Skip to content
Novot AI
NLBook a call
← Back to knowledge base

AI and guest contact in hospitality

In hospitality an AI assistant can take, move, confirm and remind bookings and answer questions about opening hours or the menu. Allergen questions are different: they fall under the information duty in Regulation (EU) No 1169/2011 and touch food safety. Let the assistant relay only what the kitchen recorded, word for word, and hand the conversation over as soon as someone asks whether they can safely eat something. Allergy notes are health data and belong to that one booking; no-show history is a judgement about a person and should not be written by an AI.

Published on

An AI assistant in a restaurant looks like an AI assistant in a web shop, but it is not the same thing. In a web shop, things go wrong when a customer is told the wrong delivery date: annoying, fixable. In a restaurant, things go wrong when a guest with a nut allergy is told the sauce is safe while it contains pesto. That is not a customer service question handled badly. That is a food safety incident.

That difference shapes how you set up AI in hospitality. The general privacy rules apply to you as well, of course, and they are explained elsewhere in this knowledge base. This article is about what is different in hospitality: the duty to provide allergen information, what you record when a table is booked, what you do with no-shows, and where the line runs between an assistant that takes pressure off your floor staff and a system that misinforms guests.

Why is an allergen question not an ordinary guest question?

Most questions that reach a restaurant are logistics. Are you open on Sunday, can you seat eight, is there parking, do you have a high chair. If an AI assistant gets one of those wrong, someone stands in front of a closed door or you have to reshuffle a table. Awkward, but recoverable.

An allergen question sits in a different category. The answer you give decides whether someone puts something in their mouth. With a peanut or nut allergy, a wrong answer can lead to a serious reaction within minutes. With coeliac disease the damage is slower but real. The guest cannot check your answer: they do not see the sauce being made and they have no idea which brand of stock sits in your kitchen.

That is why this is also legally different. It is not only a question of handling data properly, but of meeting the duty to inform people about food. And it is a question of whether you are stating something that is factually incorrect.

What does the law say about allergens in non-prepacked food?

The European rules on food information are in Regulation (EU) No 1169/2011. Article 9(1)(c) covers the declaration of substances that cause allergies or intolerances. Those are listed in Annex II to the regulation, from gluten and nuts to lupin and sulphites. Article 44 extends that duty to food that is not offered prepacked:

Where foods are offered for sale to the final consumer or to mass caterers without prepackaging, or where foods are packed on the sales premises at the consumer's request or prepacked for direct sale: (a) the provision of the particulars specified in point (c) of Article 9(1) is mandatory

That is exactly what a restaurant, a sandwich bar, a baker, a butcher or a caterer does. The plate your chef sends out carries no label, so the information has to come from somewhere else: from the menu, from a folder behind the bar, or from a member of staff who knows.

Ordering online sharpens the point. The Dutch Food and Consumer Product Safety Authority puts it like this (translated from the Dutch original):

Do you sell food online? Then you must state in the web shop which allergens are present in the products. You may also refer to a telephone number where people can request allergen information directly.

Note what that says. The alternative to stating it in the ordering environment is a phone number where people can obtain the information directly. If that number reaches an AI phone assistant, the assistant is no longer a nice extra. It is the way you meet a statutory duty to inform. What it says is what you say.

Which guest question may an AI answer on its own?

The useful rule of thumb is not "easy versus hard", but: can the answer come from a source that is correct today, without anyone having to make a judgement? Opening hours hold until you change them. Whether the fryer was also used for croquettes this afternoon is known only to someone standing in the kitchen.

Guest questions and whether an AI assistant may handle them on its own
Guest questionHandled by AI alone?Why
Are you open on Boxing Day?YesA fixed fact from your own calendar. Getting it wrong is a nuisance, not a danger.
Can you seat four at eight o'clock?YesComes straight from the booking system. The guest sees immediately whether it is right.
What does the three-course menu cost?Yes, if the menu is currentQuoting the wrong amount is an incorrect statement to a consumer.
Are there nuts in the pesto?No, only relay what the kitchen recordedThe assistant cannot check the preparation and the supplier may have changed.
I have coeliac disease, can I eat here safely?NoThis is about cross-contamination in today's kitchen. That is not a fact, it is a judgement.
Can the chef make the risotto without butter?No, pass it to the kitchenAgreeing to change a dish is a promise on behalf of the brigade.
Can I order a bottle of wine in advance?Taking the order is fine, serving it without a check is notAge has to be established when the drink is handed over, not assumed.
I want to move my bookingYesAn action in your own system, with a confirmation to the guest.
Why was my booking refused?NoIf a no-show history sits behind it, that is a judgement about a person.

The right-hand column is the point. Above the line the assistant retrieves something. Below the line it would have to judge, promise or reassure. That last part belongs to a human being who is actually in the building.

How can you let an AI talk about allergens at all?

Doing nothing is not a solution either. If nobody picks up the phone in the evening, the guest with an allergy gets no information at all. The workable form is an assistant that relays rather than advises. In practice that means three things.

  • One source, maintained by the kitchen. The allergen matrix per dish belongs to the chef, not to the marketing copy on the website. The assistant reads from it and invents nothing. What is not in there is not known, and "not known" is a valid answer.
  • Quote literally, do not summarise. "This dish is recorded as containing gluten, egg and nuts" is something else than "this dish is fine for you". The first relays information, the second delivers a verdict.
  • A hard handover. As soon as someone uses the word allergy or intolerance together with a decision ("can I eat this?"), the conversation goes to a person or to a call-back arranged before the order is final. Not after the confirmation.

The trap you see most often in practice: the menu changes, the data behind the assistant does not. A new supplier of brioche, a different stock, a seasonal dish added. Allergens sit at product level, not at menu level. Change the menu without changing the matrix and you leave behind an assistant that gives outdated information with full confidence.

What happens if the assistant says something wrong?

Alongside the food safety risk runs a consumer law line. Article 193c of Book 6 of the Dutch Civil Code states that a commercial practice is misleading "if information is provided that is factually incorrect or that misleads or may mislead the average consumer". It does not say you have to have done it on purpose. An assistant that confidently gives the wrong answer about the contents of a dish, a price or a cancellation condition is providing factually incorrect information. That it was a machine is no defence: you deployed it, so you are the one speaking.

The Netherlands Authority for Consumers and Markets describes the consequences plainly: in the case of an unfair commercial practice, consumers may get out of the contract by annulling it, and they can claim damages, and that applies to all customers in the European Union. For a restaurant that bites hardest with groups and packages: a party that booked on the basis of a wrong assurance stands stronger than you might think.

Practical consequence: keep a record of what the assistant said. Not to catch the guest out, but because otherwise you have nothing in your hands in any dispute.

What do you record when a table is booked, and what not?

A booking needs very little: a name, a phone number or email address, a time and a number of people. Everything on top of that has to be something you can explain. An AI that holds a conversation collects too much more easily than too little, because guests volunteer all sorts of things in a chat.

Two kinds of note deserve extra attention.

Dietary and allergy notes. "Has a nut allergy" is information about someone's health. Article 9 of the GDPR does not start neutrally there, but with a prohibition: processing of data concerning health, among other categories, is in principle not permitted unless an exception applies. In hospitality that is usually the explicit consent of the guest, who gives you the information because they want the kitchen to know. That means the note belongs to that booking and that purpose: cooking safely. It is not a profile attribute you then use to send targeted mailings about gluten-free evenings.

Free text. "No pork" can say something about religion. "Wheelchair" says something about health. "Marriage proposal" is private. An AI that summarises the conversation into a guest note records this kind of thing unasked, and such notes often sit there for years. Limit what lands in the guest profile automatically and keep sensitive points attached to the current booking only.

One more point specific to hospitality: the person booking for eight sometimes passes on the dietary needs of seven other people. Those seven know nothing about it. Keep such notes short, without names attached, and delete them once the dinner has happened.

How long may you keep an allergy note against a guest?

Two things get mixed up here that you need to pull apart. The bill is accounting and falls under the tax retention duty; that sits outside this article. The note "sesame allergy" is not accounting. It belongs to the cooking, and once the dishes are done the cooking is over.

Restaurants still like to keep such a note, and there is a good argument for it: a regular appreciates not having to explain it every time, and it is safer too. That is possible, but then keeping it is a deliberate choice the guest knows about. Simply ask: "shall we save this with your details for next time?" Someone who can say yes can also say no, which means the assistant has to be able to ask the question and record the answer.

What you should not do is let the note ride along quietly because the system stores everything anyway. A guest who requests their data will read what is there. An old note that is no longer accurate is dangerous as well: allergies change, and an outdated line in the system can lull the kitchen to sleep.

What do you do with no-shows without building a blacklist?

No-shows are a real problem, and an AI assistant genuinely helps against them: confirming, reminding, and making cancellation easy. That is the clean route, and it reduces the problem most.

Recording them is something else. The moment you note that a specific guest failed to turn up and use that on a later booking, you are no longer keeping bookings but a judgement about a person. That makes it more sensitive than the rest of your guest data. A few practical lines.

  • A count ("did not show twice") is factual. A label ("unreliable", "difficult party") is an opinion, and the guest reads it too if they request their data.
  • Do not let an AI write those labels itself. An assistant that distils "difficult guest" out of an irritated phone call is making a note nobody has checked.
  • Do not let the assistant refuse on its own. It may say there is no table at that time; a refusal based on history should be taken by a person who can explain why.
  • Put an end date on it. A no-show from three years ago says nothing any more. Set a period you can justify and have the system erase it afterwards.
  • Do not share it outside your own business. A list circulating among operators is something quite different from a note in your own system, and it demands a far heavier justification.

If you ask for payment details as a no-show guarantee, that has to be clear up front: that something is being stored, when a charge would follow, and until when cancellation is free. An assistant that slips the question in casually turns a booking into a surprise.

May an AI take orders that include alcohol?

Taking the order is fine. Serving or delivering it is the moment that counts. Article 20 of the Dutch Alcohol Act is strictly worded: it is prohibited to supply alcoholic drink commercially, or otherwise than free of charge, to a person of whom it has not been established that they have reached the age of 18. Established, not assumed. A tick box in a chat is not an establishment, and an AI on the phone cannot look at an identity document.

For a place with table service that is no problem: the check happens at the table or at the door. For delivery and collection it is, because there the driver or the counter staff are the only moment of control. So make sure an order containing alcohol arrives visibly flagged at the delivery stage, and that the assistant states in the confirmation that identification will be requested on handover. That also prevents an argument on the doorstep.

Where is the line between a useful assistant and a system that misinforms guests?

The line does not run along the technology but along the kind of sentence the assistant produces. Relaying, scheduling, confirming and reminding are safe acts: the guest can see for themselves whether it is right and nobody gets ill from it. Judging, reassuring and promising are not. "I'm sure the chef can arrange that" is a promise on someone else's behalf. "That is not in there" is a verdict about a plate that has yet to be made.

Three things keep that line standing in practice.

  1. Make clear that it is AI. At the first message or the start of the call, not somewhere at the bottom. A guest who knows they are talking to a system is more likely to push back on an answer that does not sit right. Here that is not a formality but a safety measure.
  2. Keep the way out open. One sentence is enough: "would you like to speak to one of us?" If that way out leads nowhere outside opening hours, the assistant should say so and arrange a call-back rather than take the gamble itself.
  3. Test with the questions you dread. Not with "are you open on Sunday", but with "my son is allergic to milk, can he have the pancake?" What the assistant says to that is the real measure of your setup. Run that test again after every menu change.

Set it up that way and you keep what an AI is genuinely worth in hospitality: a phone that gets answered during the rush, bookings that do not disappear into a full inbox, and floor staff who are not called away halfway through service. That is gain enough. The question of whether there are nuts in the sauce stays with the kitchen.

See what Novot AI can do for your business.

Book a call