A menu translation can sound polished while describing a different dish. “Creamy” may appear where the original says only “soup”. The useful task is a controlled language draft with every food-related statement tied to an approved record.
Start with five dishes and one target language. The workflow below is an editorial proposal for managing translation, not a validated food-safety system or a promise of fewer incidents. AI must never decide whether a dish contains an allergen, is free of it or is suitable for a particular guest.
Set the boundary before translating
Keep three separate fields: the dish’s display name, its approved description and its human-approved allergen information. A short menu description is not a complete ingredient record. Neither an appealing name nor a recipe establishes the absence of allergens or resolves cross-contact risks.
The Food Standards Agency’s best-practice guidance applies to England, Wales and Northern Ireland. It recommends written allergen information supported by a conversation for non-prepacked food, including restaurant meals. That recommendation is best practice, not a new requirement that every jurisdiction follows.
For restaurants elsewhere, check the applicable local requirements with the competent authority. This article does not interpret Ukrainian law or provide a complete compliance checklist. Ingredient verification, allergen management and guest conversations remain within the restaurant’s established procedures.
Assemble a record the reviewer can trust
Appoint a person who knows the kitchen’s ingredients, current products and allergen procedures. Give them authority to stop a release. A language reviewer should also be competent in both languages and food terminology; one person may perform both roles only if suitably qualified for both.
For each dish, prepare a source packet containing:
- A stable dish ID and the current approved recipe version
- The complete current ingredient record, including garnishes, dressings, cooking ingredients and bought-in components
- Current labels or supplier specifications for those components, with product identifiers and document dates or versions
- The approved source-language name and description
- Separately approved allergen wording and any relevant precautionary wording, with the responsible person’s approval
- A record of substitutions and the version currently used in service
GOV.UK’s business guidance on allergies recommends updating allergen information when recipes or ingredients change, informing staff and checking deliveries against orders. Our document IDs and approval fields are an editorial way to organise that checking.
If a component’s composition is unclear, obtain reliable information through the responsible person. An old supplier webpage, remembered recipe or previous menu is insufficient to approve the current record. Keep unresolved fields explicit; a blank must never become “none”.
Share only information approved for the AI service. Guest allergy reports and other personal information are unnecessary for a general menu translation. Follow your business’s data-sharing boundaries.
Separate language work from food decisions
Give AI the approved text and its source references. Ask it to preserve factual detail and list uncertainty. It may suggest readable wording; it must not invent preparation methods, origins, certifications, dietary labels or ingredient substitutions.
Allergen wording needs a separate review. Do not let AI create an allergen matrix from a recipe or “improve” an existing declaration. A translation of approved wording still requires the designated human to confirm its meaning and its match to the current dish.
Do not add “may contain” as a catch-all repair for missing data. The FSA’s technical guidance on precautionary allergen labelling connects such wording to assessed cross-contamination risk. A translation tool cannot perform that assessment or authorise a claim.
Example: one draft, two stops
Consider a document-handling example. It does not establish the composition or safety of any real meal.
Suppose dish S12 has an approved Ukrainian description: “Гарбузовий суп із підсмаженим гарбузовим насінням”. Its register points to recipe S12-v4, bought-in stock B17-v6 and topping record T03-v2. Assume the responsible person has verified those current ingredient records. The allergen-wording field is still awaiting their separate approval.
The requested English draft is “Pumpkin soup with toasted pumpkin seeds”, linked to description S12-v4. If AI returns “Creamy, gluten-free pumpkin soup”, reject the additions: neither claim appears in the approved wording. Do not ask the model to justify them from common recipes.
The acceptable language draft remains internal. Its status is “Awaiting human-approved allergen wording and final review”, so the menu entry is not released. Completing the ingredient packet does not authorise AI to fill the missing field.
Before release, the kitchen substitutes stock B18 for B17. The B18 specification is missing. Mark S12’s approval stale, record the replacement and request the current documentation. Even if the short description stays identical, the dish’s earlier checks no longer cover the current ingredients. Reopen the affected approvals before releasing the translation.
A similar product name or the same supplier does not clear the hold. The responsible person determines the next step after obtaining documentation.
A reusable source-locked prompt
Use this with approved records and adjust the language pair:
Prepare an English translation draft from the supplied Ukrainian source packet only. Work by dish ID. Translate the approved name and description without adding ingredients, preparation claims or dietary labels. Give the source ID and version for every field. Do not infer allergen presence, absence, cross-contact risk or suitability from names, recipes, ingredient lists or model knowledge. Translate allergen or precautionary wording only when a designated human has already approved that exact source field; keep it separate and mark it for human review. Do not generate or complete an allergen matrix. Report missing records, conflicting versions, unclear terms and substitutions. Use “HOLD: human clarification required” for affected fields. Missing information never means “none”. Return source text, translation draft, source reference and review question. Do not publish or alter menu systems.
The working-brief approach helps define the output. The prompt is a constraint, not proof that the model followed it.
Check the release, including every copy
- Compare every translated field directly with its approved source
- Resolve unclear culinary terms with the knowledgeable person and language reviewer
- Confirm current ingredient, supplier and substitution versions
- Have the responsible human approve allergen wording and relevant precautionary information
- Record reviewer, date, language and menu version
- Check printed menus, QR pages and delivery listings against that same release
- Remove superseded copies and reopen affected entries after changes
As a pilot measure, record review time and unsupported additions found. Do not report those counts as evidence of allergy prevention. The result is a traceable translation process with a clear human review boundary.
Questions owners often ask
Can AI identify allergens from the dish name?
Do not use it for that decision. Work from current verified records and the designated person’s approved information.
Is translating back into Ukrainian enough?
No. Back-translation can prompt questions, but it does not verify ingredients, current supplier products or the meaning of an approved declaration.
Can we publish the description while allergen checks are unresolved?
Keep the affected entry on hold in this workflow. Follow established procedures for current guest information; a draft or disclaimer does not close the gap.
