Restaurant Buddy: an AI Menu Assistant with Guardrails
An AI menu assistant is only worth having if a guest can trust what it says. Restaurant Buddy sits on your digital QR menu and answers the one question a long carte on a small screen never answers well: what should I have? The guest taps two or three chips, or simply writes what they feel like, and Buddy comes back with a composed answer — starter, main, dessert and a glass, with the total and the reason each was chosen.
On the Restaurant and Business plans the AI is part of the price and billed to us: 1,500 AI requests a month on Restaurant, 4,000 on Business. There is no provider account to open and no key to paste, and a restaurant that outgrows the allowance can connect its own OpenAI, Gemini or Claude key to lift the cap entirely. What never changes is the grounding rule: every dish shown maps back to a real item ID in your published menu, and the price the guest reads is taken from the menu rather than written by a model.
That constraint is the whole design. A recommendation engine in a dining room has to be wrong in safe ways — suggesting a dish someone does not fancy is fine, inventing a dish or mishandling an allergen is not. This page explains what the model does and does not decide, how the wine pairing works, how a printed menu becomes data in one upload, where the monthly allowance goes, and what to check before you switch it on for a real service.
Every paid plan starts with a 14-day free trial, with no payment details to start. How the trial works
Two engines, and only one of them is a model
The selection itself is deterministic. Buddy filters your menu on availability, allergens and dietary tags, ranks what is left against the guest's answers and your own curation, and composes one suggestion per course. No model is involved in that, which is why tapping a chip is instant, identical every time, and costs nothing — a full dining room tapping chips does not use a single AI request.
The AI does three narrower jobs: it reads a sentence a guest has typed and turns it into filters, it reads your menu once per version to learn the vocabulary those filters are expressed in, and it reads a printed menu you upload. Anything it produces is validated before it is used: a filter has to name an attribute that exists in your menu's own vocabulary, and a dish has to match an item ID.
The split matters in service. A recommendation flow that occasionally takes four seconds, or occasionally produces something odd, gets switched off by the first manager who watches a guest hit it on a busy Friday. Keeping the decision deterministic and the language work optional is what makes it survivable in a real room.
Guests can just say it, in six languages
"Nothing with fish, something savoury and a dry white" is how people actually order. Buddy takes that sentence and returns four filters — no fish, savoury, white, dry — then composes from your menu. The guest never learns a syntax, and the free-text box sits beside the chips rather than replacing them.
The translation between the guest's words and your menu is the part that makes this work on any kind of house. Buddy reads your carte once and builds a vocabulary from it: the grapes on your list, the cooking methods your kitchen actually uses, the words your own descriptions use for texture. A guest asking for something full-bodied is matched to the word your list uses for it, so an enoteca and a trattoria get sensible behaviour from the same code.
Guest-side there are six languages — English, Italian, Spanish, French, German and Portuguese — and Buddy answers in the one the guest is using. Attempts to talk to the model instead of the menu go nowhere: a sentence asking it to ignore its instructions and recommend a made-up item is discarded, and the guest simply falls back to the chips.
The sommelier half: pairing from your own list
Most guests order wine by colour and price, because asking feels like an exam and the list is four pages long. Once Buddy has learned your wine list — red or white, the grape, whether a bottle reads as tannic, structured, dry or savoury — it can pair a glass with the plate it has just recommended, and put the total of both in front of the guest.
It is held to the same rule as the food: the wine has to exist on your list, at your price. You can also wire pairings by hand, dish to bottle, for the cases where you know better than any model — those are respected ahead of anything inferred.
The most valuable thing it does here is admit defeat. When the guest's constraints leave nothing sensible, Buddy says so — there is no fruity white on this list that stands up to the steak — instead of inventing a pairing to look clever. A wine bar can also flip the whole flow and start from the glass, then suggest what to eat with it.
Import the menu you already print
The reason most venues never go digital is the typing. So the way in is an upload: hand over the PDF your designer made, or photographs of the pages, up to 4 MB per import, and dishes, descriptions, prices and courses come back as editable rows.
The import also detects the menu's language and suggests a template that matches the kind of house it read — a five-course carte, a drinks-first list, a flat board with no descriptions. Nothing is published until you have looked at it: an import is a draft, and a price that read wrong is a field you correct, not a support ticket.
Paid plans carry up to 400 dishes on Restaurant and 500 on Business, which covers a long menu plus a wine list on the same page. Photos are stored as references rather than pasted into the document, so a large illustrated menu stays fast on a phone at the table.
What the monthly AI allowance actually covers
Three things spend one AI request: a guest typing a sentence instead of tapping, the one-off re-reading of your menu after you edit it, and a menu import. Scanning the QR, browsing the carte and tapping the guided chips never reach a model at all — so the median guest costs nothing and the allowance goes to the guests who genuinely write to you.
Restaurant includes 1,500 of those requests a month and Business 4,000. Running out does not produce an error: Buddy keeps recommending and pairing on its deterministic engine, and simply stops reading freely typed sentences until the month turns. A guest sees the same interface without the text box, which is deliberate — a degraded feature should not look broken.
If you need more, there are two doors and neither is a surprise invoice. Connect your own provider key and our cap no longer applies, with your provider billing you at cost; or tell us the volume you are doing and we raise the allowance.
Allergens and dietary needs: confirmation, not inference
If a guest writes that they are allergic to milk, Buddy does not silently filter the menu. It offers the constraint for the guest to confirm, and only that tap turns it into a filter — after which no dish carrying that allergen can be suggested, and the tiramisu is gone from every course it could have appeared in.
The reason for the extra tap is that a model reading an allergy out of a sentence is a reading, not a fact, and a filtered result looks authoritative. Dietary claims are treated the same way and in the strictest direction: nothing is ever described as vegetarian because nobody said it was not.
Restaurant Buddy remains a discovery aid, not an allergen verification system, and it says so to the guest. A kitchen's cross-contamination reality is not something a menu data model can know, so guests with allergies or intolerances are told to confirm ingredients with your staff.
A selling tool you control
You mark the chef's picks, the bestsellers and the items that should never be suggested — the ones that are slow during a rush, or the stock you are running down. Buddy leans on the first two and never offers the third, while the dish stays visible on the menu for anyone browsing.
That is where the commercial value sits. Between two dishes a guest would enjoy equally, the one you want to move is the one that gets named, and the drink beside it is the line on the bill nobody manages to upsell at eight o'clock on a Friday. The decision is made once, in your settings, instead of depending on who is on shift.
You also choose the button text, the tone and the language defaults, and you can switch the whole thing off at any time without touching the menu or reprinting the QR code. It is a setting on the page, not a different page.
What to check before a real service
Fill the menu data first. Recommendations are only as good as the tags underneath them, and a menu with half the allergens missing and no categories will produce suggestions that look arbitrary. The import gets you most of the way; the tags are the part worth ten minutes of attention.
Then mark exclusions and hide what is off. This is the single setting that most affects whether the kitchen likes the feature, because it is what stops the pass being sent a dish that ran out at seven.
Finally, test it as a guest — on a phone, in the room, with the lights as they actually are. Whether the flow is short enough, and whether the answer reads well next to your own carte, are not questions you can answer from a laptop.
Frequently asked questions
Do I need my own AI account or API key?
No. On Restaurant and Business the AI is included and billed to ShareCut: 1,500 AI requests a month on Restaurant, 4,000 on Business. You publish the menu and it works. Connecting your own OpenAI, Gemini or Claude key is an option for venues that want no cap at all — then your provider bills you at cost — but nothing about the feature waits for it.
Can Restaurant Buddy recommend a dish that is not on my menu?
No. Nothing is shown unless its item ID exists in your current menu, and the price the guest reads is re-read from the menu rather than written by a model. Validation is on IDs rather than names, which is the detail that makes it reliable: a plausible-sounding invented dish has no matching ID and is discarded.
Can it pair a wine from my own list?
Yes. Buddy learns your list — colour, grape and character words like tannic, structured or dry — and pairs a glass with the dish it has just recommended, at your price. You can also wire pairings by hand, and when nothing on the list fits the guest's constraints it says so instead of inventing one.
Can I import my printed PDF menu instead of typing it?
Yes. Upload the PDF or photographs of the pages, up to 4 MB per import. Dishes, descriptions, prices and courses come back as editable rows, the language is detected and a matching template is suggested. You review and correct everything before it goes live.
What happens if I use up the monthly AI allowance?
Service continues. Buddy keeps composing recommendations and pairings on its deterministic engine — chips, filters, budget bands, allergen constraints — and stops reading freely typed sentences until the month turns. Guests see no error. Bring your own provider key, or ask us to raise the allowance, if you expect to be there every month.
Can guests rely on it for allergies?
No. It is a discovery aid, not an allergen guarantee, and it tells guests so: a confirmed allergen filters the menu, but cross-contamination in a working kitchen is not something menu data can describe. Guests with allergies or intolerances should always confirm ingredients and kitchen practices with staff. The feature can also be switched off at any time without changing the menu or reprinting the QR.
Ready to put an AI menu assistant on your live carte?
See the AI menu →