Knowledge
Why do AI projects stall after a successful prototype?
Because a prototype is allowed to assume the things production has to prove: that the data is clean, that someone owns the decision, that a wrong answer is harmless, and that no one’s job changes. Each assumption is reasonable in a demo and load-bearing in production, and they come due all at once.
The four assumptions that come due
That the data is clean
The prototype used an extract someone prepared. Production uses the live system, where the same field means two things depending on which integration wrote it.
That someone owns the decision
In the demo, the person watching approved everything. In production, it is unclear who is allowed to approve what, and the question has never been asked out loud.
That being wrong is harmless
A wrong answer in a demo is a laugh. A wrong answer in production changes a real order, and nothing was designed for that case.
That nobody’s job changes
The workflow assumed a person would keep doing the part the AI does not do — but nobody agreed to that, scheduled it, or measured it.
Why more model work does not fix it
The instinct after a stall is to improve the model, because that is the part the team knows how to improve. It rarely helps, because accuracy was not the binding constraint.
The binding constraint is usually that no one has stated what the system is allowed to do without asking, and until that is stated the work cannot go to production regardless of how good the answers are.
What actually unblocks it
Naming the decision rights, then building the controls that make them real: rules stated in advance for what AI-started work may do, a person approving where those rules require it before anything is applied, and a recorded run history rather than only a final state.
Questions buyers ask
- Should we have started smaller?
- Usually not smaller — narrower. A prototype that runs against the real system on one narrow workflow surfaces these problems while they are still cheap.
- Is this an argument against prototypes?
- No. It is an argument against treating a prototype as evidence that production is close. It is evidence about the model, not about the organization.