Who Is Responsible When an AI Agent Makes a Mistake?

When an AI agent makes a mistake, the business that deployed it is generally responsible toward its customers. The agent is not a separate legal actor, and vendor contracts rarely shift that responsibility entirely. A Canadian tribunal said so plainly in Moffatt v. Air Canada (2024), and the EU's revised Product Liability Directive extends strict liability to software, including AI, from December 2026. In practice, accountability means a named owner for every agent, permissions and approvals proportionate to risk, evidence of what the agent did, and a plan for fixing mistakes. This is practical governance, not legal advice.

Moffatt v. Air Canada (2024 BCCRT 149). A customer relied on the airline's website chatbot, which wrongly told him he could apply for a bereavement fare retroactively. Air Canada argued, in effect, that the chatbot was responsible for its own statements. The British Columbia Civil Resolution Tribunal rejected that, held the airline liable for negligent misrepresentation and noted it had not taken reasonable care to ensure the chatbot was accurate. The amount was small; the principle is not: a business cannot disown what its AI tells customers.

EU Product Liability Directive (Directive (EU) 2024/2853). The revised directive explicitly treats software, including AI systems and software delivered as a service, as a product, with strict liability for damage caused by defects. Member States must apply it to products placed on the market from 9 December 2026. It also covers defects introduced by updates and failures to provide security updates.

Other regimes (consumer protection, data protection, sector regulation, the EU AI Act for certain uses) add obligations depending on what the agent does. The common thread: responsibility follows the business that puts the system in front of customers or into its operations.

External liability is settled by law and contracts; internal accountability is a design choice. Every agent in production should have clearly named roles.

Accountability is easier when the agent cannot do much damage. Tier each agent by the worst plausible mistake and apply controls accordingly.

The Air Canada case was about information, not an action. Customer-facing agents should answer from approved, current sources (policies, prices, terms) rather than general model knowledge, cite or link to the governing policy, avoid promising things outside policy, and hand off to a person for exceptions. Disclose that customers are dealing with an AI system where rules require it. See AI customer support automation.

When something goes wrong, you need to reconstruct exactly what the agent saw and did, and to show that reasonable controls existed. Keep:

Read vendor terms for data use and retention, security commitments, service levels, liability caps and indemnities, and how model changes are communicated. Expect liability caps; plan your own controls on the assumption that you carry the customer-facing risk. Keep the ability to switch providers, so a vendor's change in behaviour or terms does not leave you exposed.

Accountability shows in the response: stop or limit the agent, correct the outcome for affected customers, find the cause, fix the control that failed and record what changed. A prepared runbook makes this fast; see AI agent incident response. For approval design, see human-in-the-loop AI; for the organization-wide framework, AI governance framework.

An AI agent's mistakes are the business's mistakes. Accept that early and design for it: named owners, controls proportionate to risk, grounded customer communication, evidence of every action and a tested response plan. Done well, this is what lets a business give agents real responsibility with confidence.