Who is liable when an AI agent causes harm?

When an AI agent causes harm, legal responsibility almost always traces back to a person or organization — not to the AI itself, which has no legal personhood. As a default, liability flows to the deploying organization: under established agency, vicarious-liability, and negligence principles, the business that puts an agent into operation generally answers to the third party it harms, much as it would for an employee or a tool it chose to use. Responsibility can extend upstream to the developer or vendor through product-liability, professional-liability (E&O), or contractual-indemnity theories — particularly where the harm stems from a defect, a misrepresented capability, or the agent's autonomous decision-making rather than the deployer's own configuration. Outcomes vary by jurisdiction, the agent's degree of autonomy, and whether it faces customers, handles transactions, or runs internal workflows. Two practical wrinkles matter: emerging laws such as Colorado's AI Act (SB 26-189, obligations from January 1, 2027) impose deployer and developer duties — interaction notice, adverse-outcome disclosure, and meaningful human review — whose breach can support a claim; and AI-specific insurance exclusions such as Verisk's CG 40 47 can strip coverage a deployer assumed it had, so who ultimately pays may differ from who is liable. In practice, liability is shaped before any incident — by where human review sits, what the audit trail can prove, and how vendor contracts allocate risk.