Jake Heller didn’t build Casetext to impress investors; he built it to do a lawyer’s job faster. That mindset — solving a real, paid-for problem — led to one of the biggest AI exits yet: a $650 million acquisition by Thomson Reuters. His Y Combinator talk lays out a blunt but elegant formula for building AI companies that endure.
1 Find the Job, Not the Novelty
The best AI ideas live where someone is already paying a human. Look for work that’s expensive, repetitive, or just miserable to do manually.
- Assist professionals with precision (CoCounsel did this for lawyers).
- Replace whole roles when the AI can outperform consistency and speed.
- Enable the impossible — things once unthinkable because of time or cost.
This reframes how we size markets. Instead of counting software seats, price against the wage bill of the humans you’re augmenting. The math explodes in your favour.
2 Reliability Beats Demos
Most AI demos dazzle, then disappoint. The difference between hype and product is evals — the obsessive testing loop that defines “what good looks like.” Heller’s team built evals for every micro-task, iterating until accuracy hit near-human reliability. That grind — not a funding round — made Casetext defensible.
3 Sell the Result, Not the Software
An AI startup wins when it sells outcomes.
- Price to value: charge for contracts reviewed, not licenses issued.
- Bridge the trust gap: let clients run your model side-by-side with their old method.
- Stay in the field: deployment teams that ensure usage turn pilots into recurring revenue.
When the product genuinely works, word-of-mouth becomes your marketing budget.
What It Means for Imbila Readers
For independent consultants and builders in Africa, the lesson is clarity: don’t chase “AI” as a label. Chase the job your client already values — the repetitive brief, the data slog, the analysis no one has time for. Then build reliability so strong they forget it’s AI at all.