Who's Liable When Your AI Agent Makes a Bad Call?

I have noticed that the moment you put an AI agent in front of a customer, everyone gets very interested in the upside and very quiet about one question: when it gives bad advice or approves the wrong thing, who eats it? AI liability is the part of the business case that gets waved off with "we'll put it in the terms of service," and that hand-wave is exactly where the risk lives. So let me walk through where the fallout actually lands, what you can and cannot push onto someone else, and how to build the kind of record that keeps a bad call from becoming a bad quarter.

Where AI Liability Actually Lands

Liability lands on the deployer first. If your agent tells a customer something wrong, approves a refund it shouldn't, or quotes a policy that does not exist, the person on the other end has a relationship with your business, not with your vendor or the model weights. That relationship is where the duty lives, and duty is what liability attaches to.

This is not a fringe legal theory. As the Cloud Security Alliance put it, "Deploying an AI agent does not transfer liability to the agent; it concentrates accountability on the deploying organization." The agent is not a scapegoat you can point at. It is a thing you chose, wired into your operations, and pointed at your customers, and that chain of decisions is yours.

It also is not waiting on some exotic new statute. The RAND Corporation notes that "absent new law from Congress or state legislatures, tort law will be applied to harms caused by artificial intelligence." Translation: the same negligence and product-liability rules that already govern a faulty product or a bad piece of advice will be reached for, and they were written long before anyone shipped an agent. The novelty is in your deployment, not in the law.

What Your Contracts Can and Can't Shift

Contracts shift money, not duty. A well-drafted vendor agreement can decide who pays when a model misbehaves, and you absolutely want indemnification, a limitation-of-liability clause, and clear warranties from whoever built the thing. What that paperwork cannot do is dissolve the obligation you owe to the customer standing in front of you. You can be made whole by your vendor later and still be the one who gets sued today.

Disclaimers get oversold in exactly the same way. Burying "outputs may be inaccurate" in a terms-of-service page feels like armor, and it is closer to a raincoat. Courts still ask whether you behaved reasonably, and a disclaimer does not erase liability for negligence, a defective product, or a misleading representation your system made to someone who trusted it. If this sounds like the same discipline as a real risk assessment, that is because it is; I wrote about doing that honestly in an AI risk assessment that isn't just theater.

The stakes here are not academic, especially once you cross a border. Under the EU AI Act, fines for prohibited practices reach up to €35 million or 7% of global annual turnover, whichever is higher, according to the European Commission. No indemnification clause with a small vendor is going to absorb a number like that. Contracts are for allocating the losses you can see coming; they are not a force field.

The Accountability You Can't Outsource

The cleanest illustration of the accountability you keep is a small case about a big airline. In Moffatt v. Air Canada (2024 BCCRT 149, decided February 14, 2024), a grieving customer followed refund advice invented by the airline's chatbot. Air Canada argued, more or less, that the chatbot was a "separate legal entity" responsible for its own words. The British Columbia Civil Resolution Tribunal was not impressed, treated the bot as part of Air Canada's own website, and held the airline liable for negligent misrepresentation, per the American Bar Association.

The damages were tiny, about CAD $812 all in, which is precisely why the case matters. Nobody is scared of the money. What should get your attention is the principle: the "our AI did it" defense was tested in a real tribunal and lost on day one. Your customer does not care that a language model generated the sentence. They care that your logo was on the page when they read it.

This is the emotional core of the whole topic, and it is worth saying plainly. You cannot outsource accountability to a thing that cannot be accountable. As legal scholar Nathalie Smuha observes, cited by the NTIA, "AI systems do not have legal personality and cannot bear legal responsibility." Responsibility does not vanish into the model. It rolls back uphill to the humans who deployed it, which means it rolls back to you.

Human Oversight Is Your Legal Cover

Meaningful human oversight is the best cover you can build, because it goes directly to whether you acted reasonably. When a person reviews the high-stakes calls, sets the boundaries of what the agent may decide on its own, and can step in before harm reaches a customer, you have a real answer to the standard-of-care question a court will ask. The oversight is not decoration; it is the argument.

The trouble is that most companies claim oversight they cannot prove. 65% of organizations now report regularly using generative AI in at least one business function, roughly double the 33% who said so a year earlier, according to McKinsey. Deployment is racing ahead of control. The same research found 56% of organizations have a formal generative-AI policy while only 32% actually track or audit how those systems get used. A policy no one monitors is not oversight, it is a press release, and it is thin evidence on a bad day.

Oversight also is not a synonym for "a human clicks approve on everything," which just recreates the bottleneck you bought the agent to remove. The move is to scope authority: let the agent decide the low-stakes, reversible, high-frequency stuff, and route the consequential and irreversible calls to a person. I dug into how to draw that line in giving an agent a budget. Oversight that survives contact with a courtroom is oversight you designed on purpose.

Building a Defensible Record for AI Liability

If AI liability is the disease, a defensible record is the treatment, and it is boring on purpose. When something goes wrong, and 46% of organizations already report at least one negative consequence from generative AI, per McKinsey, the difference between a manageable incident and a losing case is whether you can reconstruct what the system did and what oversight existed at the time. You cannot narrate that from memory. You reconstruct it from records.

The record worth keeping is unglamorous: audit logs of prompts and outputs, model and version history so you can say which system produced which answer, human-review notes, the vendor terms in force, your risk assessments, and an incident report when things break. This is the same instinct behind watching your systems closely in the first place, which I made the case for in agent observability, and behind having a plan for the moment it fails, which I covered in what to do when an agent gets it wrong.

None of this is the fun part of shipping an agent, and that is the point. The company that keeps the log, scopes the authority, and can show its work is not trying to prove its AI never errs. It is trying to prove it behaved like a responsible adult when the AI did, and that, far more than any disclaimer, is what actually holds up.

Frequently Asked Questions

Who Is Liable When an AI Agent Makes a Bad Decision?

The deploying business is usually the first place courts and regulators look, because it owed the duty to the customer who was harmed. Liability is then allocated among vendor, integrator, and developer through contracts and negligence or product-liability law, but the deployer rarely escapes it entirely.

Can a Company Disclaim AI Liability in Its Terms of Use?

A disclaimer helps manage risk, but it usually does not eliminate liability for negligence, a defective product, or a misleading representation your system made. Courts still ask whether you acted reasonably, and a buried disclaimer is weak evidence that you did.

Is the AI Vendor or the Business User More Likely to Be Sued?

Both can be named, but the deploying business often faces the most immediate exposure because it chose the tool, integrated it, and owed duties to the customers or counterparties affected. Vendor indemnification may reimburse you later, but it does not stop you from being the defendant now.

What Records Should a Business Keep to Defend an AI Claim?

Keep audit logs of prompts and outputs, model and version history, human-review records, vendor terms, risk assessments, and incident reports. The goal is to be able to reconstruct exactly what the system did and what oversight existed when it did it.

Does the EU AI Act Change Who Is Responsible for AI Harm?

It does not make the AI itself responsible. It increases obligations and penalties for providers, deployers, importers, and distributors, especially for high-risk and prohibited systems, with fines reaching up to €35 million or 7% of global annual turnover.

References

• Cloud Security Alliance: AI Liability, Enterprise Accountability in the Age of AI Agents
• RAND Corporation: Liability for Harms from AI Systems
• American Bar Association: BC Tribunal Confirms Companies Remain Liable for Information Provided by an AI Chatbot
• European Commission: Liability Rules for Artificial Intelligence
• NTIA: Liability Rules and Standards
• McKinsey: The State of AI in 2024