Automated Processing and AI

The ordering assistant is an artificial intelligence system. It is not a person, it is not a member of restaurant staff, and it does not become one at any point in a conversation. We tell you this at the start of every conversation rather than relying on it being obvious.

Article 50 of the EU AI Act requires exactly this disclosure for a system that interacts directly with a person, and it applies from 2 August 2026. We make the disclosure in both our markets rather than only where it is compelled.

It proposes. It reads the menu, reads what you asked for, and suggests a basket of dishes. You confirm it or you do not, and nothing is ordered until you do.

It cannot price anything. The model is given no field in which to return a price, a total, a discount or a fee — the shape of its answer has nowhere to put one — and every figure you see is calculated afterwards by the same pricing code that runs at checkout. There is no model on the payment path at all.

It cannot place an order, apply a refund, or change anything about your account. Its output is reduced to menu item identifiers and quantities before anything else happens.

It does not give medical, dietary or allergen advice, and it will not tell you a dish is safe for you. It can repeat what a restaurant has stated about a dish and tell you what the restaurant has not answered. It cannot go beyond that, because we hold no recipe and no ingredient list to go beyond it with.

Nothing you say to the assistant is stored as a fact about you unless you tap to confirm it. The assistant may notice something and offer to remember it; the offer is a question with a button, and only the button writes anything. The code that writes has no connection to the model at all, so there is no sequence of model outputs that can produce a stored fact.

It never records an allergy it was not told. It cannot infer one, and this is enforced by the shape of the software rather than by a rule somebody might change: the structure that holds a behavioural observation has no field for severity, so an inferred allergy has nowhere to live.

Everything the assistant knows about you is visible to you in your account, each item labelled with how it was learned, and each item individually deletable.

Building a picture of what you like from what you order is profiling within the meaning of Article 4(4) of the GDPR. We do it, and we say so.

Article 22 of the GDPR gives you the right not to be subject to a decision based solely on automated processing which produces legal effects concerning you or similarly significantly affects you. Our view is that suggesting dishes does not reach that threshold: a ranked list of suggestions determines nothing, you can ignore it, the whole menu remains available to you, and nothing is added to your basket without you doing it.

That answer would change if the profile started doing more. If it set your price, refused you service, suppressed restaurants or dishes from what you can see, or decided whether you could pay after delivery, then it would be making a decision about you and Article 22 would engage. We do not do any of those things, and none of them exists in the software.

You can ask us for an explanation of how a suggestion was arrived at, and we will answer in terms of what was actually taken into account.

A suggestion that fits something we know you like is a suggestion. A suggestion designed to exploit something we know is wrong is not, and we do not make the second kind.

Concretely, and as commitments rather than aspirations: we do not use knowledge of a vulnerability — a person's age, a disability, or a difficult financial situation — to push them to order more or to order at all. We do not use a declared allergy or a health condition to target anything at anybody. We do not manufacture urgency or scarcity that is not real. We do not make it harder to leave than it was to arrive, and we do not use a dark pattern to obtain a consent.

Article 5 of the EU AI Act prohibits AI systems that use manipulative or deceptive techniques to materially distort behaviour in a way likely to cause significant harm, and separately prohibits exploiting vulnerabilities arising from age, disability, or a specific social or economic situation. Those prohibitions have applied since 2 February 2025. Ordinary commercial persuasion is not caught by them and we do not pretend otherwise — what is caught is the second kind of suggestion, and the line above is where we hold ourselves.

Suggesting a side dish with a main course is upselling. Suggesting a larger order to somebody a model believes is in financial difficulty is not something we will build.

No. The EU AI Act designates specific uses as high risk, and helping somebody choose lunch is not one of them. Nothing here does biometrics, credit scoring, employment decisions, education, law enforcement or access to essential services.

Profiling does not by itself make a system high risk. It removes an exemption that only applies to systems already listed as high risk, which this is not.

Two adjacent systems would be a different question if we build them, and we would treat them as such: anything that allocates work to couriers or evaluates their performance, and anything that scores a customer for deferred payment. The first is listed under employment and workers' management; the second under access to essential private services.

Where the assistant produces text for you to read, it is the assistant's text and it is labelled as such in the interface.

A person confirms every order. If you want a person from our side, ask for one and a complaint about an order is answered by a human being, not a model.

The software wins, and the notice is wrong and will be corrected. We maintain an internal record mapping every claim in this notice to the part of the system that supports it, and to the change that would make the claim false, so that a change to the software is visible as a change to what we may say.