The One Treaty Clause Your AI Strategy Lives or Dies On

Standard reinsurance treaty wordings include disclosure warranties triggered by changes in the cedent's underwriting practices, rating philosophy, or claims-handling approach.¹ The clauses are older than most of the boards reading them. They have not been amended for the arrival of machine learning in pricing models. Most carriers have never formally asked their CRO whether weekly ML model retraining counts as a disclosable change inside one of those clauses.

Their reinsurer will eventually answer the question for them. That answer is most likely to arrive at a renewal meeting, in the form of a non-renewal notice, a repricing, or a quiet conversation about scope.

This is the clause family your AI strategy lives or dies on. We think most boards have not yet read it through that lens.

What the Clause Actually Does

Standard reinsurance treaty wordings include a family of disclosure and termination warranties. IRMI catalogues them in its coverage of special termination provisions² and its companion piece on underwriting and claims clauses in reinsurance agreements.³ The warranties typically grant the reinsurer rights to non-renew, cancel, or rescind based on:

• A change in the cedent's financial condition or rating below a contractual threshold (commonly AM Best A- or S&P BBB+).⁴
• A change of control.
• A change in the cedent's underwriting practices, rating philosophy, or claims-handling approach.

The third bullet is the live one. It rests on the doctrine that a reinsurer underwrites the cedent, not the underlying policies. The reinsurer's price assumes the cedent will keep doing what it was doing at the time of placement. If the cedent changes how it selects and prices risk, the reinsurer's exposure changes, and the contract gives the reinsurer remedies. IRMI's underwriting-and-claims article includes verbatim language for the claims-handling variant: "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."³

For most of the clause's history, the changes that triggered it were obvious. A new line of business. A move into a new state. A pricing committee that loosened guidelines. Auditable, dated, signed by a chief underwriting officer. The clause assumed underwriting changes would be slow and visible.

That assumption no longer holds when a machine learning model is in the rating algorithm.

What Counts as a Change in Philosophy

When an ML rating model retrains, the risk weights shift. The book of business gets resorted into different rating cells. Two policies the prior model would have priced identically now diverge. The cedent's actual selection behavior on the next thousand bound risks is materially different from what it was on the last thousand, even though no human committee voted to change anything.

Whether that constitutes a change in underwriting practices or philosophy under one of these warranties is a question almost no cedent has formally answered for its reinsurer. We have not seen a major US carrier publish a position. We have not seen a treaty wording amended to address it. The question almost never gets asked at a renewal meeting because the renewal meeting covers loss ratios and exposure shifts, not the governance of the model that produced the loss ratios.

There is a concept worth naming here. Treaty disclosure was designed for episodic underwriting changes: discrete, dated, attributable to a person. Modern ML pricing produces continuous underwriting changes: every retraining cycle is a small, possibly unattributable shift in selection behavior. The clause family was never tuned for that frequency or that opacity. Cedents are operating inside a contractual regime that assumes a tempo their technology no longer matches.

A reinsurer reviewing the cession three years from now will not be evaluating whether each weekly retraining was material. They will be evaluating whether the cumulative drift between the model the reinsurer underwrote and the model the cedent is now running is a change the cedent should have raised. That is an after-the-fact judgment, made by a counterparty with an incentive to find one.

The CRO Often Cannot Answer the Disclosure Question

Ask a chief risk officer at a regional or specialty carrier what their company has formally communicated to its reinsurance panel about the ML models in its pricing stack. In most cases, the honest answer is some version of "we have told them we use models, but we do not have a register of what we said about which model, when, or to whom."

Treaty disclosure registers, where they exist at all, tend to be built around schedule attachments and aggregate exposure summaries. They were not designed to track model versioning, training data updates, or the specific risk-weight changes a retraining produced. The artifact that would resolve a future disclosure dispute, a contemporaneous written record of what the cedent communicated about its model and when, simply is not being produced by most carriers' current governance.

The legal default in reinsurance does not favor the cedent. The doctrine of uberrimae fidei, the duty of utmost good faith, applies more strictly in reinsurance than in primary insurance. Under that doctrine, a cedent's failure to disclose a material fact can entitle the reinsurer to avoid the contract ab initio, even where the non-disclosure was innocent and unintentional.⁵ Modern US courts (notably the Second and Eighth Circuits) typically require the reinsurer to also show actual reliance or prejudice before voiding the policy, but the burden still falls on the cedent to demonstrate that the fact was either disclosed or not material, and "not material" is a hard argument to win retrospectively, after a loss has already developed.

The practical posture this creates for a P&C carrier deploying ML in rating is uncomfortable. A strict disclosure standard sits over a technology that produces continuous undocumented changes, with no contemporaneous record being kept to defend the cedent's position three years later. A reinsurer with a developed loss and a careful coverage attorney has the better hand.

Reinsurers Are Already Underwriting AI Directly

Cedents who assume reinsurers have not yet noticed the AI question should look at what reinsurers are building.

Munich Re sells aiSure,⁶ a performance-guarantee insurance product that underwrites the performance of a customer's AI model directly. The pricing inside aiSure is built on Munich Re's internal models for AI failure modes, drift, and degradation. When a reinsurer writes a contract that depends on understanding how an AI model performs over time, that reinsurer has built capability to evaluate AI models. That capability does not stay siloed inside the aiSure product. It informs how the same reinsurer evaluates the cedents it accepts as treaty counterparties.

The implication for a P&C cedent is direct. The reinsurer accepting your catastrophe XL treaty has, or is building, the analytical machinery to form an opinion about the ML in your pricing stack. They are forming that opinion whether your CRO sits down with them at renewal to walk through it or not. The opinion they form in the absence of cedent-led disclosure is unlikely to be the most generous one.

We covered the broader version of this argument in the canonical post on why reinsurance, not regulation, is the existential AI risk for insurers. The treaty disclosure warranties are the specific contractual hinge that makes that broader argument operational. Everything in the canonical thesis routes through them.

The Question for the Next Board Meeting

To be precise about precedent: no published case has yet rescinded or non-renewed a reinsurance treaty for AI model drift specifically. The contractual mechanics, the uberrimae fidei doctrine, and the reinsurer-side analytical capability are all in place, even though the litigated case has not yet been written. We are describing a train on the tracks, not a wreck.

The boards we think will be best positioned in 2027 and 2028 are the ones whose CEO is asking a short, specific question this quarter:

Do we know what we have disclosed to our reinsurers about our models, and could we produce the disclosure register at next renewal?

Three artifacts answer that question well. Most carriers do not currently produce any of them.

The first is a model change log written for a treaty underwriter rather than a data scientist. The audience matters. A commit history in a Git repository will not satisfy the disclosure standard, because a treaty underwriter cannot extract from it what changed in the model, when the change took effect, or how the change is expected to affect the loss curve by line and territory. The disclosure document is the page that states those three things in plain English.

The second is a treaty disclosure register that records what was formally communicated to which reinsurer at which renewal, with the document attached. The test is whether, three years from now, the cedent could pull a single record showing exactly what its panel was told about a specific model version. If that record cannot be produced, the cedent's defense to a uberrimae fidei challenge has a gap the reinsurer's coverage counsel will find.

The third is a board-level acknowledgment that ML retraining is a treaty-relevant event, codified in the carrier's model governance policy. Putting that acknowledgment in writing changes the internal conversation. It moves model versioning out of the data science team's calendar and into the company's reinsurance placement workflow, where it should have been from the start.

None of this is regulatory. No state insurance department is asking for these artifacts today. Boards should build them anyway, because the contractual remedy for getting this wrong is the loss of the reinsurance protection that lets the carrier survive a one-in-fifty cat year, and that remedy operates without any fine ever being issued.

The CEO question is short. Most carriers cannot answer it yet. The renewal meeting where they find out whether the answer mattered is somewhere on a calendar between now and 2028. The window to build a defensible disclosure register sits before that meeting, with no equivalent window after it.

Related

• The Reinsurance Treaty Is Where AI Risk Becomes Existential
• The 89-Day Death Clock: What FedNat Teaches Every Carrier Building AI
• Merced County, Camp Fire, and What 112 Years of Underwriting Buys You

Footnotes

1. "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 ↩

2. "Special termination clauses allow early contract exit when 'certain contingencies arise.'… Common Triggers: Rating Downgrades… Change of Control… Underwriting Changes… Financial Deterioration… Operational Issues." — IRMI: Special Termination Provisions in Reinsurance Contracts ↩

3. "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 ↩

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

5. "Under common law principles codified in the Marine Insurance Act 1906 (sections 17-20), such non-disclosure, whether innocent or intentional, entitles the reinsurer to avoid the contract ab initio, rendering it void from the outset and potentially leaving the ceding insurer exposed to unrecovered claims without refund of premiums in cases of deliberate breach." — Lexology: Uberrimae fidei — contracting with the utmost good faith ↩

6. "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. When your AI solution comes with aiSure™, Munich Re bears the risk — if a solution does not perform as promised, they will step in and help compensate clients." — Munich Re: aiSure — More AI Opportunity. Less AI Risk ↩