Every AI idea meets the same objection: a vendor will ship this as a feature and there will be nothing left. It is a fair question. On the one idea-sourcing run we publish unedited, it is not the one doing the killing.
The run: 69 candidates in, 36 dropped, every drop carrying the checks it failed. The wrapper check was named on 6 of them. The check asking whether there was a documented problem behind the idea at all was named on 21.
That is 3.5 to one, and it is an argument about ORDER rather than about whether the wrapper question is soft. If you are holding an AI idea and worrying about a vendor roadmap, you are working on a question far down the list while the first one on it goes unexamined.
The stronger detail is in how those 6 drops were composed. Only 1 died of the wrapper question by itself. The other 5 each failed something else in the same pass.
What accompanied it on this run: hands-on support load, a missing documented problem, and a compliance barrier. Read that back and it stops sounding like an AI problem.
An idea that is a thin layer over a model is also, usually, an idea nobody documented a problem for, or one that needs a human in the loop for every customer, or one that cannot launch without a compliance review. The wrapper is the visible symptom because it is the part you can see from outside.
So the cheap checks go first, and here that is concrete. Is there a documented problem behind it — not a plausible one, one you can point at on a page somebody else wrote, with a link that opens. Can it run without you in the loop, because a model that needs supervision on every output is a support load wearing a product's clothes. Can it launch without a licence or an audit.
Each of those is settled by reading. The wrapper question, by contrast, is a forecast — and a forecast is the most expensive kind of question to answer first.
Where the AI question does bite: when the answer to what is left if they ship it is genuinely nothing. If the whole product is convenient access to a capability somebody else owns, the thing keeping you alive is their inattention, and pricing that as durable is how founders get surprised by a changelog. That is real. It is just not the common cause of death.
Limits: one run, one set of markets, one day. These are our checks, the ratings underneath are model output that samples fresh each run, and nothing here claims a rate.
The counted version, with the per-drop composition: https://whittleos.com/guides/ai-startup-ideas
Top comments (0)