Underwriting
the Agent Economy:The Blueprint
for an AI Insurance Stack

Extended Summary

From maritime trade to commercial nuclear power, insurance has enabled economic growth and major technological developments by limiting downside risk for companies and their investors, quantifying risk, supporting safety research, spreading best practices, and ensuring third parties are compensated after an accident [1], [2], [3].

We believe AI agents — AI systems that receive high-level natural language instructions, form plans, and then take consequential actions in the world with limited or no human oversight — represent another such major technological development. Analysts expect AI agents to be producing hundreds of billions of dollars in economic value by 2030 [4], [5], [6], [7]. Realizing that promise will depend on enterprises having the confidence to adopt these systems at scale — confidence that, in turn, will depend on insurers’ willingness to help absorb and manage the risks.

Developers of AI agents, the enterprises that deploy them, and the foundation model providers they depend on all need insurance. They need coverage for both first-party operational losses, where an agent damages the policyholder’s own systems, data, operations, or reputation; and for legal liability, where the policyholder’s AI agent product or service harms a business partner, customer, or member of the public, who then sues.

But insurers are not ready. Today, the vast majority of AI agent risk sits in “silent coverage” within existing cyber, professional liability, and general liability policies — unpriced, invisible, and potentially destabilizing [8], [9], [10]. Underwriters also appear to be underestimating the risk posed by their policyholders’ usage of AI agents: nearly 50% of Lloyd’s underwriters surveyed believe their policyholders have “adequate” AI risk management [11], while only 1 in 5 surveyed businesses report having a mature model for governance of autonomous agents [12]. This suggests a significant disconnect.

Insurers are beginning to write exclusions to correct for this unpriced risk [13], [14], [15], [16]. While perhaps necessary, this only widens the coverage gap, leaving businesses to fend for themselves when they most need support. Some 60% of business leaders admit having “intentionally slowed implementation due to concerns over potential errors and malfunctions” [17]; it is no coincidence that over 90% report wanting insurance tailored to frontier AI risk [18].

This report argues that insurers can and should reassume their role as enablers of innovation. We claim that affirmative coverage for AI agents, with enterprise coverage towers reaching the billions, is achievable by 2030, but only if the insurance industry works together to make AI agents more insurable. Coverage will remain theoretical unless insurers can reliably profit from it at premiums that customers are willing to pay.

The challenges are several. First, frontier AI is too fast-moving and too general-purpose for conventional actuarial approaches. Consider: the length of tasks that agents can accomplish without human intervention is doubling roughly every four months [19], [20], and new use cases and vulnerabilities evolve on a similar timeframe. This means the insurer’s usual playbook — offer limited coverage, then expand as loss data and actuarial models improve — will not work: actuarial models will struggle to remain predictive, always lagging behind the shifting reality on the ground [21], [22].

Second, while the alignment and accuracy of AI agents have so far been improving — published benchmarks tracking helpfulness, harmlessness, and honesty show sustained gains across frontier model generations [23], [24], [25], [26] — reliability improvements (consistency, robustness, predictability) are being outpaced by capability improvements [26], [27]. As a result, the upper bounds of incident severity appear to be rising. Hallucinated refund policies in 2022 [28] have given way to wrongful death suits and unauthorized practice of law cases in 2025 and 2026 [29], [30], [31], [32].

Recent developments reinforce this trend. In April 2026, Anthropic unveiled Claude Mythos Preview, a frontier model capable of autonomously discovering and exploiting zero-day vulnerabilities across every major operating system and browser [33], [34], [35], [36]. Experts fear future models will gain other dangerous dual-use capabilities (e.g. in synthetic biology) [37], [38], [39]. Deployed responsibly, such models promise prosperity; but mistakes will be made, and the transition to a society resilient to such powerful technology will likely be disruptive.

The situation is also not helped by concentration in the foundation model market, where three providers account for over 80% of deployments [40]: a defect at any one propagates as a correlated shock across thousands of policyholders. This and other single points of failure create textbook conditions for heavy-tailed loss distributions. In short, if little effort is made to control tail risks, premiums and capital requirements will be not only harder to calibrate but higher — mirroring challenges that have plagued cyber insurance for decades [41], [42].

Getting ahead of the challenge means developing a technical understanding of the risks and controlling their tails. The bulk of this report describes the insurance infrastructure stack needed to do so — to guarantee agentic AI remains insurable and can be adopted with confidence. We break down the stack into eight components:

1. Incident Data Collection and Analysis (§II.1): Pooling data on incidents and near misses, combining public reports with private policyholder data and claims data, in standardized formats. Treated as a shared resource, such data can benefit all parties, as demonstrated by the Closed Claims Project for anesthesiology malpractice: this program uses insurer claims data to develop new safety measures for anesthesiologists, reducing premiums and expanding the insurance market in the process [43], [44], [45], [46], [47], [48]. However, such a resource depends on insurer coordination: government intervention (e.g. incident disclosure mandates, and information sharing safe harbors) may be necessary [3], [49].

2. Accumulation Risk Research and CAT Modeling (§II.2): Modeling catastrophe scenarios to better understand accumulation risk arising from single points of failure (such as the handful of frontier model providers), correlated triggers, and systemic risk (such as emergent multi-agent failures). Insurers are also uniquely incentivized to develop and disseminate macro-level mitigations for such risks (e.g. “circuit-breakers” for the agent economy) — public goods that would mirror their investments in cyber CAT scenario modeling and safety R&D at the Insurance Institute [50], [51] for Highway Safety and Insurance Institute for Business & Home Safety [52].

3. Standard Setting (§II.3): Supporting auditable, prescriptive standards that set the minimum requirements for insurability and play three reinforcing roles: as underwriting tools that speed binding and supply contractual hooks for claims handling; as controls that bound correlated legal risk by anchoring the duty of care; and as concrete loss control guidance for policyholders. Wary of both the preventable damages they might pay for and the compliance burden on their customers, insurers are well-incentivized to balance stakeholders’ competing interests in standard-setting. Insurers have been involved in standard-setting before, most famously in 1894 when they founded the Underwriters Laboratories (UL) to manage the novel fire hazards of electrical equipment [53]. The UL mark became a litmus test for insurability and was later codified into law [54], [55], [56], [57], [58].

4. Contract Design (§II.4): Agreeing on clear, consistent definitions of covered AI risks, appropriate exclusions, aggregation clauses, and trigger mechanisms. These terms can end ambiguous silent coverage and move toward the affirmative coverage that gives portfolio managers and actuaries visibility into AI risk exposures and loss distributions, respectively. This is critical for avoiding the decade of costly ambiguity that plagued cyber insurance [10], [59], [60].

5. Risk Selection and Evaluation (§II.5): Developing AI specific underwriting practices that go beyond good-faith, annual questionnaires to incorporate organizational audits, realistic performance evaluations, red teaming, and other recurring technical tests as capabilities and vulnerabilities shift month to month. Standards can accelerate the process, providing technical test data and harmonizing questionnaires. In the limit, they enable streamlined standards-based rating like the widely used Fire Suppression Rating Schedule from the Insurance Services Office [61], cf. [62].

6. Pricing (§II.6): Developing an expected-loss formula that leans on performance evaluation results and usage telemetry to supplement immature actuarial tables, feature-rates on system specifications, safeguards, and deployment sector, before adding accumulation risk loading. Because this approach is system-specific and forward-looking, it enables genuine price differentiation between high- and low-risk deployments. This expands the addressable market and turns premium signals into a lever for driving safety improvements.

7. Ongoing Monitoring and Loss Control (§II.7): Investing in ongoing risk management guidance and oversight for policyholders, to keep pace with the rapid evolution of AI agents. Like cyber, the AI insurance market might mature in two stages: first, frequent, labor-intensive performance evaluations that build institutional knowledge of which mitigations are effective; then, scaling through partnerships with cloud service providers and foundation model providers who supply telemetry for low-friction underwriting and monitoring [63]. With appropriate leadership, AI insurers can compress cyber’s decades-long learning curve.

8. Incident Response and Claims Management (§II.8): Developing an AI-literate claims management process that pays valid losses promptly and enforces exclusions, along with an AI-specific incident response function layered onto the existing breach coach and crisis management ecosystem for cyber and recall insurance lines. Finally, rigorous AI-native post-incident forensics will turn every loss into a learning opportunity. As cyber proved, a technology moving as fast as AI requires a feedback loop — from claims processing to forensics to revised underwriting criteria and actuarial models — that turns far faster than the traditional annual cycle.

We categorize the components by their role in the value chain: foundational functions, supporting functions, product & underwriting functions, and functions for servicing policies (see Figure 1). The components are complements in important ways. Incident response and claims management, for example, are key sources of data, whose analysis then feeds into virtually every other layer, especially standard setting, risk evaluations, pricing, and accumulation risk research. Likewise, contractual exclusions of losses caused by upstream model failures, aimed at controlling accumulation risk, are unenforceable without appropriate logs for post-incident forensics — exactly the kinds of technical controls mandated by robust standards and underwriting practices.

Of course insurance has inherent limits, and cannot manage all of frontier AI’s risks. Insurers will always struggle to control or price criminal misuse of AI systems, much as they have with cyber (in part due to third-party moral hazard): insurance cannot replace law enforcement.

Looking ahead, we foresee the need for purpose-built institutions and alternative capital structures dedicated to covering and controlling catastrophic risk from frontier AI, or “AI CAT.” (For reference, in the world of insurance “catastrophic risks” refer to low-probability, high-severity loss events roughly in the hundreds of millions of dollars or more). Catastrophes reaching the low tens of billions will strain but not exceed private-market capacity. However, beyond this threshold lies a class of catastrophes whose magnitude and structure render them all but uninsurable by private markets. We refer to these as societal-scale risks. Picture critical infrastructure collapse, systemic economic disruption, or CBRN¹ risks.

As we approach artificial general intelligence, many experts warn of more societal-scale risks on the horizon. Besides dual-use capabilities, such as automated zero-day discovery and synthetic biology capabilities, which could enable widespread cyber attacks and bio-terrorism (respectively), some experts also warn of the possibility of losing control over advanced AI systems that develop goals misaligned with their operators’ intentions [39]. In the limit, they warn, this could lead to human extinction [39], [64], [65]. Of course no institution can insure extinction risks, but covering risks correlated with extinction may help control them.

Private markets can provide first-layer coverage and help with pricing, distribution, and claims management, but the bulk of coverage capacity for societal-scale risks will depend on governments, society’s de facto insurer of last resort [3], [66].

We sketch several complementary mechanisms for handling AI CAT that are worth exploring further: industry mutuals modeled on self-regulatory structures in commercial nuclear power and other industries [3], [67], [68]; catastrophe bonds (“CAT bonds”) for additional risk transfer [69]; bespoke liability regimes to correct for externalities and improve insurability [70], [71], [3], [72], cf. [73]; and government backstops to take on the risks only government can manage, while giving private insurers the confidence to shoulder more catastrophic risk than they otherwise would [66], [3], [74], cf. [21] (see Figure 2).

If insurers fail to quickly get a handle on AI agent security, safety, and reliability, an AI insurance coverage gap will rapidly balloon, with consequences for the broader economy. In the best case, the Artificial Intelligence Underwriting Company estimates that the resulting drag on AI adoption could leave some $200 billion in US GDP on the table over a decade, based on IMF productivity-diffusion estimates under a scenario of lower institutional readiness. In the worst case, a disaster whose direct damages only cost some hundred billion dollars could wipe out several trillions of US GDP over five years due to an economic slowdown cf. [75], [76]. The mechanisms that might drive such a slowdown include: a violent withdrawal of insurance coverage for AI-related risk, stalled enterprise adoption of AI, regulatory backlash, and investors suddenly pulling out of AI as trust in the technology craters [77], [78], [79], [80]. There is historical precedent for this: after 9/11 the abrupt collapse of the insurance market for terrorism risk froze major construction projects and grounded commercial aviation until ad hoc government intervention stabilized the situation [81], [82], [83], [84]. If the AI disaster is an accident, we could see adoption set back a decade given negative public sentiment about AI [85], much like nuclear power post Three Mile Island or Fukushima [3].

Building the stack needed will require coordinated action across brokers, Managing General Agents (MGAs), primary insurers, reinsurers, risk modelers, standard-setting bodies, oversight bodies and government. The insurance industry must overcome a cold-start problem: many of the components described are complements, working best in synergy. The alternative for insurers — guarding what little proprietary data one has while relying on low limits and blanket exclusions — may protect today’s fragile margins, but cedes the opportunity to expand the serviceable market, enable responsible AI adoption, and protect the broader economy from shocks.