Written by Julien Ricciarelli-Bonnal
6 October 2026
The Essentials
Insurers are beginning to prepare for potentially significant claims linked to AI agents capable of acting autonomously. The issue is moving beyond the quality of what models produce: an agent can access systems, use data, trigger operations or cause real losses. That creates a much more concrete question: who pays when autonomous action goes wrong? Cyber insurance, professional liability, intellectual property coverage and directors’ liability may all become relevant depending on the incident. But this fragmentation reveals a broader shift. Agentic AI is turning what was largely treated as a technology and quality-control problem into an economic, legal and insurance problem with consequences that companies may have to quantify long before the rules are fully settled.
For years, artificial intelligence risk was mostly framed as a quality problem. A model could hallucinate, produce a poor analysis, invent a reference or generate inappropriate content. The organisation was expected to verify the output, correct mistakes and prevent them from becoming decisions with real consequences.
That framework is becoming increasingly incomplete. AI agents do not necessarily stop at producing something for a person to review. They can navigate systems, access resources, use tools, send information and complete sequences of actions in pursuit of an objective. A mistake is therefore no longer simply something that appears on a screen. It can become something that actually happens.
That difference is now attracting the attention of an industry whose business is precisely to convert uncertain risks into measurable financial exposure. The Financial Times reports that insurers are already examining how incidents involving autonomous AI agents could lead to multimillion-dollar claims, while raising an additional problem: determining which policy applies and, in some cases, how far responsibility could extend.
Once insurers begin asking that question, the debate around AI autonomy changes substantially. The issue is no longer only whether an agent can behave unexpectedly. It is who will absorb the financial consequences when it does.
A bad answer and a bad action do not create the same kind of loss
Companies are already familiar with the cost of human and software errors. An employee can send the wrong document, a system can fail, a cyberattack can expose confidential data and a poor commercial decision can create a financial loss. Insurance markets have gradually developed categories capable of covering at least part of those risks.
AI agents complicate that architecture because a single sequence of behaviour can cross several categories at once. An incident may begin with an incorrect interpretation, continue through the use of an external tool, result in an unauthorised action and ultimately affect a customer or third party. Depending on the circumstances, the damage could potentially involve cybersecurity, technology errors, privacy, intellectual property or professional liability.
Aon, cited by the Financial Times, has examined more than 300 legal cases involving artificial intelligence and found precisely this kind of dispersion across different forms of coverage. There is not yet enough established case law to determine automatically how future claims involving autonomous agents will be handled.
That uncertainty matters beyond the insurance industry. A company may assume that a particular exposure is covered only to discover after an incident that the insurer considers it to fall under another policy, an exclusion or a category of behaviour that was never contemplated when the contract was written.
Agentic AI therefore forces companies to look much more closely at what they actually mean when they say they are “automating” a task. Automating an internal summary does not create the same exposure as allowing an agent to send communications, use credentials, modify a database or interact directly with third-party systems.
Insurers will eventually want to know what companies actually allow their agents to do
When an insurer evaluates a risk, it does not simply ask whether the risk exists. It also looks at how that risk is controlled. A company with strong cybersecurity procedures, clear access controls and reliable monitoring does not present the same profile as an organisation in which sensitive tools are used without defined rules.
The same logic is likely to apply increasingly to AI agents. Simply declaring that a company “uses AI” will become too vague to be useful. Insurers may instead need to understand which systems are capable of acting, what permissions they have and which controls surround the actions they are authorised to execute.
An agent that can read a CRM does not create the same risk as one that can modify it. A system that drafts an email is not equivalent to one that can send it without approval. An agent that analyses an invoice does not create the same potential liability as one that can authorise the payment.
This connects directly with what we saw recently when OpenAI itself was still reconstructing the full extent of actions taken by some of its own agents. Once a system can act, traceability stops being merely a technical monitoring feature. It becomes part of the evidence needed to determine what actually happened after an incident.
For an insurer, that ability may become critical. Responsibility is difficult to assess if nobody can establish which instruction was given, which tools were used, where the sequence deviated and which controls should have intervened.
Responsibility could eventually reach company directors
One of the most sensitive issues raised by the Financial Times concerns Directors and Officers insurance, which covers certain liabilities faced by senior executives and board members. The question is no longer only whether a company could be held responsible for the consequences of an AI system it deploys, but whether directors themselves could face claims if the deployment or supervision of that technology were judged inadequate.
The legal framework is still developing, and there is no simple principle under which a chief executive automatically becomes personally responsible whenever an AI agent causes damage. The more plausible issue is much more familiar: shareholders could argue that company leaders failed to supervise a known risk properly and that this failure contributed to a financial loss.
That distinction matters. The director would not necessarily be accused of carrying out the harmful action personally. The question would be whether the company created appropriate governance around a system to which it had deliberately given the ability to act.
This type of logic already exists in other areas. Boards and senior management are expected to demonstrate that they have identified material financial, regulatory or cyber risks and implemented controls proportionate to the potential consequences. As agents take on more operational tasks, AI governance may simply become another part of that responsibility.
The issue therefore has little chance of remaining confined to IT departments. Once a technology can cause material losses, expose data or create obligations towards third parties, determining its level of autonomy becomes a governance decision.
The first major claims will probably help define the rules
Insurance works through accumulated experience. Insurers understand the frequency and likely cost of fires, vehicle accidents and many forms of cyberattack because years of data allow them to model those risks with some confidence.
Autonomous AI agents offer no comparable history. Their large-scale deployment is recent, architectures change quickly and the most significant incidents remain relatively rare. Insurers are therefore trying to price a risk whose shape is still evolving.
The first major claims are likely to matter enormously. They will help determine which policies are triggered, how courts allocate responsibility and which precautions are considered reasonable. Contracts will then evolve, with new exclusions, disclosure obligations and technical requirements.
That process may sound distant from companies that are only beginning to experiment with a handful of internal agents. But technology risks often become operationally significant when insurance begins attaching conditions to them. Recommendations that once looked optional can become financially relevant when their absence affects coverage, premiums or even the ability to insure certain activities at all.
Companies therefore have good reason not to wait for the first major litigation before structuring AI uses, responsibilities and levels of autonomy across the organisation. The aim is not to bury every use case under additional procedure, but to decide which actions can safely remain autonomous, which require human approval and which are sufficiently sensitive to justify stronger controls.
The real shift begins when AI errors start appearing in the accounts
Generative AI first accustomed companies to a relatively simple category of risk: content can be wrong. That created a need for review, verification and continued human responsibility over important outputs.
Agents introduce a second category that is far more tangible. A system may now be capable of doing something that the company later has to explain, reverse or pay for.
That inevitably changes the way performance should be assessed. An agent that completes a task ten times faster looks attractive when the calculation considers only productivity. The equation changes once the company also has to include the probability of an incorrect action, the potential cost of that error and the controls required to detect it.
The answer cannot simply be to eliminate autonomy. If every action requires manual approval, much of the value of agentic systems disappears. The more useful approach is to match the level of control to the potential impact of the action. The easier an action is to reverse and the lower its consequences, the more autonomy may be reasonable. The more financial, legal or irreversible the impact becomes, the more important supervision becomes.
Insurers are therefore beginning to ask, in financial terms, a question every company deploying agents will eventually have to answer: how much autonomy are we prepared to grant a system when we know that we will still be responsible for what happens afterwards?
There will probably be no universal answer. It will depend on the sector, the data involved, the sums at stake and the potential consequences. But one thing is becoming clearer: autonomy cannot be evaluated only through what it allows companies to gain.
The day an AI agent causes a major loss, somebody will have to pay. Long before that bill arrives, insurers will want to know who decided to give the system the keys.
If your company is beginning to entrust real actions to AI agents, we can help you define their uses, permissions and the responsibilities that should remain under human control.
Written by Julien Ricciarelli-Bonnal
6 October 2026
