Who is responsible when an AI agent acts for a user?
Written by Brooke Ash, lawyer at For Founders Law.
Short answer
Usually the person or business that deploys the agent. Ontario's Electronic Commerce Act, 2000 says a contract can be formed by the interaction of an electronic agent and an individual, or between electronic agents, so an agent's purchase or booking can bind someone. Section 21 lets an individual avoid a transaction with another person's electronic agent for a material error if the agent gave no chance to correct it. If you build agents, define what they're authorized to do, confirm consequential actions, keep logs, and allocate responsibility clearly in your terms.
An AI agent doesn’t just answer questions. It books, buys, sends, signs up and changes settings, often without a person reviewing each step. That raises a simple question with a less simple answer: when the agent does something, who is bound and who pays if it goes wrong?
Canadian law doesn’t treat an AI agent as a legal person. Responsibility falls on the people and businesses involved: the user who set the agent loose, the company that built or deployed it, and the business on the other side of the transaction. Contract terms decide much of the split.
Can an AI agent form a binding contract?
Yes, in Ontario. The Electronic Commerce Act, 2000 defines an “electronic agent” as a computer program or other electronic means used to initiate an act or respond to electronic documents or acts, in whole or in part, without review by an individual at the time. Section 20 says a contract may be formed by the interaction of an electronic agent and an individual, or by the interaction of electronic agents. Section 19 lets an offer or acceptance be expressed electronically, including by clicking, and says a contract isn’t invalid only because it’s electronic.
The Act was written for simpler automated systems, and courts haven’t yet tested it much with today’s AI agents. But the starting point is that an agent’s purchase or booking can create a real contract for whoever deployed it.
What happens when the agent makes a mistake?
It depends on whose agent it is.
Your user deals with your agent. Section 21 says an electronic transaction between an individual and another person’s electronic agent isn’t enforceable by that other person if:
- the individual made a material error in the electronic information or document used;
- the agent gave no opportunity to prevent or correct the error;
- the individual promptly notified the other person on becoming aware of it; and
- where consideration was received, the individual returns or destroys it as instructed, or deals with it reasonably, and doesn’t benefit materially from it.
The practical lesson is to build in a confirmation and correction step before your agent completes a transaction.
Your agent acts for your user. If your product’s agent buys the wrong thing from a third party, section 21 doesn’t obviously help, because the error was made by an agent, not an individual. The user may be bound to the third party, and will look to you. Your terms and your product design decide whether that claim succeeds.
Your agent says something wrong. In Moffatt v. Air Canada, 2024 BCCRT 149, a British Columbia tribunal held the airline responsible for its chatbot’s incorrect statement and rejected the argument that the chatbot was a separate entity responsible for its own actions. Expect the same reasoning for agents. See is my company liable for what its AI chatbot says.
How should authorization and payments work?
The clearer the user’s instructions, the easier it is to say the agent acted within them.
- Scopes. Let users set what the agent may do: which sites, which actions, which accounts.
- Spending limits. Cap amounts per transaction and per period, and require confirmation above a threshold.
- Confirmation for consequential actions. Purchases, cancellations, sending messages to third parties, accepting terms and sharing personal information should need explicit approval, at least by default.
- Payment credentials. Use your payment provider’s tools for delegated or tokenized payments rather than storing card details, and follow its rules on customer authorization.
- Third-party terms. Many websites prohibit automated access in their terms of use. An agent that ignores them can expose you and your users to claims or blocked accounts.
What should our terms say?
Your AI product terms should address agents directly:
- The user authorizes the agent to act on its behalf within the settings it chooses, and is responsible for those settings.
- The agent can make mistakes, and the user should review actions and confirmations.
- What happens when the agent exceeds its instructions, including how users report errors and what you’ll do to help unwind them.
- Which third-party services the agent may interact with, and that their terms apply.
- Your limitation of liability for agent actions, recognizing that limits are harder to enforce against consumers.
Accept these terms with an active step, such as a checkbox. See are clickwrap agreements enforceable.
Do consumer protection and privacy rules apply?
Yes. If consumers use your agent, Ontario’s Consumer Protection Act, 2002 applies to your agreement with them whatever your terms say, and its internet agreement rules include giving consumers an express opportunity to accept, decline or correct errors. See online terms and Ontario consumer law. Marketing that overstates what the agent can do safely can be a misleading representation under the Competition Act.
Agents often read inboxes, calendars and accounts. Under PIPEDA, collect and use only the personal information the task needs, explain it clearly, and protect it. If the agent shares a user’s information with third parties, that should be clear to the user before it happens.
Why does logging matter?
When something goes wrong, the first question is what the user instructed and what the agent did. Keep records of user instructions and settings, confirmations, the actions taken and the responses received. Logs help you show the agent acted within its authorization, support refunds and dispute handling, and find recurring errors. Treat them as personal information where they contain it, with sensible retention limits.
Federal policymakers are paying attention: ISED’s 2026 consultation on AI transparency asked about monitoring AI agents. Our AI Law Tracker follows what comes next.
What to do next
We prepare AI product terms and terms of service that cover agent authorization, errors and liability, at fixed fees on our fees page. See also legal help for AI startups and what legal work an AI startup needs.
This guide is general information about Ontario and Canadian law, not legal advice for your situation. To talk through an agent product, book a free 15-minute consult.
Sources
- Electronic Commerce Act, 2000 (Ontario), s. 1 (definition of electronic agent) and ss. 19 to 21
- Moffatt v. Air Canada, 2024 BCCRT 149
- Consumer Protection Act, 2002 (Ontario)
- Competition Act, s. 52 (false or misleading representations)
- Personal Information Protection and Electronic Documents Act, Schedule 1
- Innovation, Science and Economic Development Canada: Consultation on AI transparency (July 2026)
How We Can Help
- SaaS or Master Services AgreementYour customer contract, ready for business customers. From $920
- AI Product TermsInputs, outputs and training use, in plain terms. From $760
Fixed fees. Government filing fees are extra.