The Reinsurance Treaty Is Where AI Risk Becomes Existential

Most of the conversation about AI risk in insurance right now sits in the regulatory column. The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers.¹ Colorado SB21-169.² New York DFS Circular Letter No. 7 (2024).³ EU AI Act penalties of up to seven percent of global turnover for prohibited practices.⁴ The rules are real, the enforcement appetite is growing, and compliance budgets are climbing across the industry.

The spending is going to the wrong threat.

Fines are a P&L event. Carriers absorb them, adjust pricing the next renewal cycle, and continue operating. The largest US insurance regulatory penalties of the past decade hurt the affected carriers without closing any of them.

Reinsurance loss is a balance sheet event. When a treaty walks, the next catastrophic loss is not ceded. It lands on surplus. In the wake of a single one-in-fifty cat year, surplus does not absorb that hit on its own.

This is the threat boards have not yet named. We think they should.

The Asymmetry CEOs Are Missing

Compare two scenarios for a regional carrier writing $400M in homeowners' premium.

In the regulatory scenario, a state insurance department fines the carrier under a NAIC-aligned bulletin for inadequate AI model documentation. The fine is large by industry standards: $5M, plus a consent order requiring quarterly reporting and a six-month outside review. Reputation takes a hit. The carrier raises homeowners rates 4% at the next filing. Twelve months later, the consent order is lifted and the company is whole.

In the reinsurance scenario, the same carrier's catastrophe excess-of-loss treaty is non-renewed at the next June 1 placement, citing a "change in underwriting practices and philosophy" the carrier did not adequately disclose. Premium for a replacement layer, if a replacement layer is even available at the price the carrier can afford, comes in at three times the prior cost. A Q3 hurricane lands. The retention attaches. Without the treaty above it, $480M in gross loss flows through to surplus that was not built to absorb it.

The first scenario costs five million dollars. The second one costs the company entirely. Boards are spending most of their AI-risk attention on the first.

The mechanics of the second scenario, which is more likely to actually end the carrier, are mostly absent from the conversation in board packets.

Why a Treaty Walks

Standard treaty wordings include a class of disclosure warranties that have been in the market for decades. IRMI catalogues them in its coverage of special termination provisions⁵ and underwriting and claims clauses⁶ in reinsurance agreements. The clauses typically grant the reinsurer rights to non-renew, cancel, or rescind based on:

• A change in the cedent's rating below a contractual threshold (commonly AM Best A- or S&P BBB+).⁷
• A material drop in surplus.
• A change of control.
• A change in the cedent's underwriting practices, rating philosophy, or claims-handling approach that the reinsurer was not formally advised of.

The fourth bullet is the one that interacts directly with AI. Disclosure warranties around changes in the cedent's underwriting and claims-handling approach sit in nearly every treaty in the market. Most carriers entered into them long before the cedent's underwriting model retrained itself weekly on new claims experience.

When a machine learning model is in the rating algorithm, every retraining cycle is potentially a change in underwriting philosophy. The risk weights shift. The book of business gets resorted into different rating cells. Two policies that the prior model would have priced identically now diverge. Whether that constitutes a disclosable change is a question almost no cedent has formally answered for its reinsurer, because the question almost never gets asked at a renewal meeting.

Reinsurers have started to notice. Munich Re sells aiSure,⁸ a product that underwrites the performance of a customer's AI model directly. Swiss Re signed a memorandum of understanding with RIQ in December 2025 to develop AI-powered risk transfer.⁹ The reinsurance market is building internal models for AI failure modes. Cedents using AI without disclosing the governance practices around it are buying coverage from a counterparty that is forming an opinion about their model whether they participate in the conversation or not.

What the Florida Failures Teach

The cleanest evidence that "treaty loss closes carriers" is not theoretical comes from the recent Florida market.

FedNat's catastrophe reinsurance program expired on June 30, 2022. Without it in place at the layers and limits the carrier's book required, the company was insolvent eighty-nine days later, on September 27.¹⁰ There was no fine in the chain. The treaty's absence alone was sufficient.

UPC (United Property and Casualty) saw Hurricane Ian losses escalate from an early estimate of $660M to $1.5B as the loss developed. The reinsurance program exhausted. The carrier was insolvent by February 6, 2023.¹¹ Again: no fine in the chain. Treaty exhaustion against an underwritten exposure that turned out to be larger than the model said it would be.

Merced Property and Casualty, founded in 1906, was AM Best A- rated when the Camp Fire ignited on November 8, 2018. The carrier was insolvent twenty-five days later, on December 3.¹² One hundred and twelve years of underwriting cleared in less than a month, because the cat exposure exceeded the reinsurance retention and surplus combined.

None of these failures was caused by AI. They are pre-AI examples. We cite them because they establish the mechanism that already exists. The question for the next decade is whether AI-driven misclassification produces enough rated-but-uncovered exposure to trigger the same outcome, faster, and at carriers nobody currently has on a watchlist.

The Causal Chain Boards Should Map

The chain runs from model retraining to insolvency in five steps. Each step is mechanically supported by existing law and treaty practice. None of them is hypothetical individually. The unique claim is that AI shortens the cycle and makes the disclosure question harder to answer.

1. The carrier's ML rating model retrains. The risk weights shift in ways the prior actuarial framework would have classified as a change in philosophy.
2. The reinsurer is not formally advised of the change, because the cedent's internal governance does not treat retraining as a treaty-disclosable event.
3. A cat loss develops larger than the new model priced for, because the model's training data had not yet absorbed climate-driven secondary peril intensification. Per Aon's 2023 climate report, secondary perils accounted for roughly 86% of global insured natural catastrophe losses in 2023.¹³
4. The reinsurer reviews the cession. It identifies the model change as a breach of the disclosure warranty. The treaty is rescinded, non-renewed, or significantly repriced. The Second Circuit's decision in Utica Mutual v. Clearwater, 906 F.3d 12 (2d Cir. 2018) is on point: the court declined to bind the reinsurer to follow the cedent's settlement absent the contractual basis for it.¹⁴ The cedent does not get the benefit of an automatic ride-through.
5. A subsequent cat loss attaches at the carrier's net retention. With the treaty layer above it gone or repriced beyond what the surplus can support, the loss flows through to capital. The carrier is insolvent.

To be precise: no published case has yet found a reinsurer rescinded a treaty for AI model drift specifically. The mechanics are in place even though the precedent has not been written. The honest framing is that the train is on the tracks. The wreck has not happened yet.

What This Means at the Board Level

The CEO question is short. If our AI drifts and our reinsurer walks, are we still a company?

Answering it requires three things most carriers do not currently produce:

• A model change log written in language a treaty underwriter can read, rather than a data scientist's commit history. The document should state what changed, when, and how the change affects the carrier's expected loss curve by line and territory.
• A treaty disclosure register that records what has been formally communicated to which reinsurer at which renewal. If the answer is "we communicated a model existed but not how it changes," the disclosure warranty is exposed.
• A solvency stress test that assumes the catastrophe XL treaty is rescinded, not just exhausted. Most carrier ORSA exercises model treaty exhaustion. Few model treaty rescission. The two are different events with different recovery paths.

These are not regulatory deliverables. No state insurance department is asking for them. Boards should be asking for them because the failure mode they protect against is the one that closes the company.

The conversation about AI risk in insurance is largely happening one floor below where the existential risk sits. Compliance, legal, and product are doing important work on the regulatory layer. The treaty layer is largely unsupervised. Reinsurers are forming opinions about the carrier's AI in the absence of cedent-led disclosure, and those opinions will eventually express themselves at a renewal.

That renewal is the moment the company finds out whether it has been managing the right risk.

Related

• The 89-Day Death Clock: What FedNat Teaches Every Carrier Building AI
• Merced County, Camp Fire, and What 112 Years of Underwriting Buys You
• The One Treaty Clause Your AI Strategy Lives or Dies On

Footnotes

1. "On December 4, 2023, in Orlando, FL, the National Association of Insurance Commissioners (NAIC) Membership voted to adopt the Model Bulletin on the Use of Artificial Intelligence Systems by Insurers." — NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers ↩

2. "On July 6, 2021, Governor Polis signed Senate Bill (SB) 21-169 into law, which protects Colorado consumers from insurance practices that result in unfair discrimination… in the application of artificial intelligence and big data within the insurance sector." — Colorado Division of Insurance: SB21-169 — Protecting Consumers from Unfair Discrimination in Insurance Practices ↩

3. "On Thursday, July 11, 2024, the New York Department of Financial Services ('NYDFS') issued Circular No. 7 Re: Use of Artificial Intelligence Systems and External Consumer Data and Information Sources in Insurance Underwriting and Pricing." — Locke Lord QuickStudy on NY DFS Circular Letter No. 7 ↩

4. "Non-compliance with the prohibition of the AI practices referred to in Article 5 shall be subject to administrative fines of up to 35 000 000 EUR or, if the offender is an undertaking, up to 7% of its total worldwide annual turnover for the preceding financial year, whichever is higher." — EU AI Act, Article 99 (Penalties) ↩

5. "Special termination clauses allow early contract exit when 'certain contingencies arise.' They address concerns about solvency, management, and reputation changes that may negatively affect the reinsurance relationship." — IRMI: Special Termination Provisions in Reinsurance Contracts ↩

6. "If there is any change in the Company's approach, method or guidelines in the processing, settling, administering or paying of claims, the Reinsurer shall be entitled to an adjustment of the portion of the claims which is reimbursable or an adjustment to Premium." — IRMI: Underwriting and Claims Clauses in Reinsurance Agreements ↩

7. "A.M. Best's financial strength rating has been suspended or withdrawn or downgraded below 'A-'… S&P Global Ratings…downgraded below 'BBB+'… Moody's Syndicate Performance Rating…downgraded below 'B+'." — IRMI: Special Termination Provisions in Reinsurance Contracts ↩

8. "Munich Re offers aiSure™, a performance-guarantee insurance to cover the performance of AI solutions and protect against losses caused by inadequate or unreliable AI solutions." — Munich Re: aiSure — More AI Opportunity. Less AI Risk ↩

9. "On December 10, 2025, Swiss Re and RIQ signed a Memorandum of Understanding (MoU) to develop and scale innovative capacity solutions, risk origination opportunities and AI-enabled capabilities." — Swiss Re Press Release: Swiss Re and RIQ partner to advance risk transfer powered by data and AI ↩

10. "On September 27, 2022, FedNat Insurance Company ('FedNat') was ordered into receivership for purposes of liquidation by the Second Judicial Circuit Court in Leon County, Florida." — Florida Department of Financial Services: FedNat Insurance Company ↩

11. "United was deemed insolvent on February 6, 2023… UPC was heavily concentrated in the Southwest Florida region and received approximately 25,000 claims from Hurricane Ian with a gross estimated loss of over $1 billion… UPC's outside actuaries determined that its losses would exceed the prior estimate of gross losses, which resulted in UPC exceeding its catastrophe reinsurance coverage." — Insurance Journal: Orderly Runoff Didn't Work; Florida's United P&C Now Insolvent, Headed for Liquidation ↩

12. "On November 30, 2018, California Insurance Commissioner Dave Jones filed a petition to place Merced Property and Casualty Company into liquidation. On December 3, 2018, the Superior Court of the State of California, County of Merced [issued the] Order Appointing Liquidator… The company was formed in 1906 by a group of farmers in Merced County… Still, the company carried a solid A-minus rating from insurance consultant A.M. Best before the fire." — CDI Press Release release141-18 and United Policyholders coverage ↩

13. "Aon pegged insured natural catastrophe losses at $118bn for 2023, with perils traditionally considered as primary or peak — tropical cyclone, earthquake and European windstorm — accounting for just 14 percent of the total." — The Insurer: Aon estimates 2023 insured nat cats at $118bn as protection gap rises to 69% ↩

14. "On the follow-the-settlements issue, the court vacated and remanded the cross-appeal because Utica failed to demonstrate its entitlement to a judgment that Clearwater was bound to indemnify Utica according to Utica's settlement with its insured." — Utica Mutual Insurance Co. v. Clearwater Insurance Co., 906 F.3d 12 (2d Cir. 2018) — Justia ↩