19. When an AI agent causes an accident, who is liable? — Agents without legal personhood
Contents
If an AI agent given permission to send emails misdirects a message and leaks internal documents to thousands of business partners, or if an agent given payment authority makes an unapproved payment — the liability for that incident falls not on the AI, but on the organization that deployed it. The law has not yet been rewritten for the new concept of the "AI agent"; instead, existing agency law, tort law, and contract law are being applied as they are.
Background — why this question is coming up now
AI agents no longer just generate answers. Cases of them performing real, hard-to-reverse actions on an organization's behalf — sending emails, making payments, calling external APIs, submitting documents — are increasing rapidly. The question is whether an organization can avoid liability for a mistake by saying "the AI did it on its own."
Regulators and courts around the world are answering this question in almost the same direction. At Reuters' Momentum AI event on September 25, 2026, US Federal Trade Commission Chair Andrew Ferguson said, "Giving software the authority to act doesn't make that software the party responsible for it," adding that AI agents should not be treated as independent actors (as reported by Unite.AI). In the UK, the UK Jurisdiction Taskforce (UKJT) released the final version of its "Legal Opinion on Liability for AI Harms" in July 2026, maintaining its conclusion that AI harms can be addressed under existing negligence principles without granting AI systems legal personhood.
Legislation to fill the regulatory gap has already begun. Under a California law that took effect on January 1, 2026, the fact that a human did not directly approve a specific harmful action is not, by itself, enough to escape liability (per The Lyon Firm's analysis). In the end, the question narrows from "is the AI liable" to "on what basis, and to what extent, is the organization that deployed the AI liable."
Key data & current state
| Issue | Current state | Source |
|---|---|---|
| Legal personhood | The UK UKJT's final opinion (2026.7) maintains that AI systems are not granted legal personhood, concluding that negligence liability is the primary path to resolution | UKJT final opinion (reported by SCL, 2026) |
| Blocking the liability-evasion defense | California law (effective 2026.1.1) provides that the mere fact a human did not directly approve a specific action is not enough to escape liability | The Lyon Firm, 2026.4.29 |
| Regulator's official position | FTC Chair Ferguson stated, "AI agents should not be treated as independent actors" (2026.9.25, Reuters Momentum AI event) | Unite.AI, 2026.9 |
| Court precedent | The British Columbia Civil Resolution Tribunal in Canada rejected Air Canada's defense that "the chatbot is a separate legal entity" (Moffatt v. Air Canada, 2024 BCCRT 149, 2024.2.14) | American Bar Association, 2024.2 |
Interpretation: These four cases all point in one direction. The defense that "the AI is an independent party responsible" is being rejected almost without exception, regardless of region or institution. What remains at issue is not the AI, but who within the organization is liable, and on what grounds.
In-depth analysis
1) Why the "organization that deployed the AI" is liable
The basic principle of agency law is that if one party authorizes another to act on its behalf, the party that granted the authority bears responsibility for the outcome, as long as the action falls within the scope of that authority. The Lyon Firm's analysis applies this directly to AI agents, explaining that from the moment an organization deploys an agent for a task and grants it authority, the organization is liable for "what happens within the scope of that authority." It identifies three types of especially high exposure: ▲negligent deployment, where an agent is put into high-risk tasks without sufficient vetting; ▲negligent oversight, where anomalous behavior goes unfiltered; and ▲breach of contract, where the agent enters into unapproved obligations with a third party.
2) Why an AI agent cannot be a party to a contract
For agency law to apply, the agent must have "legal capacity" — the ability to enter into a relationship by mutual agreement, form intent, and give and receive consent. An analysis by the UK law firm TS&P finds that AI systems do not meet this requirement. Its core argument is that "AI cannot form intent or give and receive consent, and cannot bear legal responsibility." As a result, an AI model is legally treated not as a person or a legal entity created by people, but more like software or intellectual property that a developer provides under a license or service. A contract is not made with the AI — it is made with the person or company that supplies or operates it.
3) What the Air Canada case shows about the failure of the "scope of authority" defense
Moffatt v. Air Canada (2024 BCCRT 149) is the precedent that shows how this plays out in an actual case. Air Canada argued that, regarding incorrect fare information its chatbot had provided, "the chatbot is a separate legal entity responsible for its own information." According to reporting by the American Bar Association (ABA), tribunal member Christopher Rivers rejected this, ruling that "while a chatbot has an interactive component, it is still just a part of Air Canada's website." Because this was a small-claims case, its weight as a higher-court precedent is limited, but it has become a reference point that many later analyses cite repeatedly, since the defense of "setting up the AI as a separate liable party" did not work in front of the tribunal.
4) The gray area that still remains — actions outside the scope of authority
An organization is not always fully liable. Under the agency law framework, when an agent clearly acts outside the scope of its granted authority, determining liability can become complicated. The problem is that an AI agent's "scope of authority" is often designed to be far vaguer than a human employee's job scope. Which APIs it can call is defined in code, but unless implicit expectations like "don't make a payment in this situation" are also spelled out in code, disputes over whether an action fell inside or outside the scope of authority only surface after an incident has already occurred. This is why documenting the scope of authority concretely and in advance becomes the practical standard that determines the boundary of an organization's liability when a dispute arises.
In practice — 4 principles to check before granting an agent authority to act
- Document the scope of authority in human-readable form. Record "which account can this agent use, what tasks can it perform, and up to what amount or scope" not only in code configuration but also in a separate document. This becomes the primary basis for judging whether an action was within the scope of authority when an incident occurs.
- Log and retain every action. A record of what was done, when, and under whose approval is necessary for the organization to prove in a dispute that it "was in control." Without logs, whoever bears the burden of proof is at a disadvantage.
- Require human approval, without exception, for irreversible actions. No matter how narrowly authority is scoped, actions that are hard to undo — mass emails, payments, signing contracts — should still go through a final confirmation step.
- Include a liability-allocation clause in vendor and supplier contracts. Unless the contract specifies in advance which failures the agent's vendor is responsible for versus the deploying organization, liability after an incident tends to fall entirely on the organization that deployed the agent.
References
Contents
Related posts
15. Why Granting AI Agents Broad Permissions Is Dangerous — Blast Radius
When an AI agent is given broad permissions ranging from sending emails to deleting files, a single small error can escalate into a major incident. This piece examines why, through the concepts of blast radius and least agency, and lays out design principles for adding confirmation steps around irreversible actions.
18. Where does the code you upload to a free AI tool go — read the terms
Free AI tool terms often allow training on your input, but paid isn't always safer. OpenAI, Anthropic, Google cases show training opt-out and confidentiality are separate, and terms change.
17. If you train an AI on your company's data, can that data be pulled back out — model inversion
Fine-tune an AI on company data and it stays retrievable by probing with varied queries. How model inversion works, and mitigations like data sanitization, dedup, and output filtering.
16. When a document AI reads hides an instruction — the risk of invisible commands
A document hiding instructions invisible to humans can make an AI agent leak files to an attacker or make a wrong call. Real incidents and checks to run before handing one over.