Can an AI enter a contract on your behalf and legally bind you?

Imagine asking your artificial intelligence to find a hotel for a meeting. You give it the dates, the location and an approximate budget. The agent compares options, exchanges messages with the hotel and finds a suitable room.

So far, it looks like a slightly more sophisticated search.

But then it confirms a non-refundable rate.

Or consider a company using an agent to manage its email. A supplier proposes renewing a service on new terms. The system interprets the offer as consistent with its instructions and replies: “Agreed, you may proceed.”

Has it booked the hotel? Has it renewed the contract? Can the user subsequently reject these transactions on the grounds that an artificial intelligence accepted them?

Answering requires examining how each transaction took place. An automated action can create legal obligations even when a person has not individually reviewed every step.

This issue starts to affect us as soon as we give a system the ability to act outside the conversation.

Meta provides a concrete example. In its presentation of Muse, it describes a personal agent that can browse, complete forms, manage tasks and negotiate for its user. It also announces purchases through Stripe’s Link and requests for approval before certain actions, such as sending emails or buying something. These are capabilities and safeguards described by the provider, rather than a guarantee of flawless operation or universal availability. [1]

The economic dimension also deserves attention. Zuckerberg has expressed an expectation of generating revenue through a small transaction fee. That describes a business intention; it does not establish a particular tariff or general charging terms. [2]

If the assistant participates in our purchases and its provider earns revenue from them, another question arises: how do we know which interests influence its recommendations? This should be examined alongside the service terms and the information given to users.

First, we should distinguish between recommending, negotiating and accepting.

An agent may simply compare offers. It may request quotations without committing anyone. It may discuss terms subject to approval. Or it may issue an acceptance intended to close the transaction.

These are different actions, even if they appear within the same conversation on a screen.

Important decisions also arise without a purchase:

Agent action: replying to an offer by email. Question to examine: whether the message constitutes acceptance. Agent action: renewing a subscription. Question to examine: duration, price and cancellation options. Agent action: negotiating a complaint. Question to examine: whether it accepts compensation or waives rights. Agent action: purchasing for a company. Question to examine: authority, budget and authorised terms. Agent action: cancelling a booking. Question to examine: penalties and loss of rights. Agent action: accepting a platform’s terms. Question to examine: obligations assumed and data that may be used.

These are scenarios for analysis. Their effects depend on the instructions, communications and applicable rules.

In Spain, electronic contracting already has an established legal basis.

Article 1262 of the Civil Code addresses contracts concluded through automatic devices and locates consent in the manifestation of acceptance. Using a machine therefore does not, in itself, prevent a contract from existing.

Article 1259 addresses contracting on another person’s behalf without authority and provides for subsequent ratification. However, an AI agent requires us first to clarify how its action is attributed: it may operate as the user’s instrument, execute previously defined authority, or produce a declaration exceeding its instructions. Its commercial description as an “agent” does not resolve that classification. [3]

The Information Society Services Act recognises the effectiveness of electronic contracts when consent and the other requirements for validity are present. It also admits electronic records as documentary evidence and preserves special requirements for certain transactions. An agent capable of completing a form does not eliminate a legal formality. [4]

There is a practical consequence: we should not assume that “I did not personally press the button” will provide a sufficient defence. We will need to reconstruct what was authorised, what the system did and what the other party received.

Norway raises similar questions, although the analysis must begin with its own law.

The avtaleloven governs contract formation and authority to represent another person. Section 10 provides that an action performed in the principal’s name and within the authority granted can directly create rights and obligations for that principal. Section 11 distinguishes between the limits of that authority and internal instructions, taking account, among other circumstances, of what the counterparty knew or ought to have known.

These rules help identify the issue, but we should not automatically apply the entire legal construction of a human representative to AI. We need to determine whether an automated declaration is attributable to the user and how the relevant rules operate. Provisions concerning mistake do not provide an automatic escape from every unexpected result either. [5]

An organisation working between Spain and Norway also needs to establish which law governs a transaction. The assistant’s language or the server’s location is not sufficient to answer that question.

The European Union adds several layers of protection, but does not provide a single answer to every contract made through AI.

The Electronic Commerce Directive requires legal systems to allow electronic contracting and not deny a contract legal effect solely because that medium was used. This principle facilitates automation; it does not, by itself, cure missing consent, insufficient authority or an unmet legal formality. [6]

In consumer relationships, prior information, the total price and transaction terms remain important. In Spain, Article 98 of consumer legislation requires express acknowledgement that an order entails an obligation to pay and establishes consequences if this requirement is not met.

How that acknowledgement should be evidenced when an agent is involved deserves specific examination. We cannot presume that every automated click satisfies every requirement. Nor can we assume that every purchase can be undone: withdrawal rights have conditions and exceptions. [7]

The Artificial Intelligence Act addresses, among other matters, transparency for systems interacting directly with people. Disclosing that someone is talking to AI may be mandatory in the circumstances covered. But that information does not establish the system’s authority to accept an offer. Identifying the interlocutor and establishing its authority are separate checks. [8]

In Norway, incorporation of each instrument into the EEA framework and its domestic application must also be checked. The EFTA entry consulted keeps the AI Act under scrutiny for incorporation; it should not automatically be described as applicable Norwegian law. [9]

Email may become one of the most sensitive areas.

An assistant with email access may encounter offers, complaints and proposed contractual amendments. Before allowing it to reply, we should distinguish what it may prepare, what it may send and what it may accept.

Imagine the instruction: “Manage renewals that do not exceed 200 euros.”

Is that 200 euros per month, per year or per transaction? May it accept a three-year commitment? May it change suppliers? What happens if the price is within the limit but the terms include a substantial penalty?

The amount alone is a poor description of the authority granted.

We also need to decide how incoming messages are treated. An email may contain instructions intended to manipulate the agent. I therefore propose a design rule: third-party content should be treated as information to examine, without automatically converting it into an authorised instruction from the user.

My practical proposal is to define authority before expanding autonomy.

A useful mandate should specify permitted transactions, suppliers, budget and time limits. It should also identify unacceptable terms and when a person must intervene.

Approval should display the transaction that will actually be executed: counterparty, total amount, duration, renewals and relevant conditions. It should be tied to that version. If the terms subsequently change, I propose requiring fresh approval.

As a design reference, Muse’s own connector guidelines call for relevant information before commitment, a distinction between pending and confirmed transactions, avoidance of duplicates and the provision of receipts. These are platform rules, rather than a general legal standard, but they illustrate operational problems worth resolving. [10]

Sufficient evidence of what happened must then be retained: instruction, offer, authorisation and executed action, with appropriate access controls and retention arrangements.

If a dispute arises, that record helps separate two questions: whether an obligation exists towards the supplier, and whether a claim may be brought concerning the AI service’s performance. The second does not automatically resolve the first.

International legal work is already addressing these issues.

In 2024, UNCITRAL adopted a Model Law on Automated Contracting addressing the use of AI in contract formation and performance, the attribution of outputs and certain unexpected outputs.

It indicates the direction of the debate. As a model law, it is not automatically applicable legislation in Spain or Norway. [11]

In the agent projects we are building, these questions have architectural consequences from the outset. If Alonides is to coordinate my projects and dispatch instructions, I need to distinguish between coordinating work and entering commitments on my behalf.

I may want it to search, compare and negotiate. I may authorise certain routine transactions. And I may reserve other decisions for my own review.

That allocation needs to be defined and verifiable.

The answer to the initial question is therefore that a transaction performed through AI can potentially bind you legally. Establishing whether it does in a particular case requires examining attribution of the declaration, the authority granted, acceptance and the applicable rules.

Before asking AI to take care of something, it is worth specifying how far it may go.

“Find me a good option” and “accept that offer on my behalf” deserve different instructions.