AI in hotel restaurants and bars: menus, allergens and costing, with clear limits

AI in restaurants means using a general-purpose AI assistant to draft menu text, translate labels and check arithmetic on records the hotel already keeps. It is a drafting tool with a named human sign-off: descriptions get a chef's check, allergens come only from verified recipes, and costs are recomputed in the hotel's own sheet.

The season is close and the menu has to appear in four languages. Someone pastes the dish list into an AI chat, and the text comes back clean, fluent and confident, including an allergen note that no one in the kitchen ever wrote. That moment is where this article starts. After reading it, you will know which F&B tasks an assistant can draft, which documents decide the answer, and who signs each output before a guest or a ledger sees it.

What do hotel F&B teams already ask AI to do?#

Four jobs, mostly: write menu and buffet text, put that text into other languages, scale a recipe, and lay out a cost sheet. All four work on material the hotel already owns. Name each output plainly: a draft description, a translated label, a scaling formula, a cost table. None of them is a decision.

Draw the dividing line once and keep it. Text and arithmetic on the hotel's own records sit on one side. Facts about the world, such as what is in a sauce, what a supplier charged last week, or how many guests will arrive tomorrow, sit on the other. The line matters because an AI assistant writes fluent sentences whether or not they are true. Fluency is not evidence, and a confident paragraph is not a check.

A draft, in practice, is a document that no guest and no ledger sees until a named person has read it against the source. The source documents are the ones the kitchen already keeps: the standard recipe card, the purchase invoice, the supplier's product specification and the current menu. Give each output one owner, so that a wrong allergen line or a wrong plate cost has a person behind it rather than a team.

How should a hotel use AI for menu and buffet text in several languages?#

Give the assistant the dish's components and a fixed term list, ask for one language at a time, and let the chef approve the result against the plate. Feed it components, cooking method and what the guest actually receives, not just the dish name. A description written from a name alone is a guess dressed as a menu.

Keep a house glossary. It holds the dish names that stay in Turkish, the local product names, the unit words the hotel uses, and the spelling of the hotel's own outlets. Work one language per pass from the same source text, so that a correction in one language can be carried to the others instead of being rediscovered in each.

Ban promises the kitchen does not make. Words such as gluten-free, vegan or sugar-free belong in a description only when the recipe card says so, not when the translation sounds better. Fix a short buffet label format: dish name, the two or three main components, and the allergen line taken from the verified list, in the same order on every card.

The chef's check happens every time: does the text match the plate, the portion size and the garnish that actually leaves the pass. If you want a fuller treatment of when machine translation is safe for guest-facing text, see when AI translation is good enough for guest-facing text.

Why must AI never work out allergens from a dish name?#

Because allergen content is a property of the recipe and of the ingredients the kitchen actually bought, and a dish name carries neither. A name is a label, not a formulation. The same dish name covers different recipes, different suppliers and different substitutions in two hotels, and sometimes in the same hotel across two seasons.

Allergen facts come from two documents only: the standard recipe card, line by line, and the supplier's product specification or label for each bought-in item. Service-time substitutions and cross-contact on shared boards, fryers and utensils change the answer, and both are kitchen facts that no text model can see.

An assistant asked to list the allergens in a dish will produce a plausible list. Plausible is the failure mode here, not the safeguard, because the guest cannot tell a correct line from a fluent one. What the assistant may do is format, shorten or translate an allergen statement that a person has already verified against those documents. That saves typing without moving the responsibility.

The food safety adviser signs the statement, under the official rules the hotel has to follow. Confirm the scope, the labelling duties and your own case with that adviser rather than with a chat window. For the service routine around allergic guests, see a guest safety routine for food allergies in hotel restaurants and buffets.

Can AI scale a recipe and calculate portion cost?#

It can write the method and check the arithmetic, but the number that reaches the cost sheet is recomputed in the hotel's own sheet from its own purchase prices. Scaling is not multiplication. Trim loss, cooking loss, yield, batch size, pan and oven limits all bend the straight line, and the chef knows where.

Unit traps decide the result. Kilogram against liter, piece against gram, drained weight against net weight, and cost per usable gram rather than cost per bought gram. Ask the assistant for the formula, the conversion steps and a second opinion on a result that looks wrong. Do not ask it to be the place where the price lives.

Treat purchase prices and unpublished recipes as a confidentiality decision for the hotel, not a default. Decide what may leave the building before anything is pasted into a public tool. Build the sheet so each cell points at the current purchase price, which means the plate cost moves when the supplier invoice moves, without anyone retyping the recipe.

Rounding, currency and tax treatment get a human eye, and the cost controller owns the sheet that other people quote from. The underlying discipline is described in standard recipes and portion cost in hotel kitchens.

Where do AI suggestions help in purchasing and waste, and where do people decide?#

Suggestions help when they are questions asked of the hotel's own records; the order, the supplier and the write-off stay with a named person. Useful work includes summarizing last month's usage lines, grouping waste entries by reason, drafting a par-level worksheet, and turning a messy log into a readable table.

Not useful work: knowing tomorrow's covers. The covers figure comes from a person or from the reservation system, and the assistant reads it rather than producing it. Log the reason before asking for ideas. Overproduction, spoilage, trim and guest returns are different problems, and a reason code is what makes a summary worth reading.

A purchase order commits money and a supplier relationship, so the person who signs it owns the decision and the explanation that follows. Keep the loop small: one question, one answer, one person, one record, so that a suggestion can be traced back to the numbers it came from. Review the pattern monthly rather than daily. A single day's usage line is noise, and a month of them is a conversation with the chef.

The wider context for these numbers is food and beverage cost control in hotels, and the setup pattern, one task, defined data, a person who approves, is covered in setting up an AI assistant for one hotel task, with a person who approves.

What belongs on a one-page F&B rule card?#

Five rules that fit on one page and name a person for each output, so the kitchen does not have to remember a policy it never read. Pin it in the kitchen office, add a name to each line, and review it at the start of every season when the menu changes.

What to do next#

Write the rule card this week and start with the allergen line, because that is the one output where a fluent wrong answer reaches a guest directly. The rest of this article is one pattern repeated four times: the assistant drafts, the source document decides, and a named person signs, whether the output is a translated menu, a scaled recipe, a plate cost or a purchase suggestion.