We've watched fifty AI startups ship in the last two years. The winners share five habits.
1. They ship a v0.1 in two weeks. Not a landing page — a working thing. The feedback loop starts immediately.
2. They pick one narrow job and do it embarrassingly well. 'AI code review for Python Django repos' beats 'AI for software engineers' ten times out of ten.
3. They treat evaluation as the product. A 10-question eval suite, run on every model upgrade, tells them if they're actually getting better.
4. They design for the failure case. AI will be wrong 10-30% of the time. The UI shows confidence, offers correction, and never loses the user's work.
5. They price on value, not tokens. '$20 per merged PR' beats '$0.02 per 1K tokens'.
Bonus lesson: the demo army is a treadmill. Anyone can demo GPT-4-wrapper. Few can keep it reliable for 10,000 real users. Reliability is the moat.
If you're building in AI right now: narrow the job, measure ruthlessly, and ship weekly. That's the whole playbook — and it's the reason the winners keep winning.