AI for restaurants: what can a small café actually start with?

A small café doesn't need a robot or a big project to start with AI. It needs a few jobs where a bad draft hurts nobody, and clear limits where a mistake could hurt a guest. This guide separates the two. It's written for cafés and restaurants in Germany and the EU, so the legal notes apply there.

Are small businesses using AI yet?

Yes, but still a minority. Germany's Federal Statistical Office says 26 percent of enterprises used AI in 2025, and 23 percent of those with 10 to 49 employees. That table has no separate figure for hospitality, and many small cafés are below that size. So you're not late.

What can a small café actually start with?

Start with texts and plans that you read yourself before using them. Germany's hospitality association advisers, DEHOGA Beratung, give similar examples according to Tageskarte: a restaurant creating weekly cleaning plans, a café planning Instagram posts. The table runs from least to most risky.

One rule covers all of it: what AI writes is a suggestion. DEHOGA Beratung says the same, which is why every row has a check. We're not quoting hours saved, because we found no solid figure from small businesses.

What should AI never decide alone in a restaurant?

AI shouldn't decide anything alone that could harm a guest or commit you to something. That mainly means these six areas.

Why shouldn't AI answer allergen questions alone?

Allergen information has to be right, and AI can't see your kitchen. It doesn't know about the supplier's new product or the shared fryer. For food sold loose in Germany, section 4 of the LMIDV regulation applies: the information must be clearly visible and readable before purchase, on the menu, on a sign, as a notice or electronically. Giving it verbally only works with a notice and a written or electronic record on site.

How language models handle this has mostly been studied outside restaurants. A 2023 study in the journal Nutrition had ChatGPT create 56 diets for hypothetical people with food allergies. The verdict: generally accurate, but with the potential to produce harmful diets. It tested an older model and meal plans, not your menu. For a kitchen it's warning enough: an answer that's usually right isn't good enough for allergies.

There's more in the guide to allergens on the menu.

Who is liable if a chatbot says something wrong?

Don't assume it's the chatbot's maker. The best-known case comes from Canada and not from hospitality: in 2024 a tribunal in British Columbia sided with a customer after an airline's chatbot explained a fare wrongly. The airline argued the bot was responsible for itself. The tribunal treated it as part of the airline's website. That isn't German law, but it shows the direction. DEHOGA Beratung also advises settling beforehand who is liable for errors.

What does the EU AI Act mean for a café?

For a normal business using AI for texts and plans, the effort is manageable. As of November 2026, as far as we can source it:

• Chatbots: since 2 August 2026, providers have to build AI systems so people are told they're dealing with AI, unless that's obvious (Article 50(1)). This falls on the provider first. Check that your tool does it anyway.
• Realistic fake images and video: whoever uses such AI content has to disclose it (Article 50(4)).
• AI literacy: Article 4 also covers businesses using AI. After the change made by the AI Omnibus, the duty is to support literacy, and nobody has to guarantee a set level. A short briefing for your team is a sensible start.

Data protection comes on top. The guidance from Germany's data protection authorities warns that removing a name and address often isn't enough, because the context can still show who is meant. "The Tuesday regular with the dog" may already be too much. If you're unsure about a specific case, ask your local chamber of commerce (IHK).

How can guests ask about the menu without the AI making things up?

A guest tool should only know what you've published yourself, hand allergy questions to people, and not keep what guests ask. That's how Ask the menu is built, as part of the NFCeez menu. There's nothing to set up: once your menu is published, guests ask on your Smart Page in their own words, in up to 25 languages. The answer comes only from your published menu, in three short sentences at most. On allergies it names only what you've declared, never promises that something is free of an allergen, and asks the guest to tell your team. Change a price, and the next answer has it. The app shows you how often guests ask. NFCeez doesn't store the questions themselves.

The limits: like any AI it can get things wrong, and answers aren't checked beforehand. That's why handing allergies to people matters so much. Guests can put dishes on a list and show it to your staff, but they order from your staff. Questions are limited per guest and per day.

What can you skip for now?

Wait on anything that promises more than you can check. Forecasts for purchasing and staffing, for example, need clean data over a long time. We found no independent study measuring the benefit for small cafés. That doesn't mean it never helps. It means: ask every vendor how they measured, and on whom.

Your next step: take one real job from the table today, ideally the closing checklist, and have it drafted. Read it through with someone on your team and cross out what doesn't fit. You'll learn more from that than from any webinar.