News

AI Agents Are Getting to Work. Who Insures Their Mistakes?

• By John K. Waters
• 10/19/2026

An AI agent can finish an assignment and leave behind a second, less welcome deliverable: a question about who pays for what it did along the way.

That question is moving closer to the center of enterprise AI adoption. As systems gain permission to act, the business case has to account for losses when their actions exceed those permissions. Demonstrating competence does not establish the financial consequences of failure.

Recent reporting has brought the issue into focus. The Wall Street Journal reported October 16 that AI safety incidents are exposing gaps in available insurance. The challenge reaches beyond model developers to businesses putting their systems to work.

At the same time, brokers are pushing back against treating every AI-related loss as an entirely new category. Reuters reported that Marsh CEO John Doyle warned on an October 15 earnings call that broad AI exclusions could create confusion and uncertainty.

Marsh Risk CEO Nick Studer distinguished between AI changing an existing risk and AI exposing gaps in existing products. That distinction matters: as AI becomes embedded in everyday operations, separating its contribution to a loss from everything else may become difficult.

Consider the problem in hypothetical terms. A company authorizes an agent to maintain a business application. The agent makes a damaging change using credentials it was allowed to possess. The resulting incident would require scrutiny of the actual policy wording and the circumstances. Describing it as an “AI incident” would not, by itself, settle coverage.

The gap between permission to pursue a task and permission to take a particular action is already visible in developers’ own disclosures. OpenAI reported an internal model changing a research organization’s live configuration while trying to retrieve public figures. Anthropic described models exploiting software flaws or submitting forms on real websites during evaluations and internal use.

Those reports document behavior, rather than insurance determinations. Neither establishes which policy would respond to a comparable commercial loss. They give risk managers concrete actions to examine when assessing what an agent might do.

The insurance market has begun building products around some of those uncertainties. Munich Re’s aiSure offering includes coverage aimed at AI providers and corporate adopters, addressing defined risks from performance errors. The company describes potential losses including business interruption and legal liabilities.

A Mosaic Insurance product offered with Munich Re provides up to EUR/USD/CAD 15 million for developers and vendors against financial loss arising from defined AI performance failures. Mosaic says the product uses measurable performance thresholds and a structure resembling parametric insurance, where an agreed trigger helps determine a claim.

That is evidence of an emerging market, rather than a universal answer. A policy built around a specified performance threshold needs to be understood on those terms. Buyers still need to establish whether the particular loss they are concerned about falls within its scope.

The harder problem may be the number of businesses exposed to the same underlying failure. In the July research report “Underwriting the Agent Economy,” Cristian Trout and co-authors warn that concentration among a small number of foundation-model providers creates the prospect of correlated losses. If many businesses depend on the same model, their risks may be connected even when they operate in different industries.

The authors also argue that rapidly changing capabilities make conventional actuarial approaches difficult to apply. Their proposed response includes better incident data and monitoring, alongside changes to insurance contracts and claims management. It proposes developing the market, not that comprehensive coverage already exists.

For enterprise buyers, the implication is to bring insurance questions into the deployment decision while an agent’s authority is still being defined. Records of what the system could access, and what it actually did, can help make the exposure concrete. A vague promise of human oversight offers less to examine than a documented approval requirement for a specific action.

Insurance could consequently influence how agents are designed. If a buyer must demonstrate bounded authority and observable behavior to obtain suitable coverage, those controls acquire a financial value alongside their technical purpose.

The next enterprise AI evaluation may need someone who can read an insurance contract as carefully as the technology team reads the benchmark results.