Your AI makes a booking. Who is bound by the contract?

From chat to booking: why this matters now

On 29 September 2026, OpenAI introduced “dots”: AI assistants with their own cloud computer and access to connected applications. According to the product description, actions can be allowed, blocked or made subject to approval. The announcement prompted this article; it is not a judicial finding on the validity of those actions.

A hypothetical example: a user asks her assistant to book a hotel for a weekend, with a budget of EUR 400. The AI confirms a non-cancellable room for EUR 690. Must she pay? The answer does not depend solely on whether she personally clicked the booking button.

Can a robot act as a legal representative?

Saying “AI lacks legal capacity” is imprecise. Software is not a natural person whose age or capacity is assessed under section 104 of the German Civil Code (BGB). Technical capabilities alone do not give it legal personality either. A person or company remains the contracting party.

In conventional representation, a representative makes their own declaration in the principal’s name and within their authority (section 164(1) BGB). Section 165 BGB permits representatives with limited legal capacity; it does not generally authorise software to act as a legal representative.

For ordinary automated processes, a natural starting point is the technical generation or transmission of a declaration attributable to the user. For autonomous systems, applying the rules of representation by analogy is also debated. This is a legal debate, not an established rule that a prompt turns every AI agent into an authorised representative. For the debate, see Lukas, LR 2021, p. 97, sections 2.1 to 2.3.

What the Federal Court of Justice has decided about automated bookings

In 2012, the Federal Court of Justice (BGH) held that the person or company using a computer system remains the declarant or recipient. What matters is how a human recipient may understand the declaration, not how the software processes inputs (judgment of 16 October 2012, X ZR 37/12, paragraphs 17 to 19).

In that case, “not yet known” had been entered as the second passenger’s name. Despite an automated confirmation and debit, no contract of carriage arose for that passenger. That part of the fare had to be refunded, but compensation for denied boarding was not awarded (paragraphs 20 to 25). A technical success message therefore does not itself establish a contract on the intended terms.

That judgment concerned a booking system, not a modern generative AI agent. Applying its principles here is our legal assessment. The BGH subsequently applied the general interpretation principles in its judgment of 15 February 2017, VIII ZR 59/16, paragraph 23. This does not provide a supreme-court answer for every autonomous AI booking.

EUR 400 authorised, EUR 690 booked: what matters?

First, establish whether the system merely prepared a proposal or actually made a declaration to the hotel. Next come attribution, the content of that declaration and acceptance by the provider. The general rules on offer and acceptance also apply electronically (sections 145 et seq. BGB; BGH X ZR 37/12, paragraphs 13 and 19).

A budget instruction given only to the AI is not automatically a restriction apparent to the hotel. Conversely, a saved login does not justify attributing every subsequent system action to the account holder without limits. Whether the restriction prevents contract formation, permits avoidance or only concerns the relationship with the service provider must be assessed against the actual process. For autonomous agents, this is precisely where the assessment remains unsettled.

Booking errors: avoidance is not a universal escape route

A software error can be legally significant. In its judgment of 26 January 2005, VIII ZR 79/04, the BGH permitted avoidance of an automated acceptance: a correctly entered notebook price of EUR 2,650 had been corrupted to EUR 245 during data transfer. The decisive issue was an error in the declaration, not merely an unfavourable calculation (official full text, pp. 5 to 9, II.A.1 and 2).

An AI agent making a poor choice within discretion deliberately granted to it is not the same situation. A booking error requires an assessment of whether there is a legally relevant mistake under section 119 BGB. Avoidance must then be declared without culpable delay (section 121 BGB). Section 122 BGB may also require compensation for reliance loss, capped by the interest in performance and subject to the statutory exceptions.

Is there not a 14-day withdrawal right anyway?

Not for every online booking. A hotel stay for specified dates generally has no statutory withdrawal right under section 312g(2), no. 9 BGB, unless otherwise agreed. A contractual cancellation option is a separate matter.

The requirements for a consumer order carrying a payment obligation also remain relevant. Within its scope, section 312j(3) and (4) BGB makes contract formation dependent on an unambiguous ordering process; subsection (5) contains exceptions. How these requirements are met in a particular agent workflow needs separate assessment. Saying “an AI clicked” does not replace that assessment.

What users and businesses should define before deployment

Our practical recommendation is to avoid relying solely on permissions expressed in a free-form prompt. Where technically available, binding actions should be controlled by hard limits and traceable approval. This includes:

• Separate tasks: research, prepare a booking proposal or place a binding order.
• Limit amount, duration, quantity and permitted providers; require explicit approval beyond those limits.
• Display the contracting party, total price and cancellation terms before completion.
• Retain the instructions, approvals, contract version and confirmation selectively. Unnecessary personal data should not be kept in permanent logs.
• After a booking error, stop further actions and promptly have contract formation, avoidance and cancellation assessed.

Who ultimately bears the loss?

Being bound to the hotel and a potential claim against the AI service provider are two different issues. Contractual damages require, in particular, a breach of duty, resulting loss and the further requirements of section 280 BGB. Not every unwanted outcome is automatically a compensable provider error.

We help businesses assess the role of AI agents in contracting workflows through clear responsibilities, approvals, appropriate contractual terms and analysis of specific disputes. The judgments discussed here concern conventional electronic declarations. Applying them to independently deciding AI systems remains a question of legal interpretation in material respects.