AI agents and contracting: is the business bound by what the agent agrees?

The procurement department has given an AI agent access to three supplier portals and the shared purchasing mailbox. The agent is to collect quotes for packaging and order the cheapest. One Monday there is a confirmed order for 40,000 cartons instead of 4,000, and the supplier’s own system accepted it in less than a second. The example is invented. The managing director wants to know whether the business has to pay.

The question becomes more practical every month. On 6 October 2026 OpenAI announced a partnership with the contract platform Ironclad to train agents on contract work. According to the trade press coverage, this covers reviewing documents, comparing clauses, proposing amendments and routing for approval, so not signing. But in a contract management system the step from proposal to acceptance is often a single click.

Who must cover the loss an agent causes is dealt with in the article on liability for AI agents. This article is about the contract itself.

Do contracts concluded by AI agents bind the business?

Yes, as a rule. Chapter 1 of the Norwegian Contracts Act (avtaleloven) does not require an offer or acceptance to be formulated by a human, and ordinary commercial contracts are not subject to any formal requirements. Under Contracts Act § 1, the rules also give way to what the parties have agreed and to trade usage and custom. Automatic orders from inventory systems and EDI solutions have been in use for decades without anyone arguing that they are invalid because nobody read them before they were sent.

What decides the matter is how the counterparty could reasonably understand the order. An order from the business’s account, in the business’s name and through a channel the business itself has opened gives the counterparty strong grounds for believing that it comes from the business. In our view the agent’s order is the business’s own declaration of intent, made through a tool the business has chosen.

The fact that a language model is not predictable in the way a rule-based purchasing system is does not change this. It was the business that chose a tool capable of surprising it. Norwegian law has no specific rules on contracts concluded by AI agents, and as far as we are aware no Norwegian court has ruled on the question.

Do the rules on authority apply when an AI agent acts?

Not directly. Under Contracts Act § 10(1), an agent binds the principal when acting in the principal’s name and within the limits of the authority. The rules in Chapter 2 assume that the agent is a person. Even so, they give us the best vocabulary available for drawing the lines, and several of them work well by analogy.

The last row matters. The rules on authority are built on the counterparty having someone to turn to if the authority turns out to be lacking. When the agent is software, there is no such person. In our view that points towards the business itself being bound.

A spending limit in an internal instruction stops your own employees. A wrong order is only stopped by a limit in the system.

What if the AI agent makes a mistake?

Then Contracts Act § 32(1) is the first place to look. A person who has made a declaration with a different content than intended, through a slip of the pen or a similar mistake, is not bound if the recipient realised or ought to have realised the error. With ten times the usual volume and a supplier who knows the customer’s purchasing pattern, the business is in a strong position. With an order that is unusual but not conspicuous, its position is weak.

It is unsettled what a counterparty “ought to realise” when it is the counterparty’s own system that reads the order. We assume the assessment will turn on what the business behind that system ought to have caught with reasonable controls.

It is tempting to compare the agent with the telegraph in Contracts Act § 32(2). The sender is not bound by a declaration that has been distorted in telegraphic transmission, even if the recipient acted in good faith. But the telegraph was an independent intermediary outside the sender’s control, whereas the agent is the sender’s own tool. We do not think a court will stretch the analogy that far.

Outside the rules on mistake, Contracts Act § 33 may help where the counterparty knew of circumstances that make it dishonest to hold the business to its declaration. Contracts Act § 36 allows unreasonable contracts to be set aside or adjusted, but it is applied with restraint between businesses. Either way, one practical rule applies. Notify the counterparty at once. Waiting weakens your own case.

What does the E-Commerce Act say about automated orders?

Little about whether the order binds. Under E-Commerce Act § 11(3), a party selling electronically must make it possible to detect and correct input errors before the contract is concluded. Under E-Commerce Act § 12, an electronic order must be acknowledged without undue delay. Both rules can be departed from by agreement outside consumer relationships, and the obligations do not apply to contracts concluded exclusively by email or equivalent individual communications.

The Norwegian E-Commerce Act (ehandelsloven) was written for a human in an online shop. Whether messages between two agents count as “equivalent individual communications” has not been settled. For the business the point lies elsewhere. A supplier portal that lacks an acknowledgement or a way to correct errors gives you an argument when an agent’s order is to be undone.

What does the UNCITRAL Model Law say about AI agents and automated contracting?

The UN Commission on International Trade Law (UNCITRAL) adopted a Model Law on Automated Contracting on 11 July 2024, and the UN General Assembly recommended it on 4 December 2024. The Model Law is not Norwegian law, but it shows how international trade law approaches the issue.

Under UNCITRAL Model Law Art. 5, a contract cannot be denied validity solely because no human reviewed it. Under Art. 7, an action is attributed to the person who uses the system for that purpose, including where the outcome was unexpected. Under Art. 8, however, the counterparty cannot rely on an action the user could not reasonably have expected, where the counterparty knew or ought to have known that the action was not expected. This is close to the reasoning behind Contracts Act §§ 32 and 33, and we expect Norwegian courts would reach much the same result.

What if the AI agent negotiates the terms?

Then the risk increases. An agent that accepts a counteroffer may agree to the supplier’s standard terms, with limitations of liability, automatic renewal and a foreign forum, without anyone having read them. Under Contracts Act § 6, a reply that departs from the offer is a new offer, and when two agents send amended terms back and forth it can be hard to establish which version was finally accepted.

The problem is the same as with vibe lawyering, where the text looks finished but nobody has checked it. Terms on liability and renewal should therefore never be accepted by an agent alone. See also limitation of liability in SaaS contracts.

How can a business limit the risk when AI agents enter into contracts?

Treat the agent as a new buyer with signing authority, and do not give it more authority than you would give a person in their first week. More on contract risk can be found on the contracts topic page and the artificial intelligence topic page.

1. Write an internal authority instruction for each agent, covering contract types, value limits and approved counterparties, and make it part of the company’s delegation of authority matrix.
2. Build the limits into the system, so that the agent technically cannot send an order above the value or volume limit.
3. Require human approval before acceptance above the threshold, with new counterparties and for contracts with long commitment periods, limitations of liability or automatic renewal.
4. Agree ground rules with regular suppliers in the framework agreement, for example that agent orders only bind within stated limits, that unusual volumes must be confirmed and that errors can be corrected within a short deadline.
5. Tell the counterparty that orders may come from an agent and what limits the agent has. The business can then rely on those limits.
6. Log offers, acceptances, instructions and the model version, so that it is possible to document what the agent was told and what it did.
7. Review the terms of the agent tool supplier, which often disclaim liability for exactly this kind of error.
8. Set up a procedure that stops the agent and notifies the counterparty the same day an error is discovered, with a named person responsible and a ready-drafted message.

Questions and answers

Next legal review: 15 November 2026