AI Liability and Insurance: Who Pays When It Goes Wrong
The vendor's contract caps their exposure at what you paid them. Your exposure is not capped at anything.

Erin Moore
Fractional Chief AI Officer
When an AI system you deployed causes harm to a customer, the customer's claim is against you. The vendor's liability to you is a separate matter, usually capped at some multiple of fees, and frequently far below the exposure. That asymmetry is the whole subject, and it is the part most procurement processes never surface.
This is not legal or insurance advice. Coverage terms vary enormously between insurers, policies and jurisdictions, and this area is actively changing. Use it to prepare questions for your broker and counsel, not as a description of your cover.
Where liability actually sits
Your customer contracted with you. The fact that you used a tool is not a defence they have to accept, any more than it would be for any other supplier failure. Your recourse against the vendor is governed by your agreement with them — and that agreement almost certainly contains a liability cap, an exclusion of indirect and consequential loss, and carve-outs around how the product was used.
Read those three together and the practical position becomes clear: the vendor's downside is bounded by your contract, and yours is bounded by the harm. That is not unreasonable of them; it is simply the position you are in, and it should inform how much control you insist on retaining.
The terms worth negotiating at purchase are in AI contract negotiation and question by question in the AI vendor security questionnaire. After deployment you are asking for a favour.
What to ask your broker
Insurance for this rarely sits in one place. The useful conversation covers several existing policies rather than searching for an "AI policy":
Does professional indemnity respond where the error originated in an automated system rather than a person? Does the cyber policy cover a data exposure caused by a model or a retrieval system, as distinct from an intrusion? Do media and advertising liability terms extend to content generated by a model and published by you? Where does technology errors and omissions sit relative to the other two? And are there AI-specific exclusions that have appeared at the last renewal — this is the question most likely to produce a surprise, because exclusions tend to arrive quietly.
Ask all of these in writing and keep the answers. A broker's written response is materially more useful eighteen months later than a recalled conversation.
What actually reduces exposure
Insurance is the last line. Four things do more.
Human checkpoints before irreversible actions. Most severe AI incidents involve an autonomous action nobody reviewed. This is the control with the best ratio of cost to protection, and it is the design point in AI agents for business leaders.
Scope discipline. Harm usually follows a system being used for something it was never validated for. Periodic re-attestation of actual use against approved use is unglamorous and effective.
The audit record. When something goes wrong, your position depends on being able to show what the system was given, what it produced, and who reviewed it. Without that record you cannot demonstrate reasonable care, and your insurer's first question will be exactly that. It is also the foundation of any workable incident response plan.
Not overstating capability. Claims you make about your AI to customers are performance claims, and in the US the FTC's advertising guidance applies to them as it would to any other. Overstating capability converts a product failure into a misrepresentation problem, which is worse in every respect.
Where the regulatory picture is heading
Obligations increasingly attach to the deployer, not only the developer — the EU AI Act's official text is the clearest current example of that structure. Practically, that reinforces the same conclusion: being able to evidence what you did, why, and with what controls is the durable protection. NIST's AI Risk Management Framework and the OECD's AI Principles are both reasonable standards to have documented yourself against, and naming an external framework is worth considerably more than a policy you wrote about yourselves.
Frequently asked questions
Who is liable when an AI system causes harm to a customer? You are, from your customer's perspective — they contracted with you. Your recourse against the vendor is separate and is usually capped by your agreement at a multiple of fees, which is often far below the actual exposure.
Does our existing insurance cover AI incidents? Possibly, across professional indemnity, cyber, media liability and technology errors and omissions — but terms vary widely and AI-specific exclusions are appearing. Ask your broker in writing and keep the answers.
What reduces AI liability exposure most? Human checkpoints before irreversible actions, scope discipline so systems are not used beyond what was validated, a complete audit record, and not overstating capability to customers.
Why does the audit record matter so much? Because demonstrating reasonable care depends on showing what the system was given, what it produced and who reviewed it. It cannot be reconstructed after the fact, and it is the first thing an insurer asks for.
Further reading
Tagged with:
Ready to Automate Your Business?
Let's discuss how AI automation can deliver measurable ROI for your organization in 90 days or sooner.