Handle restaurant allergen questions without an AI guessing the answer

The menu is not the kitchen

A fictional guest messages a Bangkok restaurant: “Does the stir-fried noodle dish contain peanuts? If you leave off the topping, can I order it?” The public menu mentions noodles, vegetables and a nut garnish. AI recognises a familiar dish and drafts: “Yes, simply request no peanuts.” That answer assumes the garnish is the only relevant source and knows nothing about the ingredients or preparation actually used tonight.

The useful AI job is to preserve the question and send it to someone who can check the current recipe, purchased ingredients and preparation arrangements. Keep the reply within what that person verifies. This is a restaurant operations workflow, not medical advice or a claim that any particular meal is suitable for a particular guest.

Don't train staff to treat an answer from a model as a second opinion when the kitchen is uncertain. A more confident paraphrase cannot fill a gap in ingredient evidence. If the restaurant cannot establish the requested facts, communicate that limitation plainly and let the responsible staff member handle the conversation.

Ask about the actual dish, order and preparation

Capture the dish reference, variation, requested omission and the guest's exact concern. “No nuts” and “no peanut topping” are different requests; don't silently broaden or narrow them. If the guest names several ingredients, keep each one in the handoff. An image of a menu needs its date or version because recipes and supplier products can change.

The kitchen check should cover the current recipe and relevant purchased components, including sauces or garnishes. A recipe name isn't an ingredient statement. Nor does removing a visible component answer questions about preparation contact. Ask the responsible kitchen person what they can verify about the proposed preparation, rather than directing them to declare a dish safe.

For reference, the UK Food Standards Agency publishes allergen guidance and a matrix for food businesses, checked 20 October 2026. That is a useful example of structured ingredient communication, not a statement of Thai legal requirements. Have the restaurant's local food-safety lead define its own documented process and permitted customer wording.

A worked exchange with an honest stopping point

Initial acknowledgment: “I've passed your question about peanuts in this dish, including the topping and preparation, to the kitchen. Please wait for the team's answer before ordering it on that basis.” This acknowledges the specific concern without giving the answer prematurely. Don't promise a response in five minutes unless the service team can honour that timing.

Imagine the kitchen verifies that the current sauce includes an ingredient containing peanuts and that the usual topping can be omitted. The useful internal result is: topping removable; sauce still contains the ingredient; no approved alternative established. The customer draft should state that the dish still contains peanuts in the sauce even without the topping. It must not turn “topping removable” into “peanut-free.”

Now imagine the kitchen cannot identify the ingredients in a substitute sauce delivered that afternoon. The correct answer is that the team cannot confirm the ingredient information for this preparation. Escalate to the manager's documented process. Guessing from a brand's old online page, a standard recipe or another branch's answer would hide the actual uncertainty.

If a manager offers a different dish, run the same evidence check on that dish. “Alternative” doesn't mean “verified.” Keep the guest's original question attached so it survives the handoff between front of house and kitchen.

Copy a kitchen handoff that preserves the question

GUEST QUESTION, EXACT WORDING:
Order / table / conversation reference:
Dish and menu version:
Requested modification:
Named ingredient or allergen concern:
Language for reply:

KITCHEN CHECK
Current recipe checked by and at:
Relevant supplier ingredient information checked:
Substitutions or changed ingredients today:
Preparation information the kitchen can verify:
Unknowns that remain:
Approved factual answer:
Responsible staff member:

AI DRAFTING RULE
Use only the approved factual answer.
Do not infer ingredients from the dish name.
Do not replace uncertainty with reassurance.
Do not offer medical advice or diagnose the guest.
If the answer is incomplete, prepare an escalation message.
Keep the original concern visible through every handoff.

For Thai and English service, have a native reviewer approve the repeated question patterns and response templates. The wording needs to retain the difference between “contains,” “not listed in the checked recipe,” and “cannot confirm.” Don't use a general translation instruction to collapse those distinctions into a reassuring yes or no.

Design for the dinner rush

The failure point is often ownership. Front of house sends a message to a busy group, someone reacts with an emoji, and the enquiry looks handled. A useful handoff names a responsible staff role and keeps the question open until that person supplies an answer. A chat reaction is acknowledgment unless the restaurant has explicitly defined another meaning.

Put the ingredient information and substitution record where the kitchen can maintain them during service. A perfect database nobody updates after a supplier change is worse than a simple current sheet with an accountable owner. If maintaining the evidence costs more than the team can sustain, narrow the assistant's job to capturing and escalating questions. It can still reduce missing details and repeated explanations.

Keep customer information minimal. The kitchen needs the concern and order context, not a speculative profile about the guest's health. Restrict the conversation record to staff who need it for the service process. Your privacy owner should decide retention and access for the actual system; this article does not establish a legal retention period.

Test the answer the customer will receive

Give staff fictional examples with a removable garnish, a changed sauce, an unreadable supplier label and a dish that has different versions at different branches. Include a customer who corrects the ingredient they meant after the initial question. A good result preserves the correction and prevents the old draft from being sent.

Ask the kitchen lead to mark every unsupported assurance, even if the proposed meal happens to be acceptable in the invented scenario. The workflow must succeed because the evidence supports the answer, not because a guess got lucky. Review the actual customer wording, not only an internal classification labelled “escalated.”

Success here is a complete question reaching the right person and an accurate factual response reaching the guest. A chatbot that answers fewer allergen questions automatically can be more useful than one that sounds certain about every dish.

Sources and dates

Sources checked on 20 October 2026. Worked examples are teaching examples unless explicitly attributed. Provider documentation describes product behaviour; the operating recommendations are DLVX editorial judgement.

• UK Food Standards Agency allergen guidanceOfficial example of food-business allergen guidance and matrix; not Thai legal authority. Checked 2026-10-20.