AI in restaurants: what already works and what is still a demo
AI works in a restaurant wherever the data is already collected and normalised: reading supplier documents, checking purchase prices, recalculating dish cost, answering questions about your own figures. Where data sits in separate systems and was never reconciled, AI stays a demo and answers in generalities instead of talking about your restaurant.
Restaurants have been shown dozens of AI tools in the last two years: assistants, demand forecasts, automated ordering, guest chatbots. Some of them run in clients' kitchens every day. Some fall apart on the first question about an actual restaurant.
The difference is not the model. The models are much the same. The difference is whether there is data underneath that can be queried.
What already works
These tasks share one shape: a clear input - a document, a line, a figure - and an output you can check.
Reading supplier documents. PDF, photo, scan, e-invoice, PEPPOL. Lines are recognised, matched to the restaurant's catalogue and checked by arithmetic before posting. Detail on the invoice digitisation page.
Purchase price control. Every line is converted to a price per kilo, litre or piece and compared with your own history for that supplier.
Dish cost. A recipe card plus this week's purchase prices gives cost today, not cost as of the end of last month.
Answers from your own data. «Why did margin drop in August», «which dishes went past target» - provided the data has been brought into one model.
What is still a demo
Demand forecasting for a single venue. A restaurant is too small a sample: weather, a football match, roadworks and one corporate booking break any model.
Automated supplier ordering. Technically easy, but it needs trustworthy stock. While stock disagrees with the shelf, automation repeats the error faster than a person.
Dynamic menu pricing. Works in delivery; in the dining room it runs into printed menus and guest expectations.
An «AI manager» that decides everything. Decisions about suppliers, prices and people stay with a person - and should.
The test that separates them
Ask the tool about one specific number from your restaurant and ask it to show where the number came from. A working tool names the document, the line and the date. A demo answers with a paragraph of generalities.
The second test is what happens when data is missing. A working tool says «this item has no recipe card» and lists it separately. A demo substitutes an average and shows you a confident chart.
Where to start
Get documents into one place: the supplier mailbox, e-invoices, photographs of delivery notes.
Normalise the catalogue and units - without this every figure is arguable.
Turn on checks in the flow: price against history, units, duplicates.
Only then connect an assistant. By that point it has something to read.
The order is the reverse of the usual one: data first, AI second. What that looks like step by step is on the how it works page.
Invoice digitisationFood costWhere should a restaurant start with AI?
With data, not with a model. Get supplier documents into one place and normalise the catalogue. Until that is done, any assistant will answer in generalities.
Will AI replace the manager or the accountant?
No. AI takes the manual work - reading documents, reconciling, recalculating. Decisions about suppliers, prices and people stay with a person.
How do I check that a tool works rather than looks like it works?
Ask about one specific number and ask for the source. A working tool names the document, the line and the date.