Insurance can unlock a multi‑trillion dollar AI agent economy, but current silent coverage and rising correlated risks mean coverage must be redesigned; industry-wide standards, pooled instruments and government backstops are needed to enable billion-dollar affirmative policies and manage catastrophic AI scenarios.

Main Finding

A coordinated, multi-part "AI insurance stack" is necessary to make the emerging AI agent economy insurable at scale. Without industry-wide data sharing, new standards, forward-looking underwriting tools, and public-private backstops, insurers will either silently carry rapidly growing, correlated exposures or retreat — either outcome risks stalling responsible adoption or producing catastrophic economic losses. With timely coordination, affirmative AI coverage (with limits in the billions) is achievable by ~2030; without it, systemic accumulation and tail risks will overwhelm conventional insurance mechanisms.

Key Points

• Role of insurance: Historically enabled major technologies by limiting downside, creating incentives for best practices, and compensating harmed parties. AI agents are the next such domain.
• Current gap: Most frontier AI agent risk is currently "silent coverage" hidden inside cyber, professional, and general liability policies and largely unpriced.
• Rapid adoption & concentration: Enterprise frontier AI spend rose >300% in 2025; >80% of deployments rely on three foundation model providers — creating strong concentration and accumulation risk.
• Deteriorating insurability: Agent capabilities are improving faster than reliability; incident severity upper bounds appear to be rising (examples from hallucinated policy errors to wrongful-death events). AI agents are dynamic risks that may outpace traditional actuarial approaches.
• Eight-component AI insurance stack: recommended components are (1) incident data collection & analysis, (2) accumulation risk research & CAT modeling, (3) standard setting, (4) contract design, (5) risk selection & evaluation, (6) pricing, (7) ongoing loss control & monitoring, and (8) incident response & claims management.
• Coordination & public goods: Many stack components are complements or public goods (shared incident DB, CAT scenarios, standards, model policy language); cold-start problems make industry coordination and government roles important.
• Historical precedents: Insurers previously created Underwriters Laboratories and pooled closed-claims datasets to manage new risks — similar institutions are recommended for AI.
• Tail solutions: Societal-scale frontier risks (CBRN-like, systemic infrastructure collapse, loss-of-control scenarios) will need alternative risk transfer: mutuals, catastrophe bonds, bespoke liability regimes, and government backstops.
• Macro risk: A relatively limited direct-loss catastrophe (order ~$100 billion) could cascade into multi-trillion GDP losses if it triggers economic slowdown and broad withdrawal of insurance (analogy to post-9/11 effects on terrorism insurance).

Data & Methods

• Multi-method synthesis: the report combines literature review, historical analogy, industry surveys/consultations, and empirical indices to characterize risk and inform recommendations.
• Incident-to-usage analysis: authors construct an incident index and a usage index (see Appendix 1) to estimate trends in incidents relative to frontier-AI usage; sensitivity analyses (e.g., leave-one-out bands) are used to test robustness.
• Market measures: uses observed enterprise spending growth (reported >300% in 2025) and market-concentration statistics (foundation model provider dependence).
• Qualitative evidence: interviews, industry questionnaires (e.g., Lloyd’s underwriters surveyed on perceived policyholder preparedness), and case examples of incidents (ranging 2022–2025) to illustrate rising severity and heterogeneity in harms.
• Scenario and risk-modelling prescriptions: recommends developing formal AI catastrophe scenarios and accumulation models (single-point failures, supply-chain concentration, emergent multi-agent failure modes) though the report does not present final CAT models itself.
• Comparative institutional analysis: examines prior insurance-market solutions (Underwriters Laboratories, Closed Claims Project, cyber insurance lessons, post-9/11 terrorism insurance) to derive governance and structural recommendations.

Implications for AI Economics

• Pricing and markets
  • Need for new pricing inputs: forward-looking performance evaluations (red-team/pen-test-style scores), standardized proposal forms, and continuous monitoring must be integrated into premium-setting to avoid adverse selection.
  • Rapid capability improvement implies actuarial histories will be a weaker signal; insurers should rely more on real-time/forward assessments and model-based stress testing.
  • Concentration in model providers creates strong positive correlation in losses, requiring explicit accumulation loadings, reinsurance, and novel market instruments.
• Investment and adoption
  • Robust insurance lowers adoption risk and thus accelerates productive diffusion of AI agents; conversely, failure to create credible coverage could materially slow uptake and investment.
  • Insurance can create incentives for safety investments (underwriting-backed standards and discounts), aligning private incentives with social risk reduction.
• Market structure and public goods
  • The sector faces a collective-action problem: many crucial assets (incident databases, CAT scenarios, standards, accredited auditors) are public goods. Without coordination, private markets may under-supply these, producing market failure or fragile equilibria.
  • Government and regulatory roles are pivotal: mandatory disclosure rules for severe incidents, support for anonymized reporting systems (NIST-like), backstops or reinsurance for systemic tail events, and guidance on recognizing private standards.
• Systemic and macroeconomic risk
  • AI-related catastrophes have non-linear macro impacts because of amplification via lack of insurance, financial stress, supply-chain effects, or confidence shocks — making conventional micro-insurance solutions insufficient for extreme tails.
  • Policy instruments (government backstops, cat bonds, industry mutuals) will be needed to transfer systemic tail risk; their design will shape incentives and distributional outcomes across firms and consumers.
• Competition and concentration
  • As insurers require more granular telemetry and model-level controls, large model providers and well-resourced enterprise adopters may obtain better terms — potentially increasing incumbency advantages and raising barriers for smaller entrants unless standards and access are managed equitably.
• Regulatory economics and legal frameworks
  • Clearer legal rules and model policy language reduce ambiguity and litigation cost; regulators can accelerate market formation by endorsing standards, requiring incident reporting, and curtailing silent coverage.

Summary recommendation (economic takeaway): creating a functional AI insurance market is both feasible and necessary to unlock the economic benefits of AI agents. Doing so requires coordinated investment in shared data, CAT modeling, prescriptive standards, and novel risk-transfer mechanisms; absent that coordination, market failures and systemic tail risks could impose outsized macroeconomic costs.