Your Restaurant Can Have a Sommelier at Every Table
It is twenty past eight on a Friday. The room is full, two tables are waiting to order, and a guest at table six looks up from their phone and says the four words that cost you the most money all night: “what do you recommend?”
You know exactly what happens next. Your best server stops working the floor and starts working one table. Three minutes go by. The guest, not wanting to be difficult, orders the thing they already recognised, skips the wine, and asks for tap water. Nobody did anything wrong, and the check is fifteen euros lighter than it should have been.
That moment is the whole reason Restaurant Buddy exists. Not to make your menu prettier. To answer that question, at every table, in the two seconds before the guest gives up and orders the safest thing on the page.
The expert you cannot afford to put on the floor
Think about who you would hire if payroll were not a constraint. Not another runner. You would hire the person who can look at a table, hear “something savoury, no fish, and a dry white”, and come back with a plate and a glass that make sense together.
That person exists in maybe one restaurant in twenty, usually on the nights they happen to be working. Everywhere else, wine gets ordered by colour and price, because asking about a four-page list feels like sitting an exam in public.
Buddy is that expert, published behind the QR code already on your table. A guest taps two or three answers, or simply types what they feel like, and it comes back with one composed suggestion: the dishes, a glass to go with them, the total, and the reason each was chosen. It does that at every cover at once, in six languages, and it does not need a shift.
How the sommelier half actually works
Buddy reads your wine list once and learns it in your list’s own words. Not a generic wine ontology downloaded from somewhere — your list: the colour, the grape, and the character words your own descriptions use, whether that is tannic, structured, dry or savoury.
After that it can pair. A guest who has just been recommended the tagliata gets the Chianti next to it, at your price, with a line explaining why the tannin works against the fat. A guest who wants to stay lighter gets the Vermentino instead. And the most valuable thing it does is the thing a model told to always find an answer will never do:
“There is no fruity white on this list that stands up to the steak.”
An honest dead end is the most credible sentence in the whole feature. Guests forgive a recommendation they do not fancy. They do not forgive being sold a pairing that was obviously invented to fill a slot — and neither does your wine buyer.
Wine bars and enotecas can run the flow backwards, which is how those rooms actually work: start from the glass, then let Buddy suggest what to eat with it.
It cannot invent a dish. That is the design, not a promise
Here is the part worth understanding before you put anything AI-shaped in front of paying guests. Every dish Buddy names is checked against an item ID in your published menu before it reaches the screen, and the price the guest reads is taken from your menu rather than written by a model.
The validation is on IDs, not on names, and that detail is the difference between reliable and plausible. A model that returns a convincing dish name your kitchen has never cooked has no matching ID, so the answer is discarded instead of accepted because it read well.
Allergens go one step further, deliberately. If a guest writes that they are allergic to milk, Buddy does not quietly filter the menu. It offers the constraint back for the guest to confirm, and only that tap turns it into a filter — after which no dish carrying that allergen can be suggested to them again. A model reading an allergy out of a sentence is a reading, not a fact, and a filtered list looks authoritative whether or not it deserves to. Buddy stays a discovery aid and tells guests to confirm ingredients with your staff, because cross-contamination in a working kitchen is not something any menu database can describe.
Same for sold out. Hide the last two portions and they leave the recommendations immediately, cached answers included. Nobody gets sent to the pass for something that finished at seven.
You still decide what gets sold
This is not an assistant with its own taste. You mark the chef’s picks, the bestsellers, and the dishes 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 perfectly visible to anyone browsing the menu normally.
Between two plates a guest would enjoy equally, the one you want to move is the one that gets named. That decision is made once, in your settings, instead of depending on who is on shift. And the glass beside it is the line on the bill that nobody manages to upsell at half past eight.
Your printed menu becomes the AI menu in one upload
The honest reason most venues never went digital is not scepticism. It is the typing. Forty dishes, prices, sections, allergens: an afternoon of data entry nobody has.
So do not type it. Upload the carte you already print — the PDF from your designer, or photographs of the pages, up to 4 MB per import — and the dishes, descriptions, prices and courses come back as editable rows. The language is detected, a template that matches your kind of room is suggested, and nothing goes live until you have read it and fixed whatever came back wrong.
For context on how early this still is: in Italy, roughly 14.6% of restaurants and 9.3% of bars had a digital menu as of 2024. Nine venues in ten are still reprinting. Turning a PDF into data is now table stakes — several tools do it. What happens after the upload is where the difference sits: a PDF that has become a menu can be filtered, translated, repriced, hidden, and recommended from. A PDF that stays a PDF is a photograph of a menu that a guest has to pinch and zoom on hotel Wi-Fi. We wrote a longer comparison in QR menu vs paper menu.
What it costs, including the AI
Restaurant is 19 € a month and the AI is inside the price. We pay the model bill: 1,500 AI requests a month on Restaurant, 4,000 on Business. There is no provider account to open and no key to paste.
Only three things spend a request: a guest typing a sentence rather than tapping, the one-off reading of a new version of your menu, and a PDF import. Scanning the QR, browsing the carte and tapping the guided suggestions never reach a model at all — which is why a full dining room can cost nothing.
If the allowance runs out, service continues. Buddy keeps composing and pairing on its own engine and simply stops reading free text until the month turns; your guests see no error. If you need more than that every month, you can connect your own OpenAI, Gemini or Claude key and our cap stops applying, with your provider billing you directly.
Before your first real service
Three things, in this order, and they take about twenty minutes:
- Fill the tags. The import gets you most of the way, but allergens and categories are what recommendations are built on. A menu with half of them missing produces suggestions that look arbitrary.
- Mark the exclusions. Hide what is off and flag what is slow during a rush. This single setting decides whether your kitchen likes the feature.
- Test it as a guest. On a phone, in the room, at the light level the room actually has. Whether the flow is short enough is not a question you can answer from a laptop.
Then leave it alone and watch what gets ordered. The interesting number is not how many people opened the menu — it is how many of them stopped asking your servers what to have.
You can build the whole thing today: upload your PDF, check what the AI read, print one QR code for the tables. See how the AI menu works, or go straight to the plans.