Who Writes AI Insurance Coverage Conditions and Who Absorbs Correlated AI Loss: The Evidence Bottleneck, the Capital Bottleneck and the Public Policy Choice Above Them

Two Foresight-to-Duty Conversion papers on insurers, reinsurers and Congress after the 2026 AI exclusions

Actors and jurisdictions: AI deployers · Foundation model developers · Enterprise customers · Integration partners · Lenders · Property and casualty insurers · Reinsurers · Lloyd's and Munich Re · Verisk · Cloud platforms · State insurance commissioners · State attorneys general · California · Illinois · Washington · Congress · European Union

Companion line: Installments VI and VII of the Foresight-to-Duty Conversion Series start where Who Sets the Agentic Standard of Care stopped: insurers run closest to the deployment clock, so the two papers ask which AI losses insurers will cover on conditions and who holds the losses they will not.

Read the full publications: Who Writes AI Insurance Coverage Conditions — Procurement, Enforcement Decrees, Partner Terms and the Two-Axis Condition · Who Absorbs Correlated AI Loss — Reinsurance Aggregation, Industry Pools, the Public Backstop and the Liability Shield

The Regulator That Governs by Withdrawal

In January 2026 Verisk's Insurance Services Office released optional endorsements that remove generative AI from the commercial general liability form. More than sixty property and casualty carriers filed exclusions of their own during 2026, and state insurance commissioners approved more than eighty percent of the requests. The Center for Strategic and International Studies called insurance "the most important de facto regulator of U.S. AI deployment" in September 2026 and observed that it was regulating by withdrawal.

The same year produced the opposite move. A Fortune 100 buyer wrote a clause into a vendor agreement requiring the provider to carry insurance for AI-related risks, and Munich Re published limits of up to $100 million across models with similar use cases. The fastest institution in the series now covers some AI risks on conditions and leaves the rest with firms and the people their agents harm.

Central Finding

Insurability of AI loss has two bottlenecks, and a different set of actors controls each. Credible evidence about which losses a deployer can control is the first, and buyers, attorneys general and distribution partners supply it faster than any statute. Sufficient capital for a correlated loss from one widely deployed model is the second, and no carrier has published capacity for it. Above the private layers the loss falls on firms and injured parties, and whether anyone else absorbs it is a public policy choice that three precedents show Congress making both before and after a salient loss.

Four Loss Categories Sort Into Two Bottlenecks

Ordinary deployment loss is an identifiable error within a specified system and within policy limits: a hallucinated report, a misrouted payment or an agent that calls the wrong tool. Carriers write it today on conditions because frequency is observable at the insured and the maximum loss is bounded by the deployment. The coverage dispute is a loss that sits inside a policy and goes unpaid over an exclusion or a notice clause, which makes it a wording problem rather than a capacity problem.

Correlated loss is the third category. One defect in one foundation model can surface at many insureds at once, and Gallagher Re's March 2026 white paper said such failures could accumulate across ceded portfolios in ways existing models cannot capture. A June 2026 survey of 26 public AI insurance products found foundation-model concentration failing two insurability criteria, loss independence and economic feasibility, and treated the maximum possible loss as effectively unbounded because the base of insureds using one model keeps growing.

The compensation gap is the fourth: losses above limits or outside any coverage, which fall on the firm and the injured party. An audit mandate supplies information about the first two categories and no claims-paying capital for the last two. The Geneva Association's 2025 assessment found generative AI failing three of Baruch Berliner's nine insurability criteria, and those failures bind hardest on correlated loss.

Affirmative Cover Arrives by Peril and Underwrites on Different Evidence

Affirmative AI cover now reaches the market through four models that are not interchangeable capacity. Armilla's Lloyd's-backed policy writes up to $25 million per organization against model errors and agent failures with independent certification attached. Testudo writes up to $9.25 million against hallucinations and property damage on a claims-made form and states that its application requires no invasive technical audit.

Munich Re's aiSure performance guarantee launched through Mosaic in February 2026 and pays a pre-agreed amount when a model's accuracy falls below a threshold set at binding. CFC added explicit AI wording across seven existing coverages in June 2026, which removes a coverage dispute before it starts and adds no capital. Google Cloud's Risk Protection Program shares a customer's security posture with Munich Re, Beazley and Chubb in place of a manual application. Harbor raised $10 million on October 13, 2026 for a platform that rates every agent a client runs and blocks rule-violating decisions before execution.

The products differ in whether measurement continues after binding. Armilla's certification, Munich Re's thresholds and Harbor's re-rating on change all keep measuring after the policy attaches, while Testudo's scheduled systems and Google's posture sharing describe how cover is obtained. None binds cover separately before each agent action, so the per-action design sketched in Agentic Foreseeability and the Modern Duty of Care for AI Agents remains a design.

Buyers, Enforcers and Partners Create Demand Where Statutes Have Not

A September 2026 review of 69 AI liability bills across 27 states found none with an insurance mandate. States have legislated verification instead. California's SB 1119 requires companion-chatbot operators to obtain an independent child-safety audit every two years from 2029, and Illinois's frontier-model law requires annual audits and 72-hour incident reporting from large developers. Both passed in 2026 and neither carries an insurance term.

Procurement has moved further than any legislature. Armilla reports an unnamed Fortune 100 vendor agreement that requires the provider to obtain and maintain insurance for AI-related risks covering bias and hallucination and copyright and regulatory penalties. The same agreement requires annual compliance testing, prior approval for AI deployments and two years of audit rights. Enforcement has reached the threshold without crossing it: on October 14, 2026 a coalition of 39 state attorneys general filed notice in the Character.AI bankruptcy of the terms they will press on any buyer, and the notice contains age verification and crisis-response terms but no audit or insurance term.

A mandate binds only where a policy exists to buy. A requirement to carry cover that no carrier writes is a requirement not to deploy, which is why mandates in other sectors followed the market rather than led it.

The Provider's Liability Cap Is the Hinge of the Upstream Chain

Pressure runs from users to enterprise customers and attorneys general, through distribution partners to the providers that supply the model, and ends at the insurer that decides which provider commitments it will cover. Two links already exist in published terms. Microsoft's commercial marketplace requires agents listed for Microsoft 365 Copilot to block harmful content and act on user reports as conditions of listing, and OpenAI's service terms indemnify API and Enterprise customers against third-party intellectual-property claims arising from output and exclude that indemnity from the liability cap.

Most AI vendor agreements cap total liability at twelve months of fees, which leaves the deployer holding customer loss it cannot recover. In the simulation a provider keeps that cap whenever no partner or customer requires more and accepts liability above it in two cases of three when both require it. Once a partner requires disclosure of material model changes as an access condition, providers disclose in nearly every simulated case.

Lenders sit beside the chain. A lender can write AI cover or audit evidence into credit terms the way it writes insurance covenants for property, and the simulation finds lenders doing so in about half the cases once customer requirements exist. No lender class has published such a covenant.

A Coverage Condition Carries the Reassessment Duty on Two Axes

Who Sets the Agentic Standard of Care proposed a duty to reassess a deployment after each material permission change or incident, and insurers can impose that duty through a coverage condition. A condition that produces correction needs four elements: a material-change trigger rather than a named statute, a required response within a defined time when a finding surfaces, verifiable evidence that the control worked and a stated consequence for noncompliance.

Prevention and compensation sit on separate axes. A condition reduces risk whatever the consequence, because an insured that wants to keep cover changes behavior before any breach. If noncompliance voids the policy, the injured party loses the payment the policy would have made, so a condition that prevents well can compensate badly. Compulsory auto insurance resolves that tension by preserving third-party payment and letting the carrier recover from the insured afterward, and an AI condition can adopt the same split.

Stated Limits Stop at $100 Million and Reinsurers Decide Whether They Grow

Munich Re's corporate factsheet states limits of up to $50 million per model and $100 million across models with similar use cases, and its separate provider factsheet caps standard cover at $5 million. Advertised limits do not establish what buyers obtain, and no bound placement at those limits has been publicly reported.

Reinsurers hold the choice that decides whether those limits grow. A reinsurer can treat foundation-model failure like a named storm with an aggregation trigger and a sublimit, or exclude it and require standalone cover. The named-storm path collects premium on measured accumulation that exclusion forgoes, and the June survey recommends a Lloyd's-style market study on AI aggregation triggers of the kind that resolved silent cyber in 2019 and 2020.

Nuclear, Vaccine and Terrorism Precedents Offer Three Instruments and Two Triggers

Congress reached the same point with nuclear power in 1957 and acted before any accident. The Price-Anderson Act combined primary insurance per site at the commercial maximum, now $500 million, with a retrospective industry assessment of about $158 million per reactor that forms a secondary layer near $15 billion across 95 reactors. Three Mile Island came twenty-two years later.

Vaccines reached it in 1986 after manufacturers began leaving the market, and the National Vaccine Injury Compensation Program funds no-fault awards under an injury table through an excise tax on each dose while keeping a route back to tort. Terrorism reached it fourteen months after the attacks, on November 26, 2002. The Terrorism Risk Insurance Act shares qualifying losses between insurers and the federal government above a deductible. The three precedents supply a backed market, a funded compensation system and a loss-sharing formula, and each kept a route to tort.

The Shield Removes the Exposure Without Reducing the Harm

OpenAI backed an Illinois bill, SB 3444, that would shield frontier developers from catastrophic-harm liability if they published a safety protocol before release. Anthropic opposed it, and the bill's sponsor said sweeping relief was unlikely to survive. A shield earned by publication pays for files rather than correction and moves the residual loss to the injured party alone.

The simulation plays twenty Cognitive Digital Twins (CDTs) against one correlated-loss scenario: developers and carriers and reinsurers on one side, state legislatures and congressional committees on the other. State legislatures choosing on their own payoffs prefer financial responsibility to a shield by five to one while preferring neither in half the cases. Developers back a pool in one case of four, reinsurers publish an aggregation limit more often than they exclude, and Congress introduces a public layer in fewer than one case in four without a loss and in more than three of four after one.

Game theory supplies the payoff structure and behavioral economics the decision rules. Underwriters facing a loss they cannot measure exclude and wait, and once a standard exclusion form exists no carrier gains by being the one that keeps writing the risk. A buyer's contract then supplies the template an attorney general can copy into a decree, and a qualifying loss changes every player's payoff at once.

What the Full Publications Add

The two papers carry nine MindCast Foresight Simulation Predictions (five primary and four secondary) with a deadline and a falsifier and a named verification source for each. The papers frame seven competing hypotheses separately before any result prints, three on who writes the condition and four on who absorbs correlated loss, and leave three further questions open without bands. Nine Risk Mitigation entries carrying nineteen actions name an owner function and a deadline for each action. The audiences run from deployers and enterprise customers through partners and lenders to developers, carriers and state and federal officials.

The evidence base includes the 26-product survey's 55-peril matrix (53 percent affirmative and 20 percent excluded among stated positions), Aon's 112-cell coverage matrix (90 excluded against 13 available) and the Verisk forms CG 40 47, CG 40 48 and CG 35 08. The instruments include California SB 1119 and Illinois SB 315 and SB 3444, the Character.AI bankruptcy notice and Microsoft's marketplace certification policies. OpenAI's service terms and services agreement and Munich Re's corporate and provider factsheets supply the contract and capacity terms. The precedent analysis draws on the Congressional Research Service's Price-Anderson figures, and a dated watch calendar runs from the January 2027 treaty renewal through the end of 2028.

Read the full publications: Who Writes AI Insurance Coverage Conditions — Procurement, Enforcement Decrees, Partner Terms and the Two-Axis Condition · Who Absorbs Correlated AI Loss — Reinsurance Aggregation, Industry Pools, the Public Backstop and the Liability Shield

MindCast Simulation Predictions

The combined forecast set holds nine Simulation Predictions. P marks a Primary Simulation Prediction and S a Secondary Simulation Prediction, and all four secondary entries are conditional and print as P(A given B). No entry in either paper has reached its deadline.

From Who Writes AI Insurance Coverage Conditions:

• P-1 · A state attorney general consent decree or assurance of discontinuance with an AI developer or deployer requires an independent audit or an insurance term by December 31, 2027 (17-62%). Falsifier: no such decree or assurance by the deadline.

• P-2 · A major model or tool provider publishes standard enterprise terms accepting liability above twelve months of fees for third-party claims arising from outputs other than intellectual-property infringement by December 31, 2027 (2-53%). OpenAI's existing intellectual-property indemnity forms the baseline and does not resolve the entry.

• S-1 · Given a Washington AI product-liability statute enacted from the 2027 regular session, the enacted text or implementing rules adopted by December 31, 2028 reference insurance or financial responsibility or set a coverage condition (40-86% given the statute). The statute's own likelihood is forecast in Who Sets the Agentic Standard of Care.

• S-2 · Given a lender class publishes an AI insurance or audit covenant for AI-dependent borrowers before December 31, 2028, a state enacts a financial-responsibility requirement for AI deployers by that date (48-95% given the covenant). No covenant leaves the entry unresolved.

From Who Absorbs Correlated AI Loss:

• P-1 · A single carrier or syndicate publicly confirms a bound AI liability policy of $100 million or more for one insured by December 31, 2027 (27-48%). Advertised limits and layered placements across multiple carriers do not count.

• P-2 · A reinsurer or Lloyd's publishes an aggregation trigger or sublimit for foundation-model failure in a market bulletin or treaty wording by December 31, 2027 (16-56%). Falsifier: no published trigger or sublimit by the deadline.

• P-3 · A state enacts a liability shield for frontier developers conditioned on publishing a safety protocol by December 31, 2027 (2-34%). Falsifier: no such enactment by the deadline.

• S-1 · Given no qualifying loss through December 31, 2027, a member of Congress introduces a bill creating a federal pool or backstop or compensation system for correlated AI losses by that date (3-24% given no loss).A qualifying loss is one foundation-model failure producing claims against three or more unrelated insureds that together exceed $1 billion.

• S-2 · Given a qualifying loss, a member of Congress introduces such a bill within twelve months of its first public report (75-96% given a loss). Falsifier: a qualifying loss occurs and no such bill follows within twelve months.

Read together, the two sets describe a market that grows on evidence while the public layer waits for a loss. Attorneys general and provider terms are the nearest tests of the demand side, a bound $100 million placement and a published aggregation trigger are the nearest tests of the capital side, and the shield stays a minority route in every legislature the simulation models.

Stakeholder Callouts and Mitigation

🤖 AI deployment leaders. The exclusion that removed AI from general liability arrived as an optional form, so no deployer knows its position without reading the issued policy, and correlated loss from the model it deploys stays with it above the carrier's limit. Mitigation: list each AI exclusion by form number at the next renewal; model three retained-loss scenarios for the largest deployed agent and size a reserve by the next audited statement; renegotiate the twelve-month-fees cap for output claims at the next vendor renewal or document the exposure the cap leaves. Residual: correlated losses above the company's capacity remain with the company and its claimants.

🛒 Enterprise customers and procurement. Vendor AI losses pass through the contract to the buyer in the gap between the vendor's cap and the buyer's loss. Mitigation: require evidence of AI-specific cover rather than a certificate of general liability and name the perils it must cover in the next master services agreement; carve third-party claims arising from AI outputs out of the consequential-damages waiver and the twelve-month cap. Residual: a vendor that cannot obtain cover signals the risk the buyer is accepting, and the buyer either declines or retains it.

🔗 Integration and distribution partners. A provider failure travels through the partner's channel to every business that embeds the agent. Mitigation: add disclosure of material model, tool and data changes to listing conditions with a notice period before the next policy revision; require evidence of AI-specific cover or an accepted liability commitment from providers above a usage threshold in the 2027 program cycle. Residual: a provider that leaves the program rather than disclose moves the exposure to businesses that keep its agent through other channels.

🧠 Foundation model developers. A public layer drafted after a loss arrives on terms set by others, measured in the retention and assessment the statute assigns. Mitigation: share incident data with carriers or a database in a form that supports underwriting through lawful and confidential channels before the 2027 state sessions open; contract audits that meet Illinois's standard a year before it binds on January 1, 2027. Residual: a shield-for-disclosure bill enacted elsewhere resets the terms of every later negotiation.

🛡️ Carriers and reinsurers. A carrier that excludes while competitors collect data re-enters a market written on their terms, measured in renewal cycles before competitive re-entry. Mitigation: write conditions with a material-change trigger, a required response and a consequence that preserves third-party payment at the next form filing; set an aggregation sublimit for foundation-model failure and name the trigger at the January 2027 treaty renewal rather than excluding the peril. Residual: correlated losses can accumulate beyond the carrier's available capacity whatever the conditions.

🏦 Lenders and investors. A borrower's or portfolio company's retained AI loss reaches the lender as a covenant breach or impaired collateral and reaches the investor as a valuation the market will not carry at current terms. Mitigation: add AI exclusions to the insurance covenant review at origination and renewal; require affirmative cover or a retained-loss reserve as a covenant for borrowers whose revenue depends on a deployed agent; place a correlated-loss retained-loss scenario in diligence for every AI deployer and developer in the portfolio until a pool or public layer exists. Residual: covenants and diligence discover a retained loss and do not pay it.

🏛️ State legislators and attorneys general. A liability statute that expands exposure into an excluded market strands deployers, and a decree template set by another state or a bankruptcy sale binds the office to terms it did not write. Mitigation: pair any AI liability bill with an incident-reporting duty carriers can use and allow a bond or asset alternative to any cover requirement before policy-committee action; direct the insurance commissioner to report on AI exclusion filings and available capacity in the same session; draft a standard audit-and-cover term with a named recipient who can halt the product before the first 2027 AI matter and coordinate it through the multistate coalition. Residual: capacity depends on carriers outside the state's jurisdiction, and a decree term binds one company rather than the deployers that never draw a complaint.

🏛️ Federal committee staff. A public layer drafted in the week after a loss absorbs a share of that loss the committee did not choose. Mitigation: draft the nuclear, vaccine and terrorism structures now with each costed against a correlated-loss scenario before the 120th Congress organizes; authorize the incident database as a standalone measure that needs no public layer to pass. Residual: a drafted option still waits for the vote that enacts it.

Conclusion

Insurance governed AI deployment in 2026 by withdrawing from it, and the two papers show what the withdrawal left behind. Buyers, attorneys general and partners can supply the evidence that lets carriers write ordinary deployment loss on conditions. A condition built on a trigger and a required response and verified control carries the reassessment duty onto the deployment clock while preserving third-party payment. Correlated loss from one model sits above every stated limit, and no carrier, reinsurer or legislature has yet absorbed it.

The January 2027 treaty renewal, the first attorney general decree with an audit or insurance term and the first bound $100 million placement will show whether private capital grows into the gap before a qualifying loss hands Congress the pen.

Working With MindCast

MindCast AI applies Predictive Behavioral Economics + Dynamic Game Theory through MindCast AI Proprietary Cognitive Digital Twin Foresight Simulations (MP CDT FS). Behavioral Economics supplies the decision rules. Game Theory supplies the payoff structure. Predictive simulations emerge from the combination.

For the questions these two papers raise, MindCast builds Cognitive Digital Twins of carriers and reinsurers, of developers with their customers and partners and of the enforcers and committees that govern them. The engagement simulates competing coverage conditions against a specific deployment's harm pathways and tests the nuclear, vaccine and terrorism structures against the same correlated-loss scenario. For deployers the output shows which exposures no policy covers and what condition closes the gap; for buyers and partners it shows which requirement moves a provider's cap; for reinsurers it shows which trigger produces premium rather than exclusion; for attorneys general and committee staff it tests a decree term or a public structure against the market's likely response before either is negotiated.

MindCast offers a complimentary pilot built from public filings and backtested on a known case before any retainer. No engagement proceeds without a named decision to resolve. Contact [email protected].

Related Works

Publications

• Who Writes AI Insurance Coverage Conditions — Procurement, Enforcement Decrees, Partner Terms and the Two-Axis Condition · Installment VI. Carries the evidence bottleneck, the demand routes and the two-axis condition.

• Who Absorbs Correlated AI Loss — Reinsurance Aggregation, Industry Pools, the Public Backstop and the Liability Shield · Installment VII. Carries the capital bottleneck, the precedents and the public-layer forecasts.

Series

• Who Sets the Agentic Standard of Care — Common Law, State Legislation and Federal Coordination · Installment V. Finds insurers closest to the deployment clock and names the risks carriers will not write.

• Agentic Foreseeability and the Modern Duty of Care for AI Agents · Installment IV. Sets the reassessment standard and the per-action insurance design.

• Foresight-to-Duty Conversion in AI Governance, Series Introduction. Sets the series question and the commissioning offer.

• Meta Foresight-to-Duty Conversion Across Privacy, Child Safety and Scam Ads. Shows why a finding must reach an authority that can fund the fix.

• Meta and Bank of America Foresight-to-Duty Conversion: Revenue-Funded Versus Cost-Funded Deferral. Examines why deferral persists once a gap is measured.

• Consent Order Remedies at Meta, Boeing, Starling and Wells Fargo: Verified Correction Versus Recurrence. Shows why disclosure-conditioned relief produces documentation rather than correction.

• The Foresight-to-Duty Conversion — Measured Governance Debt Becomes Evidence When Institutions Keep the Gap They Found. Separates latent from measured governance debt.

Foundational

• The Duty to Foresee, AI Deployment Readiness as Prospective Governance, and the Arrival of Agentic Duty of Care. Supplies the Agentic Hand Formula the series builds on.

• AI Governance Equilibrium. Supplies the review-capacity measures behind the authority question.

• Decision Modeling and Foresight Simulation. Describes the simulation method behind the forecasts.