Use the AI tool for layout, and nothing else: build a separate allergen sheet from supplier labels and recipe cards, add icons to the menu only from that sheet, and have the owner or head chef sign it off before anything is printed. AI should never decide whether a dish contains an allergen or is vegan.
Design tools and AI menu makers are good at columns, fonts, icons and spacing. They are not good at knowing what your suppliers put in the stock, the sauce or the bread. A tidy icon beside a dish looks like a statement of fact, and a guest with an allergy treats it as one. Menu wording is a separate job; for that, see writing menu descriptions with AI that still sound like you.
Why the layout tool is the wrong place to decide
An AI menu tool will often offer to "add dietary icons" for you. It does this by reading the dish name and description and guessing. "Margherita pizza" gets a vegetarian leaf; "roast vegetable curry" gets a vegan tick. The guess may be right most of the time, which is exactly what makes it dangerous: the wrong icon on one dish in forty is the one a guest relies on.
The guess fails because the facts are not in the description. The curry may be finished with a fish-based paste. The pizza base may use a dough improver containing egg. The "gluten-free" brownie may use a flour blend that is not certified, baked on the same trays as wheat. None of that appears in a menu line. It sits in labels, recipes and kitchen practice, so that is where the data must come from.
So the process has two separate tracks. Track one is data: a matrix of dishes against allergens and dietary labels, built from supplier labels and recipes, owned by a named person. Track two is layout: the AI tool arranges the menu and places icons exactly as the matrix says. The only link between them is the sign-off.
Build the allergen matrix first
Start with a spreadsheet: dishes down the side, allergens across the top, and a "source" column that says where each fact came from. Use the list of allergens that applies to your business. As an example, one widely used rule set lists nine major allergens (milk, eggs, fish, crustacean shellfish, tree nuts, peanuts, wheat, soybeans and sesame), while some other rule sets name fourteen and add items such as celery, mustard, lupin, molluscs and sulphites. Ask your local food-safety authority which list you must use, and keep it at the top of the sheet.
| Dish | Milk | Egg | Wheat or gluten | Fish | Nuts | Soy | Source checked |
|---|---|---|---|---|---|---|---|
| Margherita pizza | Yes | No | Yes | No | No | No | Flour and mozzarella labels, recipe card v3 |
| Caesar salad | Yes | Yes | Yes (croutons) | Yes (dressing) | No | No | Dressing label, crouton recipe |
| Vegetable curry | No | No | No | Check: paste label | No | Yes (paste) | Paste label read, delivery 12th |
| Pesto pasta | Yes | No | Yes | No | Yes (pine nuts) | No | Pesto label, pasta label |
| Chocolate brownie | Yes | Yes | Yes | No | Check: "may contain" | Yes (lecithin) | Chocolate and flour labels |
Three habits keep the matrix honest. Every cell is yes, no or "check"; never blank. Every row names the label or recipe version it came from and the date it was read. And the person who fills it in is the person who reads the labels, with the product in their hand. If a label says "may contain" a given allergen, record that as its own statement rather than quietly dropping it, because it affects what you can say about the dish.
Reading supplier labels: the places allergens hide
Most mistakes come from compound ingredients, items made from other items. When you read a label, read the whole ingredients list and any "contains" or "may contain" line, and note the date. These are the usual hiding places, described in general terms; always trust the label in front of you rather than this list.
- Stocks, bouillon and gravy bases: often contain celery, milk or wheat, and some contain fish.
- Soy sauce and many marinades: usually contain soy and wheat.
- Worcestershire-style sauces and some dressings: commonly contain fish (anchovy).
- Pesto, marzipan, nut oils and some pastes: carry tree nuts or peanuts.
- Breadcrumbs, batters and coatings: contain wheat, and sometimes milk or egg.
- Mayonnaise, aioli, meringue and some pasta: contain egg.
- Chocolate and baked goods: often carry milk, soy lecithin and "may contain nuts" lines.
- Wine, beer and vinegars used in sauces: can bring sulphites (on a 14-allergen list) or gluten.
Suppliers change recipes without notice. When a delivery arrives in a new pack, a new brand or a different size, read the label again before the product reaches a dish. Add a column "label last read" to the matrix and a standing rule: any substitution triggers a re-read and an update.
Use AI as a prompt for your checks, not as the checker
AI does help with one job: spotting ingredients you might forget to check. Give it a recipe's ingredient list and ask which items commonly contain allergens, so you know which labels to read. Then read them.
Here is the ingredient list for one dish in a restaurant kitchen.
For each ingredient, say whether it is a compound product (stock,
sauce, paste, coating) that commonly contains any of these
allergens: milk, egg, wheat/gluten, fish, shellfish, tree nuts,
peanuts, soy, sesame, celery, mustard, sulphites.
Do NOT say whether the dish is free of any allergen. Only list
which ingredients I must check on the supplier's label.
Ingredients: [list]A typical response (illustrative) for a curry lists "curry paste: may contain fish, soy, wheat; check label", "stock cube: may contain celery, milk; check label" and "naan: contains wheat and milk; check recipe". Notice what it does not do: it never declares the dish safe. If a response ever says "this dish is vegan" or "contains no nuts", treat that as a failure of the prompt, delete it, and rerun.
Dietary labels need your definition
"Vegetarian", "vegan", "gluten-free" and "dairy-free" are not the same as the absence of one allergen. Decide what each label means in your kitchen and write it on the matrix sheet. Vegetarian usually means no meat, fish or fish products, which also rules out some stocks, gelatine and certain cheeses made with animal rennet. Vegan adds no milk, egg, honey or other animal-derived ingredients. "Gluten-free" often has a legal threshold in some places, and cross-contamination from shared fryers, boards and trays counts; check the rule that applies to you before using the phrase.
If you cannot meet the standard, change the label, not the standard. "Made without gluten-containing ingredients, prepared in a kitchen that handles wheat" is a different statement from "gluten-free", and a guest with coeliac disease will read the difference. Where in doubt, use the wording that promises less and say how to ask.
Main example: a 28-dish menu, checked before print
An illustration, with invented figures. A 40-seat neighbourhood restaurant with 28 dishes uses an AI menu tool to lay out a new autumn menu. The owner builds the matrix first: 28 rows, 9 allergen columns plus "vegetarian" and "vegan", and a source column. Reading labels takes about three and a half hours the first time, because each of the roughly 60 purchased ingredients has to be picked up and read. That is the cost of doing it properly once; later menus reuse most of the rows.
The check finds three discrepancies. First, the vegetable curry was drafted as vegan, but the paste lists fish; it becomes "vegetarian" only if the paste is changed, and the owner decides to swap to a plant-based paste. Second, the chocolate brownie had been marked "gluten-free" on the old menu. The kitchen bakes it on the same tray as a wheat tart, so the label is changed to "made without wheat ingredients; baked in a kitchen that handles wheat". Third, the Caesar dressing contains fish; the old menu had no fish icon on that dish. The new one does. None of these appeared in the dish descriptions, and an AI icon guess would have missed all three.
Then the layout. The owner pastes the matrix into the design tool's chat and uses the prompt below. The tool returns a two-column layout with icons.
Lay out this menu in two columns for A4 or letter-size print.
Use the icon set: M (milk), E (egg), G (gluten), F (fish),
N (nuts), S (soy), V (vegetarian), VG (vegan).
Place icons ONLY where the table below shows "Yes" for that dish.
Do not add, remove or guess any icon. Do not change dish names.
If a row says "check", leave a visible [CHECK] marker.
[paste matrix]The first output (illustrative) is nearly right but moves the fish icon from the Caesar salad to the line below it, because the dish name wrapped onto two lines. That is the typical layout fault: an icon floats to the wrong row. It is why the proof read in the next section is not optional.
The proof read against the matrix
When the layout is done, check the page against the matrix by a person who did not build either. Go line by line: dish, then each icon, then the dietary label. A second pair of eyes catches the wrapped-line fault and any icon the tool "helpfully" added. Tick each dish on a printed copy. If a single line is wrong, correct it and proof the whole page again; fixes can move other text.
A useful tactic is a reverse check: for each icon, read the matrix row and ask "is this the right dish?" rather than looking at the menu and asking "does this look right?". The first method finds misplaced icons; the second often does not.
The sign-off sheet
One page, kept with the printed proof. It records who checked what and when, so that the menu has a history. A filled-in example (illustrative) is below.
| Field | Entry |
|---|---|
| Menu version and print date | Autumn menu v4, 14 October |
| Matrix version used | Matrix v4 (28 dishes) |
| Labels re-read since last print | Curry paste, dressing, brownie chocolate, bread flour; dates in matrix |
| Changes from last version | Curry now vegetarian (paste changed); brownie wording changed; Caesar fish icon added |
| Matrix checked by | Head chef |
| Layout proofed against matrix by | Front-of-house manager |
| Open [CHECK] items | None |
| Signed (owner or head chef) and date | Signature and date |
Nothing goes to the printer without a signature on the sheet and no open "check" items. If a supplier substitutes something the day before the print, the sign-off waits.
Icons, keys and the line that invites a question
Design choices decide whether a guest can use the labels. Put a key at the foot of every page, not only the first. Use letters or short words beside the icon, since colour alone fails for guests with colour-blindness and disappears on a black-and-white photocopy. Keep icons the same size and in the same place on every row, ideally right-aligned, so the eye can scan down the column. Ask the AI tool for exactly that layout in the prompt, and reject any version that puts icons inline with the dish name, where wrapped lines scatter them.
Add one printed line that invites the conversation: "Please tell us about any allergy or dietary need when you order; our team can check each dish with you." The menu cannot promise a kitchen free of cross-contact, and the line makes clear that a person is the final check. Write the exact wording once, have it approved, and keep it on every page and on the QR version.
What staff say when a guest asks
A matrix on a shelf helps nobody at table six. Print a one-page version for the pass and brief the team for ten minutes on how to use it. The rule is simple and has no exceptions: staff answer from the sheet, never from memory, and if the sheet says "check", they ask the chef.
Three short scripts (illustrative) cover most cases. A guest asks "Is the curry vegan?": the server reads the row, says "Our curry is vegetarian, made with a plant-based paste, and cooked in a pan used for other dishes, so I will ask the chef to confirm it is suitable for you" and goes to the chef. A guest says "I'm allergic to nuts": the server notes it on the order, tells the kitchen, and does not recommend dishes from memory. A guest asks whether the brownie is gluten-free: "It is made without wheat ingredients but baked in a kitchen that handles wheat, so we cannot promise it is free of traces." Each answer states what you know and what you do not.
A bakery counter: the same method on a smaller scale
A different case to show the method scaling down. A small bakery sells 18 items on a printed board and a one-page takeaway menu. The owner bakes from a handful of mixes and flours, so the matrix has 18 rows and a short purchased-ingredients list. Reading labels takes about 90 minutes. Because the bakery makes everything in one room, nearly every item carries a "may contain" statement for wheat, milk, egg and nuts, and the honest board says so in a single line rather than pretending any item is free of them.
The AI layout prompt is the same, with one extra instruction: "put the line 'All items are made in a kitchen that handles wheat, milk, egg, nuts and sesame' under the heading, and do not use the word free anywhere". The result is less decorative than a restaurant menu, and it is true. For packaged items sold to take away, where allergens must usually be on the pack itself, see meal prep software for weekly menus, orders and AI labels.
Time and cost of doing this properly
For the 28-dish restaurant, the first run is about 3.5 hours of label reading, 1 hour to enter the matrix, 1 hour for layout with the AI tool, 45 minutes for the proof read and 15 minutes for sign-off: about six and a half hours, once. Each later menu change costs about 45 minutes, because most rows stay the same. Design tools have free tiers, and AI features in some (Canva, for example) draw on a monthly allowance, and Canva's pricing page lists up to 20 Standard or Premium AI uses for the free plan as of October 2026; a plain template needs none of that. The real cost is the owner's reading time, and it is the cost that cannot be skipped.
Weigh that against what an error costs: a guest taken ill, a closed kitchen, a reprint of the whole run. Six hours on a first pass is a small price for a menu you can stand behind. If you decide to run the matrix inside a larger planning tool later, building a seasonal menu from your supplier list shows how supplier data can feed menu planning.
After print: changes, specials and QR menus
A printed menu freezes the data; the kitchen does not. Write a rule for changes: any recipe or supplier change updates the matrix first, and if a dish's allergens or dietary label change, the menu is corrected before the dish is served as it is described. For a day's change, a clear written note at the pass and on the table can bridge the gap, but only if staff have the matrix and know to point to it.
Specials are the usual gap. A special added at 4 pm has no icon on the printed menu, so train staff to answer from a one-line matrix row written at the same time. If you also run a QR menu, the same data should feed both, so you update once. Building printable and QR menus in Canva covers the layout side, and whether an AI chatbot should answer allergen questions explains why a chatbot is the last place to answer them without the matrix behind it.
Translated menus add a further risk: a translation tool can render an allergen word wrongly. Have a fluent person check the allergen terms specifically, and read whether AI can handle restaurant menu translation without mistakes before relying on one.
Where this leaves the AI tool
The AI tool saves time on layout: a menu that took an afternoon of fiddling takes an hour. The saving stops where responsibility starts. Labels are read by a person, the matrix is owned by a named person, and the print is signed by the owner or head chef. That is also the answer if a guest or an inspector asks how your menu was checked. Rules vary by where you trade and I am not a food-safety adviser, so confirm the allergen list and wording rules with your local authority, and keep a dated copy of each signed sheet. If you would like to set up the matrix and sign-off with someone, that is what my AI implementation consultation covers; get in touch if that suits.
Questions owners ask about AI menus and allergens
Can an AI tool work out which allergens are in a dish?
It can suggest ingredients that often contain allergens, which is a useful prompt for your checks. It cannot know what is in your suppliers' products or your recipe on the day. The allergen list must come from labels and recipes you have read, not from a guess.
How many allergens must I label?
It depends on where you trade. Some rules name 9 major allergens, others name 14, and the lists differ in detail. Check the list that applies to your business with your local food-safety authority, and build your matrix on that list.
Do I need to reprint the menu when a supplier changes a product?
Possibly. A new brand of stock, sauce or bread can add or remove an allergen. Re-read the label on delivery, update the matrix, and reprint or add a correction if any dish changes. A short rule helps: no change goes live until the matrix is updated and signed.
Is a QR menu safer than a printed one?
A QR menu is easier to correct, because one edit updates every table. It is not safer unless the data behind it is right. The same matrix and sign-off apply, and you should keep a way for guests to ask a person.
Further reads
- AI in Food Service: A Four-Week Menu Cycle With Costs and Allergens — Allergen tracking across a four-week menu cycle.
- AI for Delis: Daily Specials, Platter Orders and Allergen Labels — Allergen labels for deli specials and platter orders.
- AI Menu Engineering: Which Dishes to Promote, Reprice or Drop — Decide which dishes earn their place on the page.
- AI Marketing for Restaurants: A Four-Week Plan Owners Can Follow — Promote the menu without posting unchecked dietary claims.
- AI Chatbot Online for Hospitality: Questions a Café Can Safely Automate — A checklist that sorts café guest questions into chatbot-safe answers, answers that need a check, and hand-offs to a person.
- AI for Cafés: Plan a Seasonal Drinks Menu From Your Sales Data — Export drink sales, add costs, and have AI rank profit per cup, propose specials from stock, set prices and score a four-week test with keep-or-drop rules.
- AI for a Cloud Kitchen: Keep Every Brand's Menu Consistent on Apps — A brand-by-app checklist for cloud kitchens: one master menu sheet, AI rewrites that keep the facts, sheet checks for prices and allergens, and a change log.
- Bubble Tea Staff Training: Drink Recipe Cards and Quizzes With AI — Record each drink once, let AI format it into a recipe card with sweetness and ice levels, then check it and build a ten-question staff quiz from the cards.
- Catering Event Checklist: Build the Run Sheet From One Brief With AI — A brief template, a prompt and a manager's checklist for building a catering run sheet, equipment list, staffing plan and load-in plan from one client brief.
- ChatGPT Prompts for Restaurants: 10 for Menus, Rotas and Replies — Ten prompts for menus, rotas and customer replies, with inputs, worked outputs and the checks that catch a wrong total, label or promise.
- Food Truck Vendor Applications: Draft Festival Forms With AI — A drafting method for festival and market forms: facts sheet, requirements table, answers from your facts only, and a gap check on insurance and licences.
- Free AI Restaurant Menu Generator: What the Free Tiers Allow — Canva Free, ChatGPT and Gemini compared for menu making: AI allowances, watermarks, print resolution and the free route that works.
- Takeaway Delivery App Menus: Rewrite Descriptions and Photos With AI — A description formula, paste-in prompts, per-app photo checks, a modifier-group clean-up and a two-week test for takeaway delivery menus.
- Can ChatGPT Write a HACCP Plan for a Restaurant? What to Verify — What ChatGPT drafts well in a HACCP plan, what it guesses, and the checks your food safety authority still makes: limits, logs, thermometers and training.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: FDA food allergen information page (nine major allergens, including sesame); Canva pricing page; ChatGPT, Claude and Gemini plan pages.