A founder posted something unusually honest this morning. He did not have a pain point of his own to build around, and he could not find one by looking, so he asked an AI to research the problems early stage startups run into and pick one for him. It suggested a webhook delivery tool. He knew the space was crowded. His next move was the healthy part: he asked a forum full of strangers to talk him out of it before writing any code.
Most commentary would mock the first half of that story. The second half deserves the attention. He did more diligence on a borrowed idea than many funded teams do on their own. But the sequence exposes something new about how companies get started now, and a gap that no amount of prompting closes.
Can AI find you a good startup idea?
It can find you a coherent one. Models are trained on every postmortem, every forum complaint, every landscape analysis ever written. Asking one for an underserved niche returns something that reads like the output of a good analyst, because it is a compression of thousands of them. The idea will have a market, named competitors and a plausible wedge. Coherence is not the bottleneck and never was.
The bottleneck is that a generated idea arrives with no owner. Nobody in the loop has the problem. The model read about it. The founder adopted it. The first person who will actually feel the pain is a hypothetical customer neither of them has met.
What does having the pain actually buy a founder?
A lie detector. When you have lived a problem, you can hear the difference between someone describing it and someone performing it. You know which workarounds people really build, what they already pay, which part of the week hurts. A founder interviewing customers about a problem he learned from a research summary has no baseline. Every polite answer sounds plausible, because plausible is all he has ever seen.
This is not an argument that only insiders can found companies. Outsiders have built great ones. But the outsiders who survived earned the baseline the slow way, by collecting so much contact with real users that they could eventually tell signal from courtesy. The pain can be borrowed. The calibration cannot.
Why is demand the one input you cannot generate?
Look at what is already automatable in that founder's stack. The market scan was generated. The competitive analysis was generated. The code will be substantially generated. Soon the outreach will be too. Every input to a new company can now be produced on demand except one: a real person who wants the thing enough to act.
A model can simulate that person. It can even simulate them convincingly, which is worse, because synthetic enthusiasm fails in the optimistic direction. Research on simulated respondents keeps finding the same pattern: they overstate willingness to pay, because they are built from text written by people being agreeable. The market is not agreeable. It ignores almost everything.
What should you do with a generated idea?
Treat it as a search result, not a decision. The founder with the webhook idea had the right instinct, get strangers to attack it, but the wrong jury. A forum will give you opinions, and opinions are leads at best. The test that matters is behavioural. Find people who ship webhooks to their customers today and watch what they do: have they built this in house twice, have they churned off an existing vendor, will they give you an hour of their calendar, would they commit anything at all. One person who altered their behaviour is worth a thread of takes.
And if nobody alters their behaviour, the idea was not yours anyway. Dropping it costs you a prompt.
Key takeaways
- AI generates coherent ideas, and coherence was never the bottleneck. Demand is.
- A generated idea arrives with no owner of the pain, which removes your ability to tell polite answers from real ones.
- Every startup input can now be produced on demand except a real person who wants the product.
- Treat generated ideas as search results. Validate them against behaviour from real people, not forum opinions or simulated users.
The cost of having ideas just went to zero. The value of knowing which one is wanted went the other way.