, get to perfect, then 50, then 100, tweaking the prompt relentlessly. Keep a held-out set you don't peek at, so you're not overfitting to your own evals.

What emerges is oddly reassuring: AI fails predictably. Ambiguous instructions, a consistent lean in one direction — patterns you can prompt around with clearer instructions and examples. And then comes the line that functions as the talk's spine, the true qualification for success:

The biggest qualification for success here is whether you or whoever is working on the prompts is willing to spend two weeks sleeplessly working on a single prompt to try to pass these evals.Jake Heller

The grind is brutal and demoralizing by design. You start passing 60% of tests — most people quit here, declaring "AI just can't do this task." You spend a night prompting and reach 61% — the next wave quits. But push through two solid weeks of prompting and eval-adding, Heller promises, and you land around 97%, with the remaining 3% being genuine judgment calls a human might get wrong too. His pre-production benchmark: 100 tests per prompt, 99 out of 100 passing before beta. A thousand is ten times better, if you can manage it.

Do just those two things — study how professionals really work, then test every step and the whole workflow — and "you'll be like 90% of your way there to building a better AI app than most of the crap that's out there." The flashy Twitter demos that raise capital and become momentary heroes? Heller has a warning wrapped in a benediction: "Be careful who chooses your heroes. The real people are behind the scenes quietly building, quietly making their stuff better every single day."

Selling It: Product Beats the Pitch

The final act is the one Heller calls the hardest — the part he admits his multi-billion-dollar acquirer is "still trying to figure out." And he opens it by picking a fight with startup orthodoxy. Series A and B investors, he says, will tell you product barely matters if you're great at sales and marketing. He doesn't buy it. For ten years Casetext had an okay product and okay results, cycling through well-credentialed sales leaders. Then they built something genuinely great — and everything changed. "People were referring us by word of mouth. News was coming to us because we were doing something genuinely new and interesting." Word of mouth and press are free marketing. His existing salespeople "became like order-takers." His advice for the boardroom: "When you guys have those lame VCs on your board, you can think back to this talk and push back."

Still, selling matters, and Heller offers three hard-won lessons. First, you may not be selling traditional software anymore. The companies that excite him most are simply delivering the service — reviewing a contract that a law firm would charge $1,000 for, charging $500, maybe with a human in the loop. "$20 per month versus $500 per contract. We're talking about extreme step-ups in price. Price it according to the value you're selling." But, in the same breath: listen to how customers want to pay. When Casetext floated per-usage pricing, customers pushed back — they'd rather pay more for predictable, consistent billing. So Casetext charged $6,000 per seat. "Listen to your customers."

Second, build trust across the AI trust gap. Enterprises want to dip a toe in — every Fortune 500 CEO is being asked by their board "what are we doing about AI?" — but they're used to managing people they can coach and fire, not black-box software. The smartest companies run head-to-head comparisons: keep your law firm or your accountant, run the AI side by side, and measure speed, quality, and results. Pilots and studies, done thoughtfully, are how skeptics get comfortable.

There's going to be a mass extinction event as a lot of pilot revenue... is not going to convert into real money.Jake Heller

Third — and this is where Heller sounds a genuine alarm — the sale doesn't end when the check clears. As a post-exit angel investor, he keeps seeing startups touting "$10 million ARR" that, under scrutiny, is six-month pilots paying handsomely but not converting. He has a caustic name for it: not ARR but "PR — pilot recurring revenue," or worse, revenue that isn't recurring at all. "There's going to be a mass extinction event," he warns, as that pilot money fails to become real money — a danger even for startups reporting eye-popping numbers. The antidote is relentless onboarding, training, and rollout, industry by industry. He points to a "throwaway comment" from Satya Nadella earlier that day about forward-deployed engineers — "a really fancy term for just boots-on-the-ground people to sit next to your customer and make sure the product's actually working."

It all rolls up into a single conviction Heller repeated constantly at his own company: "Your product isn't just the pixels on the screen. It's the human interactions with your support, customer success, with the founder, it's training, it's everything around it." Get that wrong, and the best pixels in the world lose to a rival who invests more in their customers.

Q&A: Competitors, Focus, and the "GPT Wrapper" Fear

The audience questions drew out some of Heller's sharpest lines. On competitors: don't care about them at all. The markets are so vast — trillions of dollars in professional spend — that no single company wins them. And often "you're going to be dumbfounded about how bad they are." His real advice for choosing a market: look at what companies already outsource, even to other countries, because that signals work they don't consider part of their identity. Nobody outsources the storytelling at Pixar; plenty of firms outsource contract review. Find big markets with a shared pain point, throw a dart, "and I think you're going to probably hit a trillion-dollar market."

Asked how his focus changed across company stages, Heller offered a self-deprecating chorus: what he should have done at seed, Series A, B, and C was "focus on making a great product that gets product-market fit" — every single time. What he actually did was get pulled into HR, finance, and fundraising as ends in themselves rather than means to the product. "I fell into this trap. Big mistake." A company, he argues, is nothing but the service it delivers through its product; obsess over that, and hiring, culture, and marketing fall into place as necessities in service of it.

There was a charming exchange when a founder revealed they had sold a startup to Deloitte at age 14. "It's kind of amazing," Heller laughed. "I exit at 40 and you exit at 14." He confessed that focusing on legal was in some ways a mistake — the pre-LLM legal software market was small, "a very small amount" of the trillion dollars lawyers earn. His advice now: chase "the biggest problem you could possibly think of that is possibly solvable with the technology and skill set that you have." A robot that cleans your house for $1,000 a year, the way the 1950s dishwasher freed a generation's time — that scale of unlock is what he'd run at today.

On pricing services that no human could ever perform — reading hundreds of thousands of documents — Heller said to start from value: if it saves a business $100 million or replaces $5 million of work, take 10 or 20 percent and negotiate from there. Competition, he added almost cheerfully, will drive prices down over time until services cost "10 cents on the dollar" — "bad for your business, good for society."

And the fear haunting every founder in the room — that they're building nothing but a "GPT wrapper" on a model anyone can call? Heller's answer was the entire talk compressed to two words, and a challenge:

Just build it. And as soon as you build it, you'll see how hard it was... you're going to find that you built something that nobody else can build, because you spent like two years just doing nothing but that. So I'm not scared. Don't be scared.Jake Heller, closing the talk

The defensibility, in Heller's telling, isn't a secret model or a patent. It's the accumulated grind — the integrations, the checks, the finely tuned prompts, the model selection, the years of eval-chasing that competitors won't stomach. It's the same unglamorous work he spent the whole talk championing: study the job, break it into steps, test everything, and refuse to quit at 61%. The $650 million, he keeps insisting, was just the receipt.