Should an AI assistant answer allergen questions? No.
Allergen questions should be a hard stop for any restaurant assistant, transferred to a person without an attempted answer, because the cost of being wrong once is somebody in hospital.
Why should an AI assistant refuse allergen questions?
Because the downside is unbounded and the upside is small. Answering correctly saves a staff member thirty seconds; answering incorrectly can put somebody in hospital and end the business. No accuracy rate makes that trade sensible.
This holds even when the assistant has the ingredient list. Cross-contamination is a kitchen practice question, not a data question, and the answer can change with the shift, the supplier or the pan that was used.
What should the assistant do instead?
Stop immediately, say plainly that this needs a person, and transfer. It should not answer part of it, hedge, or offer general reassurance about ingredients before handing over.
Partial answers are the dangerous case. A caller who hears the risotto is probably fine and then gets transferred has already been given the answer, and that is what they will act on.
How do I know whether a vendor's assistant does this?
Ask them to show you the refusal in writing, and ask whether it is enforced in code or written as an instruction in a prompt. An instruction can be ignored by the model; a hard stop tested before deployment cannot.
It is a fair question to ask of any vendor, and the answer tells you most of what you need to know about how the rest of the system was built.
If this is your trade
Restaurants and hospitality →