When a storefront assistant states a specification that is not true, the shopper does not file a bug. They buy the product, receive something that does not match, and return it.
That reframes the accuracy conversation. The cost of a wrong answer is not a worse demo. It is the return shipping, the restocking, the support contact, and in regulated categories a compliance exposure.
Refusal as a feature
The instinct is to make the assistant answer everything. The better design makes refusal explicit and measurable: when the catalogue does not contain the attribute, say so and route to a human.
We track refusal rate as a first-class metric. A rising refusal rate is a signal that the catalogue data is incomplete (which is useful information) rather than a hidden failure being papered over with plausible text.
Grounding mechanics
Every claim an assistant makes should trace to a field in the merchant’s own data. Not “the model was given context and probably used it”. It is an actual pointer to the source field, retrievable after the fact.
That constraint is unglamorous and it is the difference between a feature merchants trust and one they quietly disable after the first bad week.