The pattern is dull because it keeps happening. A company ships an AI feature. The demo is clean. Leadership is pleased. A quarter later the usage graph is a polite decline, and the write-up says the model hallucinates, or users do not trust it, or the change needs more enablement.
Usually the model was adequate. What failed was everything a person had to do with it.
the demo was a performance
Demos are run by people who know which question works, in a room where a wrong answer costs nothing. The user on a Thursday does not know the trick. They cannot see the difference between a fluent answer and a correct one. They are trying to finish a job. The gap between those two rooms is a design gap, not a model gap.
Three failures show up over and over. Every answer wears the same confidence, so one burn teaches people to trust none of it. The interface is a blank box, so people either ask nothing or ask the wrong thing. And when the system cannot help, it apologizes in a loop instead of handing the work to a person or a manual path.
People do not experience your model. They experience what the interface assumes about it.
design the failure first
Decide what the screen does when the answer is wrong, slow, or missing before you debate vendors. If those moments are honest — source, uncertainty, a way out — an imperfect model is usable. If they are not, a better model will still be abandoned.
Then put the help where the work already is. The features people keep do not announce themselves as AI. They make one painful step less painful. The model is easy to buy now. The experience is the part you still have to make.