AI deployers · Foundation model developers · Enterprise customers · Integration partners · Lenders · Property and casualty insurers · Cloud platforms · Verisk · State insurance commissioners · California · Illinois · Washington · State attorneys general
Companion to Who Absorbs Correlated AI Loss — Reinsurance Aggregation, Industry Pools, the Public Backstop and the Liability Shield · Agentic Foreseeability and the Modern Duty of Care for AI Agents · Who Sets the Agentic Standard of Care — Common Law, State Legislation and Federal Coordination
Foresight-to-Duty Conversion Series Introduction: Meta Foresight-to-Duty Conversion Across Privacy, Child Safety and Scam Ads · Meta and Bank of America Foresight-to-Duty Conversion: Revenue-Funded Versus Cost-Funded Deferral · Consent Order Remedies at Meta, Boeing, Starling and Wells Fargo: Verified Correction Versus Recurrence
Executive Summary
Insurers can change deployment conditions at renewal while legislation and appellate decisions remain pending, and in 2026 the market split. In January 2026 Verisk released optional endorsements that remove generative AI from commercial general liability, more than sixty carriers filed their own exclusions and state commissioners approved more than eighty percent of the requests. In the same year specialty underwriters wrote affirmative cover for defined AI failures, a cloud platform began sharing customers’ security posture with insurers and a Fortune 100 buyer wrote AI-specific insurance into its vendor agreement. The fastest institution in the series now covers some AI risks on conditions and leaves the rest with firms and the people they harm.
Thesis: Insurance can convert risk assessments into enforceable deployment conditions, but only where someone supplies the evidence a carrier needs. Insurability has two bottlenecks: credible evidence about controllable losses and sufficient capital for correlated losses. The present paper takes the first. Verified controls and contractual demand from buyers, attorneys general and partners address it, and the correlated-loss paper takes the second.
What distinguishes the MindCast approach: Who Sets the Agentic Standard of Care? found that insurers run closest to the deployment clock and that some agent risks exceed what any carrier will write. The paper treats the coverage condition as a legal instrument and the exclusion as a regulatory act. The paper measures prevention and compensation as separate outcomes, and it follows pressure from customers and partners upstream to the providers whose commitments insurers can cover. One simulated result carries the argument: providers accept liability above the fee cap in two cases of three once a partner and a customer both require it, and in almost none when neither does.
Game mutation is the part of the simulation that matters most here. The rules governing every player change the moment a buyer’s contract supplies a template, because an attorney general who had to draft a term from nothing can then copy one into a decree. The rules change again when a distribution partner conditions access on cover rather than safeguards, because a provider that could keep its fee cap at no cost then loses listings by keeping it. Each mutation moves the coverage condition from a term one buyer negotiated to a term the market assumes.
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.
What the paper does: The paper maps the 2026 split between exclusion and affirmative cover and the evidence each affirmative product underwrites on. The paper separates four loss categories and shows which two the evidence bottleneck governs. The paper then identifies five routes that create demand for conditions, including customers and partners. The paper closes with the coverage condition measured on two axes and with competing explanations, four Simulation Predictions and a Risk Mitigation Layer.
Simulation Prediction highlights: The forecast set holds four MindCast Foresight Simulation Predictions, two primary and two secondary, stated in full in the Simulation Predictions section. The nearest-term test is a state attorney general decree or assurance that requires an independent audit or an insurance term by the end of 2027 (17-62%). A major provider accepting liability above twelve months of fees for output claims beyond intellectual property by the same date is the test of the upstream chain (2-53%). Given a lender class publishing an AI insurance or audit covenant, a state enacts financial responsibility for AI deployers by the end of 2028 (48-95% given the covenant).
🏛️ Policymakers: A mandate to carry AI insurance binds only where a carrier will write the policy. The decree route is open now and needs no statute, and a statute can adopt the terms buyers already ask vendors to meet.
💼 Executives: Read the endorsement schedule at every renewal, and read your largest customers’ vendor agreements. The exclusion that removed AI from general liability arrived as an optional form, and the buyer that now requires AI-specific cover arrived in the master services agreement.
⚖️ Counsel: Most AI vendor agreements cap total liability at twelve months of fees, so residual loss sits with the deployer unless the contract says otherwise. A coverage condition that specifies the required response to a finding is the nearest thing to a reassessment duty a deployer can buy.
📊 Investors: A portfolio company whose general liability policy excluded AI in 2026 retains that loss. Ask for the endorsement schedule and the retained-loss scenario before the liability map.
I Insurers Change Conditions at Renewal and the Market Splits
Insurers set terms at every renewal, and a renewal arrives faster than any statute or ruling. A deployer whose policy renews in March meets an AI exclusion or an audit condition before a 2027 bill takes effect and years before an appellate court speaks. The deployment clock runs from each permission change, new evidence or incident at the deployment itself, and the institutions paper placed insurers closest to it for that reason.
The term most deployers met in 2026 was an exclusion. An exclusion removes coverage and says nothing about what conduct would restore it, so it shifts losses without specifying corrective conduct. The Center for Strategic and International Studies called insurance “the most important de facto regulator of U.S. AI deployment” in September 2026 and said it was regulating by withdrawal rather than by standard.
Withdrawal was one response among several. A June 2026 survey of 26 public AI insurance products found the market bifurcating: affirmative products expanding by peril while exclusions narrow legacy forms. Of the cells in its 55-peril matrix with a stated position, 53 percent were affirmative and 20 percent excluded. The paper therefore asks two questions: which exposures the affirmative conditions reach, and who holds the exposures they do not.
⚖️ Counsel: An exclusion is a term of the policy, not a judgment about the law, so it tells a deployer nothing about whether the duty applies.
Takeaway: Insurers change conditions at renewal while statutes and rulings remain pending, and the 2026 renewal cycle split the market between carriers that excluded AI and carriers that wrote it by peril on conditions.
II The 2026 Narrowing Arrived by Form and by Carrier
Verisk’s Insurance Services Office released three optional endorsements with a January 2026 edition date. Form CG 40 47 removes bodily injury, property damage and personal and advertising injury arising from generative AI from the commercial general liability form. Form CG 40 48 removes advertising injury only, and CG 35 08 applies the exclusion to products and completed operations.
Carriers moved on their own as well. More than sixty property and casualty insurers filed AI exclusions in 2026, including Berkshire Hathaway and Chubb. State commissioners approved more than eighty percent of the requests, and an Aon matrix of 112 coverage cells shows 90 excluded against 13 available.
The forms are optional, which creates the gap brokers now warn about. An optional form changes an insured’s position only when it is incorporated into the issued policy, so one client can hold an exclusion with one carrier and none with another. Brokers advise clients to map where AI touches customer communications, contracts and professional advice and to show what human review occurs before output is used.
Behavioral economics explains the speed. Underwriters face a loss they cannot measure and a reference point set by the last renewal, so the available move that preserves the loss ratio is to exclude and wait. Game theory explains the uniformity: once a standard form exists, no carrier gains by being the one that keeps writing the risk its competitors dropped.
💼 Executives: Map every AI touchpoint and the human review behind it before the next renewal, because that map is what a carrier now asks for before it writes anything.
Takeaway: The narrowing came through standard forms and individual filings at once, and the optional structure means no deployer knows its position without reading the issued policy.
III The Affirmative Market Writes by Peril and Underwrites on Different Evidence
Affirmative AI cover now arrives through at least four distinct models, and they are not interchangeable capacity. The first is the standalone specialty policy. Armilla’s Lloyd’s-backed policy writes up to $25 million per organization against model errors and agent failures such as wrong decisions and tool misuse, 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; applicants identify the systems to be covered.
The second model is the performance guarantee. Munich Re’s aiSure launched through Mosaic in February 2026 and was unveiled by Munich Re’s chief executive on July 8. The cover pays a pre-agreed amount when a model’s accuracy falls below a threshold set at binding or when a named regulatory finding issues. Munich Re’s stated limits and the question of capacity belong to the correlated-loss paper.
The third model is clarified wording inside existing policies. CFC added explicit AI language across seven coverages in June 2026 and extended the approach to healthcare and professional services on October 27. The wording names hallucination, AI-generated content and model drift. Clarification removes a coverage dispute before it starts, adds no capital and no limit, and counts as a separate market response from new capacity.
The fourth model is platform-distributed cover. Google Cloud’s Risk Protection Program shares a customer’s security posture from its own configuration with Munich Re, Beazley and Chubb in place of a manual application, and since May 2025 includes affirmative AI cover for Google-related AI workloads.
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 platform re-rates an agent when its model, tools or data change and insures the residual with Harbor’s own capital. Neither Google’s program nor Harbor’s has disclosed limits.
The products differ in what evidence they underwrite on and in whether measurement continues after binding. Armilla’s certification, Munich Re’s performance thresholds and Harbor’s re-rating on change all continue after the policy attaches. Testudo’s scheduled systems and Google’s posture sharing describe how cover is obtained, and neither public description establishes an ongoing reassessment obligation in the issued policy.
None of the products binds cover separately before each agent action. Munich Re’s per-output premium varies the charge with exposure, and Harbor’s blocking of individual actions is a control rather than a policy that attaches per action. The per-action design from Installment IV remains a design.
📊 Investors: A company that holds affirmative AI cover has passed some form of assessment, and which form tells you what the carrier measured. Limit and sublimit are separate questions from the existence of cover.
Takeaway: Affirmative cover arrives through standalone policies, performance guarantees, clarified wording and platform programs. Each underwrites on different evidence, and the evidence is what the demand routes in the next sections supply.
IV Four Loss Categories and the Evidence Bottleneck
AI loss falls into four categories that insurance treats differently. The first is ordinary deployment loss: an identifiable error within a specified system and within policy limits, such as a hallucinated report, a misrouted payment or an agent that calls the wrong tool. Carriers write that category today, with conditions, because frequency is observable at the insured and the maximum loss is bounded by the deployment.
The second is the coverage dispute. A loss may sit inside a policy and still go unpaid because an exclusion applies, notice came late, defense costs eroded the limit or two policies each point at the other. CFC’s wording changes address this category directly, and brokers’ advice to map AI touchpoints against every policy form is a dispute-avoidance exercise. Coverage disputes are a legal problem rather than a capacity problem, and the simulation treats them as a separate outcome.
The third is correlated loss from one widely deployed model, where one defect can surface at many insureds at once and accumulate beyond available coverage. Gallagher Re’s March 2026 white paper said such failures could cause linked losses across sectors and accumulate across ceded portfolios in ways existing models cannot capture. The June survey found foundation-model concentration fails two insurability criteria, loss independence and economic feasibility, and treats maximum possible loss as effectively unbounded because the base of insureds using one model keeps growing.
The fourth is the compensation gap: losses above limits or outside any coverage, which fall on the firm and the injured party. The gap can arise from any of the first three categories, and it is the category public policy must decide about. A public layer is a hypothesis for the third and fourth categories only, and nothing in the research establishes that one is inevitable.
Two bottlenecks run across the categories. The first is evidence: credible information about which losses a deployer can control. Three underwriting paths now supply it: technical verification by certification or pre-bind validation, litigation-based assessment of the kind Testudo builds from lawsuit data, and platform-supplied operational evidence of the kind Google Cloud shares from a customer’s own configuration. A shared incident database is a fourth path and a prerequisite for none of the others.
The second bottleneck is capital to absorb a correlated loss, and the correlated-loss paper carries it.
🏛️ Policymakers: Separate the evidence problem from the capital problem before choosing an instrument, because an audit mandate addresses the first and only a pool or public layer addresses the second.
Takeaway: Ordinary losses are insurable on conditions, coverage disputes are a wording problem, correlated losses can accumulate beyond available coverage and the compensation gap is where public policy decides.
V Statute, Procurement and Enforcement Create Demand for Conditions
Legislation is one route, and no state has used it for insurance. A September 2026 review of 69 AI liability bills across 27 states found none with an insurance mandate.
States have used legislation for verification. California’s SB 1119 of September 10, 2026 requires operators of companion chatbots to obtain an independent child-safety audit every two years from 2029 and after higher-risk modifications. Illinois’s frontier-model law of July 6, 2026 requires annual audits and 72-hour incident reporting from large developers. Neither carries an insurance term.
Procurement is the second route, and it is in force in at least one reported case. Armilla sells the cover such clauses require and reports an unnamed Fortune 100 vendor agreement reading: “Provider shall obtain and maintain, at its own expense, insurance coverage for AI-related risks.” The reported clause names bias, hallucination, copyright and regulatory penalties and sets limits “in amounts satisfactory to [Customer].”
Armilla reports that the same agreement requires annual compliance testing, prior approval for AI deployments or updates and two years of audit rights. Armilla also reports that buyers in media, human resources technology and AI services ask for evidence of cover before signing. One reported agreement establishes the mechanism; it does not establish that buyers broadly obtain the protection they request.
Enforcement is the third route, and it has reached the threshold without crossing it.
On October 14, 2026 a bipartisan coalition of 39 state attorneys general filed notice in the Character.AI bankruptcy of the terms they will press on any buyer. The terms include age verification, crisis-response measures that a click-through cannot bypass and a privacy program with risk assessments and limits on training on chat content. The notice contains no audit or insurance term. The Washington Attorney General’s office has said existing consumer-protection law reaches AI products, so a decree carrying such a term needs no new statute.
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.
🏛️ Policymakers: The procurement route shows buyers asking vendors for testing, update approval and insurance, so a statute can adopt terms the market is already being asked to meet.
Takeaway: Statute, procurement and enforcement can each create demand for conditions. Procurement is reported to require cover and statute already requires audits, while no decree yet carries either.
VI Customers and Partners Move Pressure Upstream
Enterprise procurement is one customer. Individual users, the partners that distribute an agent and the providers that supply its model and tools each decide something that changes who pays, and the institutions paper’s five institutions sit downstream of all of them. Pressure runs from users to enterprise customers and attorneys general and through distribution partners to providers. The chain ends at the insurer that decides which provider commitments it will cover.
Two links of that chain already exist in published terms. Microsoft’s commercial marketplace requires that agents listed for Microsoft 365 Copilot block harmful AI content and give users a way to report it. Listed agents must also act on reports, meet transport-security and domain standards and supply perpetual test environments as conditions of listing.
OpenAI’s service terms indemnify API and Enterprise customers against third-party claims that their use or distribution of Output infringes intellectual property, subject to listed exclusions. OpenAI’s Services Agreement excludes indemnification obligations from its liability cap. A distribution partner has written named safeguards into access, and a provider has accepted one class of output liability above the cap. Both terms predate the forecast set and form the baseline it measures from.
The provider’s liability cap is the chain’s hinge. 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 responsibility above it in two cases of three when both require it. The cap moves when someone with access or purchasing power asks.
The chain also runs in reverse. A provider that changes a model or withdraws a safeguard without disclosure can leave the deployer exposed to customer claims despite limited operational control. Partners that require disclosure of material changes as an access condition close that path, and in the simulation providers disclose in nearly every case once a partner requires it. Where no partner requires it, the deployer with the least control holds the loss.
Lenders sit beside this chain rather than in it. A lender can write AI cover or audit evidence into credit terms the way it writes insurance covenants for property, and in the simulation lenders do so in about half the cases once customer requirements exist. No lender class has published such a covenant, so the question stays an open hypothesis.
💼 Executives: Read the partner’s listing terms and the provider’s cap together, because the first tells you what safeguards the market already demands and the second tells you what loss stays with you when they fail.
Takeaway: Customers, partners and providers form a chain that can carry requirements upstream to the insurer. Two of its links already exist in published terms, and the provider’s cap moves when a partner or customer asks.
VII The Coverage Condition, Measured on Two Axes
The institutions paper proposed a duty to reassess a deployment after each material permission change, new evidence or incident. Insurers can impose that duty through a coverage condition, and the affirmative market already does in fragments. Armilla attaches certification, Munich Re validates before binding, Harbor re-rates on change and the AIUC-1 standard requires quarterly testing.
A condition that produces correction needs four elements. The first is a material-change trigger rather than a named statute or test, so the condition survives the rule changes the institutions paper forecasts. The second is a required response: when a test or incident surfaces a finding, the insured must take a defined action within a defined time, such as restricting the agent’s permissions or suspending the affected function.
The third element is evidence that the control worked, in a form the carrier can verify at renewal or after a claim. The fourth is a stated consequence for noncompliance. A condition can reduce risk whatever the consequence, because the insured who wants to keep cover changes behavior before any breach, so prevention is measured on its own axis.
Compensation is the second axis. 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 after a breach. Compulsory auto insurance resolves that tension by preserving third-party payment and letting the carrier recover from the insured afterward. An AI coverage condition can adopt the same split, and the paper’s forecasts test designs that reduce risk and preserve payment together rather than assuming one implies the other.
Installment IV sketched a per-action design in which cover attaches to a specific agent action for a specific window. Harbor’s product rates and blocks at the action level, which shows that action-level control can be built. Whether cover can attach per action, and whether such cover attracts capacity beyond a startup’s own capital, remains open.
💼 Executives: Specify in advance what the company does when a test surfaces a finding, because the next affirmative policy and the next vendor agreement will both ask for that procedure and the response time.
Takeaway: A coverage condition carries the reassessment duty when it names a trigger, a required response, evidence of control and a consequence. The effect on prevention and the effect on compensation must be read separately.
VIII Three Explanations Compete for Who Writes the Condition
Three readings compete for who writes the first binding coverage condition. Each has its own falsifier, and the Simulation Predictions test them. The correlated-loss paper carries the four readings that concern capacity and the public layer.
Hypothesis A · Demand from procurement and enforcement: Buyers and attorneys general create the first binding insurance terms for deployment classes, and carriers write against them before any statute requires cover. Falsifier: a state statute requiring AI-specific cover for a deployment class is enacted before any decree or published buyer standard imposes one.
Hypothesis B · Compulsory financial responsibility: A state requires deployers above a threshold to show cover, a bond or qualifying assets. Falsifier: no state enacts such a requirement by 2028.
Hypothesis C · Upstream transmission: Customers and partners carry requirements to providers, providers accept liability above the twelve-months-fees cap, and insurers cover the accepted commitments. Falsifier: through 2028 no major provider’s standard terms accept liability for output claims beyond intellectual property, or no carrier writes cover against a provider’s accepted commitment.
The forecasts rest on Hypothesis A, with C as its mechanism. Hypothesis B depends on A producing a market to buy from, and the simulation finds a lender covenant is the condition under which a state mandate becomes the expected outcome rather than a minority one.
📊 Investors: Under Hypothesis A the first binding conditions arrive in contracts and decrees rather than statutes, so a portfolio company’s exposure tracks its customers’ procurement terms before any legislature acts.
Takeaway: The evidence favors demand from customers, partners and enforcers with a market that follows it, and the forecasts test whether a provider’s cap moves before a statute requires cover.
IX MindCast Simulation Predictions
The forecast set holds four Simulation Predictions: two Primary Simulation Predictions (P) and two Secondary Simulation Predictions (S). The secondary entries are conditional and print as P(A given B). Each entry names an actor, an outcome, a window and a falsifier. Each resolves against a named source.
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%)
Mechanism: once a buyer’s contract supplies a template, the attorney general attaches the term in most simulated cases. Falsifier: no such decree or assurance by the deadline. Source: attorney general press releases and court filings.
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 sits outside its cap and forms the baseline the entry measures from; it does not resolve the entry. Mechanism: the cap moves when a partner or customer requires it. Falsifier: no such published standard terms by the deadline. Source: published provider terms.
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 the institutions paper. Falsifier: neither the enacted text nor rules adopted by the deadline contain such a reference. Source: Washington session laws and administrative code.
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)
A covenant is a published credit term from a bank or a lender trade body that requires AI-specific cover or audit evidence from borrowers whose revenue depends on a deployed agent. Financial responsibility means cover, a bond or qualifying assets above a deployment threshold. No covenant leaves the entry unresolved. Mechanism: a lender requirement gives legislators a private standard to codify and removes the objection that no market exists to buy from. Falsifier: a qualifying covenant is published and no state enacts the requirement by the deadline. Source: lender publications and state session laws.
Validated outcomes to date: none. No entry has reached its deadline, so the forecast set shows no hit and no miss.
Two questions remain open hypotheses without bands. The first is whether a distribution partner requires AI-specific cover or an accepted liability commitment, rather than named safeguards, as a condition of listing by the end of 2027. The second is whether a lender class writes such a covenant at all, since S-2 takes the covenant as given.
Takeaway: The decree is the nearest-term test and the provider’s cap is the test of the upstream chain. A lender covenant is the trigger that turns financial responsibility from a minority route into the expected one.
X Risk Mitigation Layer
Each entry states exposure in a unit the stakeholder controls, with severity and probability as separate axes and the band as the probability. Actions are unilateral, with an owner and a deadline tied to the forecast checkpoints. The layer sits as a standalone section so that each audience reads its own entries together; the trade-off is that a reader of one Simulation Prediction must turn here for its actions. Entries for developers, carriers and federal staff sit in the correlated-loss paper.
Multistate deployers · P-2 (2-53%): Severity: high. Exposure: AI losses retained without a reserve, measured in quarters of runway a retained-loss scenario would consume.
Action 1: read the endorsement schedule on every liability policy and list each AI exclusion by form number. Owner: risk manager. Deadline: the next renewal.
Action 2: model three retained-loss scenarios for the largest deployed agent, from a single-customer error to a correlated failure across the customer base, and size a reserve against the scenario the company can survive. Owner: chief financial officer. Deadline: the next audited statement.
Action 3: renegotiate the twelve-month-fees cap with the model vendor for third-party claims arising from outputs, or document the retained exposure the cap leaves. Owner: procurement counsel. Deadline: the next vendor renewal.
Residual: correlated losses above the company’s capacity remain with the company and its claimants, which the correlated-loss paper addresses.
Enterprise customers · P-1 (17-62%), P-2 (2-53%): Severity: moderate. Exposure: vendor AI losses passed through the contract to the buyer, measured in the gap between the vendor’s cap and the buyer’s loss.
Action 1: require evidence of AI-specific cover, not a certificate of general liability, and name the perils it must cover. Owner: procurement. Deadline: the next master services agreement.
Action 2: carve third-party claims arising from AI outputs out of the consequential-damages waiver and the twelve-month cap. Owner: commercial counsel. Deadline: same.
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 · P-2 (2-53%); partner cover requirement, open hypothesis without a band:Severity: moderate. Exposure: a provider failure transmitted through the partner’s channel to every business that embeds the agent, measured in integrations exposed per undisclosed change.
Action 1: add disclosure of material model, tool and data changes to listing conditions, with a notice period before the change reaches embedded agents. Owner: platform policy lead. Deadline: the next policy revision.
Action 2: require evidence of AI-specific cover or an accepted liability commitment from providers above a usage threshold, with a safeguard alternative for smaller providers. Owner: partner program lead. Deadline: the 2027 program cycle.
Residual: a provider that leaves the program rather than disclose moves the exposure to the businesses that keep its agent through other channels.
Lenders · S-2 (48-95% given the covenant): Severity: moderate. Exposure: a borrower’s retained AI loss that reaches the lender as a covenant breach or impaired collateral, measured in exposure at default across AI-dependent borrowers.
Action 1: add AI exclusions to the insurance covenant review at origination and renewal, so an excluded loss is visible before it is retained. Owner: credit risk. Deadline: the next policy update.
Action 2: for borrowers whose revenue depends on a deployed agent, require either affirmative cover or a retained-loss reserve as a covenant. Owner: credit committee. Deadline: the next facility renewal.
Residual: covenants discover a retained loss; they do not pay it.
State legislators and staff · S-1 (40-86% given the statute), S-2 (48-95% given the covenant): Severity: moderate. Exposure: a liability statute that expands exposure into an excluded market, measured in sessions before an insurance provision is added.
Action 1: pair any AI liability bill with an incident-reporting duty whose output carriers can use, and allow a bond or asset alternative to any cover requirement. Owner: committee counsel. Deadline: before policy-committee action.
Action 2: direct the insurance commissioner to report on AI exclusion filings and available capacity by product and limit. Owner: committee chair. Deadline: the same session.
Residual: capacity depends on carriers outside the state’s jurisdiction.
State attorneys general · P-1 (17-62%): Severity: moderate. Exposure: a decree template set by another state or by a bankruptcy sale on terms the office did not write, measured in matters resolved before the office has a term of its own.
Action 1: draft a standard audit-and-cover term for AI consumer matters, with a named recipient inside the company who can halt the product, before the next AI assurance is negotiated. Owner: consumer protection division chief. Deadline: the first 2027 AI matter.
Action 2: coordinate the term through the existing multistate coalition so that one decree becomes the template. Owner: multistate coordinator. Deadline: the next coalition action.
Residual: a term binds one company; it does not reach the deployers that never draw a complaint.
Takeaway: Every action builds evidence or conditions before a loss would otherwise build them on worse terms.
Conclusion
Insurers were the fastest institution in the series, and in 2026 the market split. Most carriers narrowed cover while affirmative cover arrived through standalone policies, performance guarantees, clarified wording and platform programs. A Fortune 100 buyer wrote cover into its vendor agreement, a marketplace wrote safeguards into listing and a provider accepted one class of output liability. The next links, cover in listing terms and output liability beyond intellectual property, are the ones the forecasts test.
The conditions are the series’ mechanism. Each one carries the reassessment duty onto the deployment clock when it names a trigger and a required response, with evidence of control and a consequence. Prevention and compensation remain separate axes, and a condition that voids cover on breach can prevent well and compensate badly.
The conditions reach only the losses evidence can describe. Ordinary deployment losses are insurable and coverage disputes are a wording question, while correlated losses from one model can accumulate beyond available coverage. The correlated-loss paper asks who absorbs that remainder.
The series now runs from measurement through institutions to evidence. Installments I through III showed what happens when an institution knows and defers, Installment IV set what a deployer should foresee and Installment V asked which arrangement turns foresight into correction. Installment VI asks who writes the condition that carries it, and the answer today is whoever holds purchasing power or access.
What to Watch
The first checkpoint falls on January 1, 2027, when Illinois’s disclosure duties take effect. The Washington session opens on January 11, and state legislatures can introduce financial-responsibility or database bills through spring 2027. The multistate coalition’s next AI matter is the first place a decree term can appear.
Each signal maps to an entry. An attorney general decree with an audit or insurance term resolves P-1. Provider terms accepting output liability beyond intellectual property above the cap resolve P-2, and Washington’s enacted text or rules resolve S-1. A published lender covenant activates S-2, which a state financial-responsibility statute then resolves.
Working With MindCast
Published analysis can describe the demand routes in general. Published analysis cannot tell a specific deployer what its exclusions leave retained, tell a buyer which cover term a vendor can obtain or tell an attorney general which decree term the market will copy. The series introduction sets out where the answers sit: policy schedules and vendor agreements, partner terms and evaluation files, and the rules each deployment meets.
MindCast builds Cognitive Digital Twins (CDTs) of carriers, of developers and their customers and partners, and of enforcers. The engagement simulates competing coverage conditions against the same deployment and harm pathways.
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 attorneys general, it tests a decree term against the carriers’ and providers’ likely response before the term 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 mcai@mindcast-ai.com or visit www.mindcast-ai-simulation.com to schedule the diagnostic session.
Sources
MindCast Works
Who Sets the Agentic Standard of Care? Common Law, State Legislation and Federal Coordination. 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. 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.
Who Absorbs Correlated AI Loss? Reinsurance Aggregation, Industry Pools, the Public Backstop and the Liability Shield. Carries the capital bottleneck and the public-layer forecasts.
Primary Sources
California’s Adam’s Law Mandates Independent AI Chatbot Audits, Cloud Security Alliance (September 2026). Describes SB 1119’s audit cycle and reporting to the Attorney General.
Verisk to Roll Out New General Liability Exclusions for Generative AI Exposures, Independent Insurance Agents and Brokers of America (October 2025). Describes forms CG 40 47, CG 40 48 and CG 35 08.
Armilla AI Raises Lloyd’s-Backed Coverage to $25M as Traditional Insurers Retreat from AI Risk, Armilla (2026). Carrier statement of limit, covered failures and certification condition.
Microsoft Commercial Marketplace Certification Policies, Microsoft Learn (page updated October 27, 2026). Agent listing conditions: harmful-content blocking, user reporting, security standards and test environments.
OpenAI Service Terms, OpenAI (checked October 29, 2026). Indemnity for third-party intellectual-property claims arising from Output for API and Enterprise customers.
OpenAI Services Agreement, OpenAI (checked October 29, 2026). Excludes indemnification obligations from the liability cap.
What’s New With Google Cloud’s Risk Protection Program, Google Cloud (May 15, 2025). Describes posture sharing with Munich Re, Beazley and Chubb and affirmative AI cover for Google-related workloads.
Munich Re Insure AI: Frequently Asked Questions, Munich Re. Carrier statements on trigger, limits and premium basis.
Covering AI: AI Insurance Becomes a Procurement Requirement, Armilla (May 15, 2026). Quotes the Fortune 100 vendor agreement clause and reports buyer requirements by sector.
Secondary Sources
Illinois Enacts AI Safety Law, Becoming First State to Mandate Independent Third-Party Audits of Frontier AI Developers, Skadden (July 2026). Describes S.B. 315’s audit and incident-reporting duties.
State AGs Set Child-Safety and Privacy Expectations for Any Sale of Character.AI, Troutman Pepper Locke (October 27, 2026). Reports the 39-state notice of October 14, 2026.
AI Insurance Gap: What It Means for Technology Contracts, Honigman (May 27, 2026). Reports the twelve-month-fees cap and recommends carve-outs.
Gregory C. Allen, The Insurance Industry’s Retreat from AI Threatens to Slow Innovation and Adoption, Center for Strategic and International Studies (September 4, 2026). Supplies the exclusion figures, the insurability analysis and the database and backstop recommendations.
Alex Leung, Rex Zhang, Ervin Ling, Kentaroh Toyoda and SiewMei Loh, The Insurability Frontier of AI Risk: Mapping Threats to Affirmative Coverage, Silent Exposures, and Exclusions, AIFT (June 2026). Supplies the 26-product survey, the four-tier frontier and the ordered candidate structures for correlated loss.
CFC Unveils Updates to Embed Affirmative AI Cover, Business Insurance (June 26, 2026). Reports explicit AI wording across seven existing coverages.
AI Exclusions Are Spreading. Harbor Just Raised $10 Million to Insure the Gap, Insurance Business (October 19, 2026). Reports Harbor’s agent-level rating and re-rating on change.
AI Liability Insurance Market Map 2026: Every Carrier, Product and Capacity, Agent Insured (August 2026). Reports specialty products and limits as stated by the desk, unverified with carriers.
AI Exclusions Are Spreading Through Commercial Policies, Brokers Warn Clients to Check Cover, Insurance Business (October 21, 2026). Reports broker guidance on optional endorsements.
Why No State Requires AI Liability Insurance Yet, CASRAI (September 2026). Reviews 69 bills across 27 states.
