Agent Liability

In 1979, an IBM training manual put the problem in a single line: “A computer can never be held accountable, therefore a computer must never make a management decision.” Agent liability is the law’s answer to that gap. When an AI agent causes harm, the software is not a legal person anyone can sue, so the real question is which of the humans and organizations behind it must answer.

Handing work to an AI agent is a lot like handing work to a human assistant: you are delegating authority, and the law has carried a toolkit for that problem for centuries. It is called agency law, the doctrine covering relationships where a principal gives an agent authority to act on their behalf. When something goes wrong, three questions separate the cases: were the instructions themselves bad, was the tool defective, or did the agent go off-script? Responsibility then lands in one of three seats: the developer that built the system, the organization that put it to work, or the user that directed it.

The software cannot stand in a courtroom. Someone in its chain can.

Civil liability is where most of the action is. California draws a bright line: under Civil Code section 1714.46, a defendant who developed, modified or used an AI system cannot escape by claiming the AI did it autonomously. The law does not make anyone automatically liable; it removes one excuse. The European Union goes further with strict product liability: defective software, expressly including AI systems, triggers no-fault liability for its maker, and courts may presume a defect when technical complexity makes it excessively hard for a claimant to prove. Colorado takes a third route, imposing reasonable care duties on both developers and deployers of high-risk AI systems.

Criminal exposure is also on the table. A pending bill in the US Senate, the AI Agent Accountability Act, would extend Computer Fraud and Abuse Act liability, criminal and civil, to operators who knowingly run an agent that recklessly causes hacking damage, and to developers who skip reasonable safeguards against hacking despite knowing what their agent can do. Another proposal, the AI AGENT Act of 2026, would hold user-authorized agents to a duty that cannot be waived by any contract: the care, skill and diligence of an ordinarily prudent person. Both are introduced, not enacted.

Three lookalikes cause most of the confusion. Agent compliance is about following the rules up front; liability is about who answers when harm happens anyway, and being compliant helps without immunizing anyone. Fiduciary duty for AI agents is one specific, elevated standard of care, for agents trusted with a user’s interests. It is a route into liability, not the whole question. Agent governance is the oversight machinery an organization builds to keep its agents in line; good governance shrinks exposure, but the legal allocation of responsibility is a separate matter. One naming note: the phrase sits on older vocabulary from principal-agent law, where the live question is whether a principal answers for what its agent did in the course of its duties.

Take a concrete case. A company gives an AI purchasing agent a budget and an approved vendor list. A fraudulent invoice arrives, and the agent pays it. Then come the old questions, the same ones courts ask of any delegation: who had authority, who controlled the agent’s behavior, and who benefited? The company can face exposure for negligent oversight, the developer can face exposure if a documented safeguard failed, and the fraudster answers for the fraud. In California at least, the company cannot wave the problem away by saying the agent acted on its own. In the end, responsibility follows control and benefit.