Who Pays When Your AI Agent Fails? The 2026 Liability Gap

The accountability gap, in one court case

Start with the case every legal team now cites. In Moffatt v. Air Canada, decided by British Columbia's Civil Resolution Tribunal in February 2024, an airline's website chatbot told a grieving customer he could claim a bereavement discount retroactively. That was wrong — it contradicted the airline's own published policy — and when the customer asked for the refund the chatbot had promised, Air Canada refused. The airline's defence was remarkable enough that the tribunal used exactly that word: it argued the chatbot was "a separate legal entity" responsible for its own statements. The tribunal rejected the idea outright, ruled that Air Canada was responsible for all the information on its site whether it came from a static page or a bot, and ordered the airline to pay the difference.

The dollar figure was trivial — a few hundred Canadian dollars. The principle was not. A company tried to argue that its automated tool was an independent actor it could not be held responsible for, and it lost. Now scale that principle from a bereavement fare to an agent that autonomously issues a five-figure credit, misquotes a regulated product, or commits your business to terms in an email thread, and you can see why "the AI did it" has become the least effective sentence in corporate legal.

This is the heart of what practitioners now call the accountability gap. Traditional software liability assumes a fault you can point to: a bug, a misconfiguration, a line of code that did the wrong thing. Agentic systems are non-deterministic by design — the whole reason you deploy them is that they choose their own path at runtime. So an agent can cause real harm through an action nobody explicitly wrote, which leaves a gap between unpredictable behaviour and the legal system's need to hold a person or company responsible. The law of 2026 is filling that gap in one consistent direction: it holds the organisation that deployed the agent accountable.

What the law now says about deployers

Two developments in particular have hardened the position over the past year, and both point at the deployer rather than the model maker.

In Europe, the EU AI Act's obligations for high-risk systems are in full effect in 2026, and they land squarely on the deployer — the business putting the system to use — not only on the provider that built the model. If your agent operates in a high-risk domain, you are expected to run a risk-management process, keep a human in meaningful oversight, log the system's activity, and maintain documentation you can show a regulator. The penalties are not symbolic: non-compliance can reach 35 million euros or 7 percent of global annual turnover, whichever is higher. We covered the operational detail of this in our guide to the EU AI Act for business automation, but the liability takeaway is simple — running the agent is now a regulated act with its own duties.

In the United States, California's AB 316 took effect on 1 January 2026 and did something narrow but pointed: it bars a defendant from arguing that an AI system "autonomously" caused the harm as a way of escaping liability. In other words, the exact defence Air Canada tried is now statutorily unavailable in California. Alongside these statutes, courts and regulators are converging on what is being described as a "reasonable oversight" standard: the deploying organisation is liable for an agent's actions unless it can demonstrate it had genuine monitoring, auditing and safety controls in place. Notice what that does — it turns your guardrails from a nice-to-have into the evidence that decides whether you are on the hook.

Why your vendor contract will not save you

The instinct, once a leader understands the exposure, is to push it onto the vendor: "surely the platform that built the agent is responsible for what it does." It is worth being precise here, because the answer has two parts that get confused.

First, to the outside world — the customer you overcharged, the regulator investigating a mis-sold product — your contract with the vendor is irrelevant. Liability to a third party flows from the relationship you have with that third party, and no clause you sign with a software company can sever it. The affected customer sues you; the regulator fines you. Second, a vendor contract can still allocate cost between you and the vendor after the fact, through indemnities and warranties. That matters, but it is a private settlement mechanism, not a shield — and only to the extent the vendor actually agreed to cover the loss.

This is where most teams get an unpleasant surprise. Read the indemnity in a typical automation or AI-platform agreement and you will usually find liability capped at something like the last twelve months of fees you paid, with consequential and indirect losses excluded entirely. An agent that runs thousands of decisions a month on a modest subscription can generate a loss many multiples larger than that cap in a single bad afternoon. The vendor's maximum exposure is your annual bill; your exposure is whatever the agent did. Those two numbers are rarely in the same universe.

The insurance market is scrambling to catch up

If the contract does not cover the gap, the natural next question is whether insurance does. For now, the honest answer is "check carefully, because the market is mid-scramble." The problem for underwriters is that autonomous agents have almost no claims history. Pricing risk without data is guesswork, so insurers are doing three things at once: extending some existing cyber and technology errors-and-omissions policies to touch AI, testing how far that wording really stretches, and in some cases explicitly excluding agentic-AI failures until they understand them. The dangerous middle ground is assuming your current cyber or professional-indemnity policy quietly covers an autonomous agent when it may say nothing about it, or may already carve it out.

At the same time, a dedicated AI-liability market is forming fast enough to be worth watching. A few of the moves already visible in 2026:

• Munich Re's HSB (Hartford Steam Boiler) offers standalone AI errors-and-omissions protection aimed at smaller businesses, treating AI mistakes as their own insurable category rather than an afterthought inside a cyber policy.
• Armilla writes AI performance cover that responds to model errors and the downstream liability they trigger, effectively insuring the gap between what the model promised and what it did.
• Counterpart embeds affirmative AI cover inside its management-liability and E&O products, and — crucially — treats autonomous agent outputs as insured events rather than ambiguous grey area.
• Specialist and broker-led carriers such as CFC are publishing their own frameworks for how autonomous AI changes underwriting, a sign the topic has moved from novelty to line-of-business.

The practical implication is not "go buy AI insurance tomorrow." It is that "are we covered if the agent causes a loss?" is now a specific question with a specific answer, and the answer lives in your policy wording, not in a general assumption. If you are deploying anything high-stakes, put that question to your broker in writing and ask them to point to the exact clause that responds to an autonomous-action loss.

Where the liability actually concentrates

Not all automation carries the same risk, and treating every workflow as a lawsuit waiting to happen is as unhelpful as ignoring the problem. Exposure concentrates wherever an agent takes an irreversible action on the outside world without a human checkpoint. Map your automations against that idea and the picture usually clarifies quickly.

The two high-risk rows are where the money and the headlines are. An agent that files an internal summary for a manager to read is a productivity tool; an agent that emails a customer a binding quote is a legal actor. The design goal is to move as much of your automation into the bottom row as the business case allows, and to wrap anything that has to live in the top rows in controls strong enough to survive a regulator's questions. Getting this mapping right is also the foundation of an honest AI agent total cost of ownership calculation — liability is a real line item, even when nobody puts a number on it until something breaks.

A pre-deployment liability checklist

You do not need a legal department to close most of the gap. You need to treat the agent's authority as something you grant deliberately rather than by default. Before any agent that can act on the outside world goes live, walk through this sequence:

1. Draw the blast radius. Write down the worst single action this agent could take and its plausible cost. If you cannot describe it, you are not ready to deploy it.
2. Gate the irreversible steps. Require a logged human approval before money moves, contracts change, prices update or customer-facing promises go out. This is the control that regulators and insurers most want to see.
3. Constrain the tools and permissions. Give the agent the minimum set of actions and data access it needs, and nothing that would let it wander into a high-risk step you never intended.
4. Log every decision. Capture the input, the reasoning trace, the tools called and the output. Under a reasonable-oversight standard, this log is the difference between "we controlled it" and "it was out of our hands."
5. Read the vendor cap. Locate the liability limit and exclusions in your platform contract and compare them to the blast radius from step one.
6. Confirm the insurance wording. Ask your broker, in writing, whether an autonomous-action loss is covered and under which clause. Treat silence as a "no" until proven otherwise.
7. Assign a human owner. Every production agent needs a named person accountable for its behaviour, with the authority to pause it. An agent nobody owns is an incident nobody is watching.

Most of these steps take an afternoon, not a quarter. What they buy you is the ability to answer the only question that matters after something goes wrong — "what controls did you have in place?" — with evidence instead of a shrug. That evidence is what the emerging reasonable-oversight standard rewards, and it is exactly the discipline we recommend when you test an AI agent before you buy it.

Why this is a strategy issue, not a compliance footnote

It would be easy to file all of this under legal housekeeping and move on. That would be a mistake, because the liability picture is quietly reshaping which automations are worth building at all. The market data explains why. MIT's 2025 State of AI in Business study found that around 95 percent of enterprise generative-AI pilots showed no measurable return, and Gartner has estimated that 89 percent of AI-agent pilots never reach production, with a substantial share of agentic projects on track to be cancelled by 2027 over runaway costs, unclear value and — the word that keeps recurring — governance failures. Liability is one of the forces pushing projects over that edge.

Read the other way, the discipline is an advantage. The businesses that keep deploying agents through the shakeout are the ones that scoped the authority, gated the irreversible actions, logged the decisions and knew exactly where their coverage began and ended. They move faster precisely because they are not afraid of their own automation. The accountability gap is real, but it is closeable — and closing it deliberately is what separates an agent you can trust with your customers from one you cannot.

The one-line version to take into your next automation review: an autonomous agent is a legal actor you are vouching for. Decide what you are willing to vouch for before you switch it on, not after the invoice arrives.

FAQ

Who is legally liable when an AI agent makes a mistake?

In almost every case, the business that deployed the agent is liable to the affected customer or regulator, not the AI vendor. Courts and tribunals have consistently ruled that a company owns the actions of its automated tools, including promises the tool makes and commitments it executes.

Can I pass liability to the AI vendor through my contract?

Not to the outside world. A vendor contract can allocate cost between you and the vendor after the fact through indemnities, but it does not remove your direct legal responsibility to the customer or the regulator. Most vendor agreements also cap that indemnity well below the loss an autonomous workflow can cause.

Does the EU AI Act change who is responsible?

Yes. In 2026 the EU AI Act places clear obligations on deployers of high-risk systems — risk management, human oversight, logging and documentation — with fines up to 35 million euros or 7 percent of global annual turnover. It reinforces that the organisation running the agent, not just the model maker, carries the compliance burden.

What is the accountability gap?

Traditional software liability rests on identifiable bugs or faults. Agentic systems are non-deterministic, so an agent can cause harm through an action nobody explicitly programmed. The accountability gap is the space between that unpredictable behaviour and the legal need to hold someone responsible — and the law is filling it by holding the deployer accountable.

Does my existing cyber or E&O insurance cover AI agent failures?

Do not assume so. Underwriters lack the claims history to price autonomous-agent risk with confidence, so cyber and technology errors-and-omissions policies are being tested, extended, and in some cases explicitly excluding agentic AI failures. Review your wording before you deploy a high-stakes agent, and ask specifically about autonomous-action losses.

Is there dedicated AI liability insurance yet?

Yes, a young market is forming. Munich Re's HSB offers standalone AI errors-and-omissions cover for smaller firms, Armilla writes AI performance cover for model errors and downstream liability, and Counterpart embeds affirmative AI cover in its management-liability and E&O products, treating autonomous agent outputs as insured events.

Which automations create the most liability?

Any workflow where the agent acts on the outside world without a human check: making promises to customers, approving refunds or payments, signing or altering contracts, changing prices, and anything touching regulated activity such as lending, insurance, healthcare or hiring. Internal, read-only or draft-only agents carry far less exposure.

What is the single most protective control I can add?

A logged human approval gate on every irreversible or customer-facing action. It converts the agent from an autonomous decision-maker into a fast assistant, and the audit trail is exactly the evidence of reasonable oversight that regulators and insurers now expect.