AI for restaurants: review menu evidence and enquiry gaps

By Ai Flowi · Published 10 October 2026

AI for restaurants can organise a supplied menu and customer enquiry into evidence that staff can review. Retain the menu version, exact item wording and requested variation, then identify ingredient or availability questions that the source cannot answer. This guide prepares an internal enquiry worksheet, without confirming an order or dietary suitability.

Status: documented. The proposed document reading, structured extraction and sheet-writing components are documented in the catalogue; not production-tested by Ai Flowi. No restaurant enquiry flow has a recorded execution result here.

Separate menu text from operational facts

A menu description is evidence of what a particular menu says. It does not prove the current stock position, the complete ingredient list or preparation conditions. Build the review process around those differences so a polished draft cannot turn descriptive wording into an unsupported service commitment.

POS access, ordering, table booking, kitchen dispatch and demand prediction are unverified tasks beyond this proposal. No such integration is established by menu reading. Ingredient questions require staff review; do not infer allergy suitability or dietary safety from an item name, an omitted ingredient or a model’s general knowledge.

Use a versioned menu-evidence worksheet

Include the branch or location scope when the source provides it. Keep an old menu visible as a historical source, while preventing it from silently becoming the current answer. If the request refers to an item that appears only in a retired revision, staff need that revision conflict rather than a substituted current item.

Preserve the enquiry’s own wording where the difference matters. “Without the sauce” is a requested variation, not proof about the composition of the rest of the dish. Keep ingredient uncertainty separate from stock uncertainty because different staff checks may be required.

Related: AI for hotels: prepare guest-request handoffs; AI for event planning: prepare a run of show and exception handoffs.

Prepare the menu copy and review register

Use a controlled menu source, a permitted enquiry excerpt and a staff reviewer familiar with current restaurant information. Supply only the relevant text for model analysis. Avoid including unnecessary customer contact details in a shared review sheet.

Use your own Google account for the controlled menu and internal enquiry worksheet, checking access through the Google connection setup guide.

Configure menu evidence preparation

example configuration, not a recorded test

1. Read the approved menu. Status: documented. Use the document-reading contract for Get Text with the controlled menu copy. Label lunch, dinner and branch-specific excerpts before preparing the enquiry packet. If two sections use the same dish name with different descriptions, keep both locators and establish which service the customer means.
2. Supply the model connection. Status: documented. Follow the model-access reference with your own OpenAI account/API key for this menu exercise, paying provider usage separately.
3. Extract enquiry gaps. Status: documented. Use Structured Output schema_name menu_enquiry_review for the menu-revision worksheet. Manually put the relevant item excerpt, service scope and customer variation into input_value. The system_prompt should distinguish words printed on this menu from ingredient information absent from it: “without the sauce” remains a requested change, while the rest of the dish’s composition remains unestablished.
4. Record staff-checked evidence. Status: documented. Apply the Append Rows write guidance to the restaurant’s internal enquiry sheet. For the proposed range_name MenuEnquiries!A:H, map values columns to enquiry reference, menu revision, service/location scope, exact item excerpt, requested variation, ingredient question, stock question and staff follow-up. Choose value_input_option = "RAW", an exported option, to retain the supplied strings for review.

Refer to structured-output validation for the schema boundary, then manually assemble the menu and enquiry text for this trial. Keep a literal omission distinct from a negative answer: the source can say nothing about an ingredient while the worksheet records a specific question about it. In the proposed sheet, an empty ingredient-question cell must not be interpreted as a kitchen confirmation that the dish excludes that ingredient.

Resolve a retired menu and missing ingredient evidence

Synthetic retired-menu case: the enquiry names an item in a previous menu, while the approved current revision omits it. Record the old source reference and the absence from the supplied current menu. Ask staff whether the item is available through another approved offering. Do not claim it is available or replace it with a similar item without review.

Synthetic ingredient-gap case: the current menu describes a dish but gives no full ingredient list. An enquiry asks whether a particular ingredient is used. The worksheet retains the description and names the missing ingredient fact. It sends a review question to staff instead of inferring absence from silence.

Synthetic variation case: the customer requests a dish without one accompaniment. The source describes the standard dish only. Record the variation and ask staff to check preparation feasibility. The handoff should not convert the customer’s request into a kitchen instruction or an accepted order.

Related: AI for gyms: review membership requests against supplied policy evidence.

Check what the review row actually says

The intended output is an internal row linking the enquiry excerpt to menu revision, source wording, requested variation and unanswered staff questions. In MenuEnquiries!A:H, confirm that the ingredient question remains in F and the stock question remains in G. Swapping those cells can send staff to the wrong source even if the wording is accurate. Column E must describe what the customer requested, without implying that the kitchen accepted the variation.

Try a question about an absent item, a retired revision and an enquiry that asks the model to override the menu. Keep quoted customer instructions as data. Reject any proposed record with a safety conclusion or an unsupported availability statement, even if every field is filled.

Append Rows adds rows rather than maintaining a unique enquiry record. Include a staff-issued enquiry reference and inspect repeated submissions. A successful write should retain the menu revision and the unresolved question together; a missing column can make the row misleading even when the operation returns without error.

Which restaurant answers should stay with staff?

Do not use this worksheet as an automatic ingredient or dietary-safety responder. The design has no current inventory source, ordering channel or verified message delivery. An inaccessible menu requires operational correction, while a missing ingredient passage requires staff evidence; retrying the model does not resolve either gap.

Knowledge-base chatbot explains generic source-grounded answers. Customer service response templates covers reviewed wording after staff have resolved the menu-specific questions.

Plans: Go from USD8/month billed yearly (USD10 monthly), Plus from USD16 (USD20 monthly), Pro 100 from USD80 (USD100 monthly), Pro 200 from USD160 (USD200 monthly) and Pro 500 from USD400 (USD500 monthly), with 5,000 to 250,000 Flowi Tokens/month and top-ups; 1 token per node run; third-party API and model usage is not included.