Most of what a restaurant phone agent does is low-stakes: hours, an order, a table. Allergy and dietary-safety calls are the exception. Get one wrong and the cost isn’t a lost order — it’s someone in an emergency room and a restaurant that’s liable. Any serious automation has to treat these calls as a category apart.
The one rule: don’t improvise
A language model is built to produce a fluent, confident answer. On an allergy question, fluent confidence is exactly the failure mode. “It should be fine” is the sentence that must never be generated, because a plausible guess about cross-contamination reads as reassurance and can be dangerous.
R.ai does not free-guess allergen and safety questions. It gives only what’s verified, and when a caller’s safety turns on the answer, it hands the call to a human — a manager or the kitchen — rather than reassuring anyone on its own. The design bias is caution: when unsure, escalate.
Verify, don’t reassure
The right behavior looks like this:
- If a fact is confirmed on the menu — “this dish contains peanuts” — the agent states it plainly.
- If safety depends on prep, cross-contact, or a substitution the agent can’t confirm, it stops and gets a person.
- It never manufactures reassurance to keep the conversation smooth. A pause for a human is the correct answer, not an awkward one.
Why this builds trust instead of spending it
Customers with real allergies are experts at spotting a place that’s hand-waving. An agent that says “that’s an allergy question, let me get someone from the kitchen” signals a restaurant that takes it seriously. That’s more reassuring than any confident bot could be — and it’s the honest answer besides.
It’s the same principle behind routing catering and delicate calls to a human: automate the routine perfectly, and escalate the calls where a wrong answer is expensive — or unsafe. Allergy calls are simply the sharpest example of the rule.