Who is liable when an AI agent causes harm?

Last verified: June 1, 2026

Answer

The organization that deploys an AI agent is usually first in line when it causes harm — AI has no legal personhood, so liability falls on people and organizations, not the software. A vendor or developer may share responsibility where a defect, misrepresentation, contract term, or degree of control points upstream. Insurance, including AI-specific exclusions, affects who ultimately pays, not who is legally liable. This is general analysis, not legal advice.

Liability for an AI agent's actions tends to resolve in layers. Default — deployer or operator: the business that puts the agent into operation is generally answerable for the harm it causes, much as it would be for an employee or a tool it chose to use, under established agency, vicarious-liability, and negligence principles. Vendor or developer: responsibility can extend upstream through product-liability, professional-liability (E&O), or misrepresentation theories where the harm traces to a defect or an overstated capability rather than the deployer's own setup. Contract and indemnity: master service agreements, warranties, limitation-of-liability clauses, and indemnities reallocate that risk between the parties and often decide who actually bears a loss. Insurance and exclusions: a policy may respond, but AI-specific exclusions such as Verisk's CG 40 47 can strip coverage a deployer assumed it had — changing who pays without changing who is legally liable. Human review and audit trail: where a person reviews the agent's decisions and every action is logged, that record shapes whether the deployer is found negligent and whether coverage responds. Outcomes vary by jurisdiction and the agent's degree of autonomy, and newer rules such as Colorado's AI Act (SB 26-189, deployer and developer duties effective January 1, 2027) can add obligations whose breach supports a claim. This is general business and insurance-risk analysis, not legal advice.

Scope

General business and insurance-risk analysis, not legal advice. Liability outcomes change with jurisdiction, the AI tool's role and degree of autonomy, whether a human reviews its decisions, how vendor and customer contracts allocate risk, and the exact wording of any policy exclusion or endorsement. Confirm specifics with qualified counsel and your broker.

Operational implication

Liability is shaped before any incident, by operating controls a business can put in place now. Gridex deploys AI inside governed workflows: human review points where a person stays in the decision on consequential actions; explicit authority boundaries that cap what the agent may do on its own; a complete audit trail that logs each action and the human sign-off behind it; vendor contracts and indemnities reviewed so responsibility is allocated, not assumed; and a coverage check against AI-specific exclusions such as Verisk's CG 40 47 so the policy actually responds. These are the controls that determine who is found responsible and whether insurance pays.

Applicable Regulations

SB-26-189

Colorado AI Act — Automated Decision-Making Technology (SB 26-189, repeal & reenactment of SB 24-205)

enacted

On 2026-05-14 Governor Polis signed SB 26-189, which repeals and reenacts the Colorado AI Act (originally SB 24-205). The new law abandons the risk-management / annual-impact-assessment model and replaces it with a disclosure-and-notice framework governing "automated decision-making technology" (ADMT) that makes or substantially influences "consequential decisions" (education, employment, housing, financial services, insurance, healthcare, government services). The statute formally takes effect 2026-08-12 (no safety clause), but all substantive compliance obligations — for both deployers and developers — begin 2027-01-01, which is the operative date for regulated businesses; the Attorney General's implementing rules are also due by 2027-01-01. The AG has stated he will not enforce until the mandatory rulemaking process concludes.

Key Requirements

Interaction Notice Deployers must give clear notice at the point of interaction when a consumer interacts with an automated decision-making technology (ADMT)
Adverse-Outcome Disclosure Provide a plain-language explanation within 30 days of an adverse consequential decision made or substantially influenced by an ADMT
Data Correction Right Allow consumers to request correction of factually incorrect personal data used by the ADMT
Meaningful Human Review Provide meaningful human review and reconsideration after an adverse consequential decision
Developer Documentation Developers must supply technical documentation (intended uses, known harmful uses, training-data categories, known limitations and risks, and instructions enabling meaningful human review), notify deployers of material updates, and retain compliance records for 3+ years. Like all duties under the act, these obligations begin 2027-01-01
Effective: 2027-01-01 Penalties: Enforced exclusively by the Colorado Attorney General; violations are treated as deceptive trade practices under the Colorado Consumer Protection Act. Before enforcement the AG must give 60 days' written notice and an opportunity to cure; this cure right sunsets 2030-01-01, after which enforcement may be immediate. The AG has stated no enforcement will occur until the mandatory rulemaking process concludes.

Carrier Endorsement Details

CG-40-47

Verisk — CG 40 47 01 26

Excludes bodily injury, property damage, and personal/advertising injury arising out of generative AI under Coverage A and Coverage B. Part of the January 2026 ISO edition; companion forms address narrower scopes: CG 40 48 (Coverage B / personal and advertising injury only) and CG 35 08 (products and completed operations).

Key Provisions

Excludes BI and PD (Coverage A) and personal/advertising injury (Coverage B) arising from generative AI
Companion form CG 40 48 limits the exclusion to Coverage B only
Companion form CG 35 08 applies the exclusion to products/completed operations
Applies regardless of whether AI is owned, licensed, or embedded
Type: exclusion Policies: CGL

Where this lands operationally

Gridex turns the compliance or coverage question into operated workflow controls: intake, review points, audit trails, and the places a person stays in the decision.

Build Your AI Governance Framework

Map the specific workflow where the agent acts, decide where human review sits, and stand up the audit trail and vendor-contract checkpoints that determine liability — start with a Gridex AI governance review.

Build Your AI Governance Framework

Related Questions

  • What should an AI governance framework include? An AI governance framework should include an AI use policy, an inventory of where AI makes or substantially influences consequential decisions, documentation requirements, incident response procedures, and regular audit mechanisms. Note that Colorado's AI Act (SB 26-189, which repealed and reenacted SB 24-205) dropped the old impact-assessment and high-risk-classification model in favor of disclosure, consumer-notice, and human-review duties — so a framework should map to those obligations rather than the repealed assessment regime.
  • What is Verisk CG 40 47? Verisk CG 40 47 is a CGL policy endorsement that excludes coverage for bodily injury, property damage, or personal/advertising injury arising out of AI systems.
  • Do AI exclusions cover shadow AI? Yes. AI exclusion endorsements like Verisk CG 40 47 and Berkley PC 51380 use broad language covering any AI use, including unsanctioned shadow AI tools used by employees without authorization.
  • What liability exposure do HR tech vendors have for AI tools? HR tech vendors face dual exposure: as AI developers under Colorado's developer obligations, and through client contracts when their AI tools contribute to employment discrimination claims covered by state hiring laws.