↑ Resources / Essay · Series
Part II of III The Confidence Infrastructure for Autonomous AI

Who Is Liable When an AI Agent Fails, and What It Takes to Show What Happened

Essay · Regulatory · 8 min read ·

Part I looked at how earlier technologies reached scale: not through engineering alone, but through the standards, inspection and insurance that grew up around them. This essay looks at the legal framework now forming around AI agents in the EU, and at the practical gap in the middle of it.

Three pieces of EU law matter most for organizations that run AI agents. They do different jobs and apply to different situations. Read together, they raise the same practical question.

The EU AI Act sets obligations for high-risk AI systems, which for the areas listed in Annex III apply from 2 December 2027. High-risk systems must record events automatically over their lifetime (Article 12) and be designed so people can oversee them effectively (Article 14). Deployers must keep the automatically generated logs for at least six months (Article 26). The Act says relatively little about how those logs should be structured or what makes them trustworthy.

The revised Product Liability Directive (EU) 2024/2853, which member states must transpose by 9 December 2026, treats software, including AI systems, as a product. It covers harm to natural persons, such as personal injury, damage to private property and loss of non-professional data. It lets courts order a defendant to disclose relevant evidence, and it allows courts to presume defect or causation in defined situations, for example when a defendant fails to disclose evidence or when technical complexity makes proof excessively difficult for the claimant. Where several operators are liable for the same damage, they can be held jointly and severally liable.

DORA has applied to financial entities since 17 January 2025. It requires them to manage the risk of third-party ICT providers (Article 28), which can include model providers, and to secure access, inspection and audit rights in those contracts (Article 30). It also sets tight deadlines for reporting major ICT incidents: an initial notification within 4 hours of classification and no later than 24 hours after detection, an intermediate report within 72 hours and a final report within a month.

None of these laws mentions AI agents by name in the way this essay uses the term. Together, though, they make one thing clear: organizations that deploy agents will be expected to answer for what those agents do, and in some situations to show it. What the law cannot supply is the record that makes that possible.

Why proof is harder with agents

When a steam boiler exploded in the 1860s, working out what happened was difficult but possible. The physical evidence survived. The metal could be examined, inspection records could be checked and witnesses could describe the sequence of events.

When an AI agent causes harm, the evidence is more fragile. Say it approves a fraudulent payment, relies on an outdated policy document, or is manipulated through prompt injection into leaking data. The agent's behavior comes from the interaction of a model, its context, retrieved documents, a sequence of tool calls and an input that may have been crafted by an attacker. The model may have been updated since. The context may never have been logged. Each downstream system may hold a record of its own part of the event, but none holds the whole sequence.

In Moffatt v. Air Canada, decided in 2024, British Columbia's Civil Resolution Tribunal held the airline responsible for incorrect information its website chatbot gave a customer, and rejected the argument that the chatbot was responsible for its own actions. It is a small-claims decision in Canada, not EU law, but its logic is widely cited: the organization that deploys an agent answers for it. The case did not have to consider what happens when the organization cannot reconstruct what its agent did.

In the EU, the Product Liability Directive makes that question more pressing. Where its conditions are met, a defendant who cannot produce relevant evidence may face a presumption of defect or causation. Missing records are not a neutral fact.

Accountability spread across many parties

In March 2026, the UK's Digital Regulation Cooperation Forum, which brings together the CMA, FCA, ICO and Ofcom, published a foresight paper on agentic AI. It is explicitly not policy. Among other points, it notes that responsibility for an agent's behavior can be split across model providers, system providers and the organizations that deploy agents.

That matches what happens in practice. The deploying organization may have configured the agent correctly, while the model provider changed the model, a tool provider changed an API, or a data source was compromised. Each party contributed, and none can show the whole picture on its own.

In our view this is less a question of governance design than of evidence. The record that would let responsibility be attributed, and that a defendant could rely on under the Product Liability Directive, is not yet a standard part of most agent deployments.

What certification can and can't show

Certification is starting to fill part of the gap. AIUC-1 is a certification standard for AI agents, often described as a SOC 2 for agents. Its requirements cover six domains: data and privacy, security, safety, reliability, accountability and society. The standard is updated quarterly, and its Q2 2026 update added controls for MCP security and agent identity. ElevenLabs, among others, has been certified and has paired its certification with AI insurance.

Certification answers a useful question: did this system meet the standard when it was tested? That gives procurement teams a concrete signal, gives insurers a starting point for underwriting and gives regulators documented evidence of diligence.

Its limit is time. A SOC 2 report describes controls at the time of the audit, which works well for software that changes on planned release cycles. AI agents change more continuously. The model is updated, behavior drifts and new attack techniques appear. An agent certified in February can carry a different risk in June without any single failure.

AIUC-1 responds with quarterly testing, which helps. But a quarterly test still shows what the system looked like on the test date, not what it was doing on the day of an incident.

What a useful record needs

No law prescribes exactly how an agent's behavior must be recorded. In our view, a record that can support incident reporting, a liability dispute or meaningful human oversight needs four properties.

It is complete. It covers the agent's actions, model calls, tool calls and data access, not a sample. Reconstructing a sequence from separate vendor consoles is slow and leaves gaps.

It is continuous. It exists from the day the agent goes live. A record assembled after an incident cannot show that controls were in place before it.

It is independent. It does not depend on one vendor. A log that exists only in a model provider's console covers only that provider's part of the work, and in a dispute involving the provider an independent record carries more weight.

It shows later changes. Ordinary logs on a system the operator controls can be edited without a trace. Chaining each entry to the previous one, and signing entries when they are written, makes later changes visible.

Certification bodies describe what good behavior looks like, and insurers price the risk that remains. Both work better with a record like this underneath them, and most agent deployments do not have one yet.

What comes next

Part I suggests how this tends to resolve. The technology exists, the legal frameworks are written and the consequences of weak records are becoming clearer. What remains is for the infrastructure to catch up.

Hartford Steam Boiler did not wait for legislation to require boiler inspections. It built inspection capacity because it paid when boilers failed. Insurers now writing cover for AI agents face a similar incentive, and we expect them to ask for better behavioral records as claims become more complex.

A continuous, independent and tamper-evident record of agent behavior is what we think that market will come to expect, less as a compliance item than as the basis on which liability and insurance can work as intended.

KYDE Gateway

Kyde scores how AI agents behave. Kyde Control writes a continuous, independent, hash-chained record of every agent action routed through it, across every provider it fronts. Signing each entry with your own Ed25519 keys is an Enterprise capability. That record can support your evidence under the Product Liability Directive and your DORA incident reporting. It does not discharge either obligation on its own.