AI Menu Import — Turn a Printed Menu Into a Digital One in 60 Seconds

Onboarding a restaurant to a digital ordering platform used to take a full day per venue — most of which was spent typing the menu into a form. With AI menu import you upload a photo or PDF of the existing menu, wait 60 seconds, and the platform has parsed every category, product, description, price, allergen, and modifier into the database, ready to be reviewed and published.

This post explains what the import actually does, where it gets confused, and how the review-and-cleanup step works in practice — because no AI import is 100% perfect, and the value is in how short the human cleanup is.

What gets extracted

• •Categories with their names ("Starters", "Mains", "Desserts", "Drinks")
• •Products with name, description, price, weight/portion where listed
• •Modifiers ("Add cheese +€2", "Choice of side: fries/salad/rice")
• •Variants ("Small/Medium/Large", "300g/500g")
• •Allergen tags when the menu uses standard EU allergen icons or text
• •Multi-language alignment — if the same menu has English and Bulgarian columns, both are captured and linked

The output is real database records — not a draft you have to retype. Each product has the same fields a manually-entered product would, including translation slots for languages the menu didn't cover (which you can later fill with the AutoTranslate button).

Where it gets confused

Three patterns trip the model up:

Decorative menus with photos of dishes overlaid on prices — the OCR sometimes attaches the wrong price to the wrong dish. Fix: re-photograph the menu with no overlap, or upload the kitchen's pricing spreadsheet alongside the menu image.

Handwritten daily specials boards — these read as approximately correct but with occasional misreads (a "9" that looks like "g"). The import flags low-confidence items for human review instead of silently saving them.

Complex modifier hierarchies on multi-course menus — "choose 2 of 5 sides, then choose dressing from 3" doesn't always nest correctly. These get imported as separate modifier groups that may need a quick consolidation step.

The review step

After import, the admin sees a preview screen with every extracted item, grouped by category, with a confidence indicator next to each field. High-confidence items (~85% of a typical menu) can be batch-approved. Low-confidence items (mostly prices and modifier counts) get a side-by-side view: the original menu image cropped to the row, alongside the extracted text. The operator confirms, edits, or rejects in one click each.

In practice, a 60-item menu takes about 8–12 minutes of human review after the 60-second import. Compare that to 90+ minutes of typing the same menu from scratch.

Multi-language menus get a bonus

Restaurants in Bulgaria and Greece often have printed menus already in two languages (local + English). The import detects this and links the two columns automatically, so the resulting digital menu ships bilingual on day one without any extra translation work. The third language (whatever the venue's enabled-languages list adds, e.g. Greek for a Bulgarian venue serving tourists) is left empty for the operator to fill via AutoTranslate — but the structural alignment is done.

Cost and limits

AI menu import is included in the platform onboarding for new venues at no extra charge. Existing venues can re-run the import to refresh their digital menu after a print-menu update — useful for seasonal menu changes where the kitchen has a new physical menu but nobody has the time to retype it into the admin.

See the AI Menu Import feature page for the technical detail and screenshot walkthrough, or the AI Menu Import architecture doc in the platform docs for how the vision model is constrained to prevent hallucinated prices.