Reinsurance AI: How Reinsurers and Rating Agencies Are Underwriting Your AI

Reinsurance AI is no longer a future topic. The largest reinsurers have spent years building the analytical capability to price AI failure, AM Best's rating methodology already has a lever that can move a carrier across the A- line on governance alone, and the treaty wordings in your binder contain disclosure warranties that were never written with weekly model retraining in mind. Your reinsurer and your rating agency are forming a view of the AI in your rating and claims stack whether or not you take part in the conversation.

This guide covers how reinsurers and rating agencies evaluate carrier AI, what the recent carrier failures teach, and what evidence to have ready before your next renewal.

Why reinsurers now underwrite your AI governance

A reinsurer underwrites the cedent, and the policies come second. The treaty price assumes the cedent will keep selecting, pricing, and settling risk the way it did at placement. When a machine learning model sits in the rating algorithm or the claims workflow, it becomes part of the risk the reinsurer is pricing.

The reinsurance market has built the tools to read it.

Munich Re has written performance coverage on AI systems since around 2018, now packaged as aiSure. The product pays if a customer's model misses a defined accuracy threshold. Pricing it required Munich Re to build a working model of how AI fails: drift, distribution shift in input data, degradation under adversarial inputs, and the rate at which a model trained on 2022 claims becomes a poor description of 2026 claims. That view does not stay inside one product. It informs how the same underwriters look at every cedent with ML in its pricing stack. A carrier placing a cat XL with Munich Re today is sitting across from a counterparty that has studied AI failure for roughly eight years.

Swiss Re signed a Memorandum of Understanding with RIQ at Abu Dhabi Finance Week on December 10, 2025, to develop capacity solutions, risk origination, and AI-enabled capabilities across the UAE. RIQ is an AI-native reinsurance platform launched in June 2025 by IHC with BlackRock and Lunate. The MoU is scoped to the UAE; the underwriting science will not stay there.

Lloyd's published its view in November 2024. The Lloyd's Market Association and Oxbow Partners released The Growth of Enhanced Underwriting in the Lloyd's Market: The New Normal?, drawing on participants representing 77% of Lloyd's 2023 GWP. The top cited barriers to algorithmic underwriting were "loss of control of underwriting decisions" (14%), algorithmic bias (12%), and regulatory concerns (10%), per the LMA.

Most cedents have never walked their panel through the ML in their rating stack, so the reinsurer assembles its own picture from public filings, prior submissions, conference talks by your CTO, your pricing vendor's disclosures, and pattern recognition across your peers. Call that reinsurer-led disclosure. The alternative, cedent-led disclosure, means you present your own documented view of your AI and the underwriter responds to it. Almost every US P&C carrier is in the first posture today.

Why reinsurance AI risk outweighs a regulatory fine

Most AI risk budgets go to the regulatory column: the NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, Colorado SB21-169, New York DFS Circular Letter No. 7 (2024), and EU AI Act penalties of up to 7% of worldwide turnover for prohibited practices under Article 99. Those rules are real. A fine, though, is a P&L event. Carriers absorb it, adjust rates, and keep writing.

Losing a treaty is a balance sheet event. Compare two outcomes for a regional carrier writing $400M in homeowners premium:

• Regulatory: a state fines the carrier $5M for thin AI model documentation, adds a consent order with quarterly reporting and a six-month outside review, and the carrier takes a 4% rate increase at the next filing. A year later the order lifts.
• Reinsurance: the cat XL is non-renewed at June 1, citing an undisclosed "change in underwriting practices and philosophy." A replacement layer, if one exists, costs three times as much. A Q3 hurricane lands, the retention attaches, and $480M of gross loss flows to surplus that was never built to carry it.

The first scenario costs five million dollars. The second can end the company. The chain from model retraining to insolvency runs in five steps, each already supported by treaty practice and case law:

1. The ML rating model retrains and risk weights shift in ways an actuary would have called a change in philosophy.
2. The cedent does not tell the reinsurer, because internal governance does not treat retraining as a treaty event.
3. A cat loss develops larger than the new model priced for, because its training data had not absorbed secondary-peril intensification.
4. The reinsurer reviews the cession, cites the disclosure warranty, and rescinds, non-renews, or reprices.
5. The next loss attaches at net retention with no layer above it.

To be precise: no published case has yet found a reinsurer rescinded a treaty for AI model drift. The mechanics are in place; the precedent has not been written.

The reinsurance treaty clause your AI strategy depends on

What the disclosure warranty says

Standard treaty wordings carry a family of termination and disclosure warranties. IRMI catalogues them in its coverage of special termination provisions and underwriting and claims clauses. They typically let the reinsurer non-renew, cancel, or rescind on:

• a rating drop below a contractual floor, commonly AM Best A- or S&P BBB+ (Moody's syndicate ratings below B+ also appear);
• a material drop in surplus or a change in ownership;
• a change in the cedent's underwriting practices, rating philosophy, or claims-handling approach.

The last trigger is the one AI touches. IRMI reproduces a claims-handling version verbatim: "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."

Episodic clauses, continuous models

These warranties were drafted for episodic change: a new state, a new line, a pricing committee loosening guidelines. Dated, auditable, signed by a chief underwriting officer. ML pricing produces continuous change. Every retraining resorts the book into different rating cells, and two policies the prior model priced identically now diverge, though no committee voted on anything. Selection behavior on the next thousand binds differs from the last thousand.

A reinsurer reviewing a cession three years from now will not ask whether each weekly retrain was material. It will ask whether the cumulative drift between the model it underwrote and the model you now run is something you should have raised. That judgment gets made after the fact, by a counterparty with a reason to find a breach.

Utmost good faith raises the stakes

Reinsurance applies uberrimae fidei, the duty of utmost good faith, more strictly than primary insurance does. Under the common-law rule codified in the Marine Insurance Act 1906, non-disclosure of a material fact, innocent or not, can entitle the reinsurer to avoid the contract ab initio. Many US courts describe the cedent's duty as affirmative, with no reciprocal duty of inquiry on the reinsurer. Modern US courts, notably the Second and Eighth Circuits, usually require the reinsurer to show reliance or prejudice too, but the cedent still has to prove the fact was disclosed or immaterial. "Immaterial" is a hard argument once a loss has developed.

Ask most CROs what they have formally told the panel about their pricing models and the honest answer is "we told them we use models." Disclosure registers, where they exist, track exposure schedules, not model versions.

Follow the fortunes after Utica v. Clearwater

Many executives still treat follow-the-fortunes (FTF) and follow-the-settlements (FTS) as background doctrine that rides along with any treaty. Both exist so a cedent can cede a claim without relitigating coverage against its own reinsurer.

In Utica Mutual Insurance Co. v. Clearwater Insurance Co., 906 F.3d 12 (2d Cir. 2018), Utica argued its facultative certificates should be read to include a follow-the-settlements obligation they did not contain. The Second Circuit vacated and remanded on that cross-appeal because Utica had not shown it was entitled to a judgment binding Clearwater to its settlement. The protection has to come from the contract. In 2022 the Eleventh Circuit, in Public Risk Management of Florida v. Munich Reinsurance America, Inc., 38 F.4th 1298, declined to imply a follow-the-fortunes clause absent express language. Two circuits now point the same direction.

Even where FTF and FTS are written in, the doctrine carries standard exceptions: cedent bad faith, ex gratia payments, settlements outside the underlying coverage, and gross negligence in claims handling. AI in the claims pipeline gives a reinsurer's counsel more to work with on each. Picture a claim that passed through an intake bot, automated coverage triage, AI-assisted reserving, and an ML fraud flag before a human adjuster signed off. If the cedent cannot reconstruct why the model recommended what it did, settlement leverage moves toward the reinsurer. If review concludes the model approved a payment the policy did not require, the cession is exposed under the ex gratia exception even with FTS in the wording.

No published case has refused a cession because AI shaped the claim decision. The defensible posture is still clear: audit every layer for express FTF/FTS language (older facultative covers and specialty layers placed through newer markets are where gaps tend to show up), keep a written record that a qualified person reviewed and adopted each material AI recommendation, and treat large or unusual cessions as if FTS were absent, with the merits written down.

How rating agencies evaluate carrier AI

AM Best's ERM building block and the comprehensive adjustment

Best's Credit Rating Methodology builds a rating from four blocks: Balance Sheet Strength, Operating Performance, Business Profile, and Enterprise Risk Management. Balance sheet anchors the analysis; the other three adjust it within published ranges. ERM is scored from Very Strong to Very Weak and, per AM Best's overview, weighs risk management capability against the carrier's risk profile. The criteria cover risk identification, framework, culture, governance, and the documented path from a breached limit to executive attention.

A machine learning rating model is a risk-bearing instrument that can put loss on the books without an underwriter touching a file. ERM already has the vocabulary to read it that way.

The lever is the comprehensive adjustment, the methodology's mechanism for moving the indicated rating on holistic judgment the quantitative blocks do not capture. Material ERM weaknesses are an explicit basis for a downward move. For a carrier at the A-/B++ boundary, a committee that reads ERM as Weak or Very Weak can move the published rating across the line even if the balance sheet looks fine. That matters because the treaty rating trigger usually sits at A-. A governance read can now open the same chain that used to require a deterioration in surplus, and it can open it sooner.

What the 2023 downgrade data shows

AM Best recorded 55 US P&C downgrades in 2023 against 35 upgrades, up from 30 downgrades in 2022. Personal lines moved hardest: 39 downgrades and nine upgrades, versus 18 and 10 the year before. The cited drivers were reserve adequacy, cat volatility, and reinsurance dependency rather than AI, but governance language has moved from footnotes into headline rationale. Rating action tends to follow treaty action with a lag, so a hard renewal and an outlook review often land in the same quarter.

What the analyst will ask

Expect the ERM portion of the next interactive review to cover:

• Ownership. Who is the named risk owner with authority to pull the model from production? A department name does not answer the question.
• Monitoring. What drift thresholds trigger escalation, and to whom? A dashboard nobody is required to act on does not count as a safeguard in this framework.
• Reinsurer disclosure. What did you tell the panel at the last renewal about model changes during the treaty period? The analyst reads this with one eye on whether the reinsurer would consider itself adequately advised.
• Incident response. Is there a documented escalation path with named owners and timing?

Commit histories, retraining notebooks, and Slack threads do not satisfy the framework. The analyst wants a written governance system with owners and evidence that it runs.

Demotech's 2022 Florida letter is the template

In mid-July 2022, Demotech sent letters to roughly seventeen Florida property carriers warning of downgrades from "A" (Exceptional) to "S" (Substantial) or "M" (Moderate), as reported by Artemis. The letters leaked within days. On July 21, OIR Commissioner David Altmaier wrote to Demotech demanding a more comprehensive review, and CFO Jimmy Patronis wrote to the FHFA, Fannie Mae, and Freddie Mac questioning their reliance on Demotech.

The letter was lethal because of what sat downstream. The Fannie Mae Selling Guide B7-3-01 rating table accepts Demotech "A" or better, alongside AM Best "B" or better, S&P "BBB" or better, and KBRA "BBB" or better (see also B7-3-02). Demotech "S" does not qualify. Policies on conforming mortgages become non-compliant at renewal, lenders force-place coverage elsewhere, and the book unwinds in weeks. A rating agency needs no regulatory authority to close a carrier; it only needs to sit upstream of an institution with a hard threshold.

Three lessons carry into the AI era:

1. The market reads the letter before the action is final. Brokers, reinsurers, and competitors hear informally. Repair work has to be in place when the early-warning letter arrives.
2. Rating analysts read governance more strictly than regulators. A carrier that just cleared a market conduct exam can still face a rating action, because the analyst's standard for effective governance sits above the bulletin's standard for satisfactory compliance.
3. The response has to exist before it is needed. The 2022 recipients that fared best had a designated executive owner for agency communications, current governance artifacts, a legal review path, and a board notification protocol. The worst spent the first week deciding who should draft the reply.

The AI version will likely come from AM Best, cite the ERM block, note missing documentation for retraining cadence, drift escalation, or treaty disclosure, and invite a response within a short window before a one-notch comprehensive adjustment. The downstream institution is usually the treaty rating trigger; for a workers' comp writer it might be a state plan administrator, and for a surplus lines writer, a broker's eligibility list. Know which doorway applies to your book.

The NAIC bulletin is the floor reinsurers hold you to

As of August 2025, twenty-four jurisdictions had adopted a version of the NAIC Model Bulletin, adopted by the NAIC on December 4, 2023. It creates no new law. It explains how departments will apply existing unfair trade practice, market conduct, and governance statutes to AI, and its expectations cluster in four areas: a board-owned written AIS program with testing, validation, and monitoring; a risk framework calibrated to the carrier's AI use in underwriting, rating, claims, marketing, and fraud; governance of third-party AI and data; and records sufficient to reconstruct the AI's role in a consumer decision. (Our NAIC model bulletin guide walks through it section by section.)

The bulletin is principles-based, so two carriers can both be compliant with very different evidence behind them. That variance is what a reinsurer reads. The bulletin gives treaty underwriters something they lacked before December 2023: a regulator-blessed baseline for reasonable carrier AI governance. A cedent at the baseline is "in the market." A cedent visibly above it earns credit. A cedent below it becomes a repricing candidate.

Consider two homeowners carriers at the same June 1 placement. Cedent A has a 200-page manual mapping every bulletin paragraph to a policy and an owner; it would survive a market conduct exam, but it has never been externally audited, model history lives in the data science team's Git repository, and nothing has gone to the panel. Cedent B has the same manual plus a quarterly external validation report, a dated model change log written for non-technical readers, and a letter to the lead reinsurer at each renewal covering material model changes. Both are compliant. In a hardening market, Cedent B's extra evidence is often what decides price and whether the reinsurer renews at all.

Two state regimes already sit above the floor and shape what reinsurers consider serious governance. New York's Circular Letter No. 7, issued July 11, 2024, covers AI and external consumer data in underwriting and pricing and expects documented testing and ongoing monitoring for unfair discrimination. Colorado's SB21-169 regulations go further, requiring quantitative testing across protected classes and reporting results: the life regulation (Reg 10-1-1) took effect in November 2023 and the auto and health rules on October 15, 2025. The EU AI Act registers too, because large reinsurers carry European exposure and have built internal frameworks against it. Benchmark your program against New York and Colorado, and the reinsurer can see and price the difference.

Case studies: what losing reinsurance support looks like

None of these failures was caused by AI. They show the mechanism AI would run through, and several show the blind spot ML rating models share.

FedNat: 89 days from treaty expiry to receivership

FedNat's catastrophe reinsurance program expired June 30, 2022; the company had said it could not secure adequate reinsurance without reorganizing. The June 1 placement did not come together at the limits and prices the book needed. On September 27, 2022, FedNat was ordered into receivership. Eighty-nine days, with no hurricane attaching and no fine in the chain. The AI-era version swaps an AM Best ERM downgrade in for the Demotech letter and runs the rest unchanged.

UPC and the Demotech seventeen: exhaustion and cascade

United Property and Casualty ran into Hurricane Ian on September 28, 2022. Expected gross losses of $660M came in at $864M by year-end SEC filing, and the estimate later rose to $1.5 billion. The cat program exhausted. On February 6, 2023, UPC filed pro forma statements showing negative surplus of about $217.6M, and on February 27 a Leon County court ordered liquidation. The AI analogue: a model concentrates the book in cells that look favorably priced on historical data while the treaty is still structured on the prior exposure view, and the first large event finds the gap.

Several of the other Demotech recipients survived through state-engineered transfers into Citizens, which passed 1.4 million policies by fall 2023 and became Florida's largest property writer. Most states have no comparable residual market; a failed regional carrier in Ohio or Tennessee sends policyholders through the guaranty fund, with caps and delays.

Merced and the Camp Fire: 112 years gone in 25 days

Merced Property and Casualty was founded in 1906 by farmers who could not buy fire insurance, and carried an AM Best A- before the Camp Fire ignited on November 8, 2018. The fire killed 85 people, destroyed more than 18,000 structures, and caused an estimated $12.5 billion in insured losses per Munich Re. Merced had about 200 policies in Butte County, concentrated enough that loss above the treaty plus the retention below it exceeded surplus. The commissioner petitioned on November 30, and on December 3 the court appointed a liquidator.

Now add an ML rating model. It finds a Sierra foothill cluster with benign recent fire history, indicates a competitive rate, and the book there grows from 4,200 to 11,500 policies while the cat XL stays placed on the old aggregation profile. Two fire events in one season exhaust the program and push $180M–$260M to net retention against $310M of surplus. California adds a structural lag: under the longstanding reading of Proposition 103, rate filings relied on roughly twenty years of historical data rather than forward-looking cat models, a regime the Sustainable Insurance Strategy is now reforming.

Hippo's 273% loss ratio: training data on an old climate

Hippo reported a 273% net loss ratio on homeowners in Q1 2023, paused new homeowners business nationwide, and reemerged selectively in areas it believed were adequately priced, while shifting toward agency distribution. Its model read certain territories as low-tail on historical loss data. Then 2023 brought record US severe convective storm losses above $50 billion.

Every ML model pricing cat-exposed property works from a climate some years old. Aon put secondary perils at 86% of global insured losses in 2023; primary perils were just 14% of $118bn (also reported by Reinsurance News). ILS manager Elementum Advisors refit its licensed wildfire model after analyzing nearly two million US wildfires, because it had been "benchmarked to historical trends and not to today's climate." Retraining quarterly on rolling claims does not fix this; it records where the climate already was when the loss occurred. This is the model drift problem expressed in a quarterly P&L.

After a quarter like that, a reinsurer asks what changed in the model since placement, what TIV sits in the loss zip codes and how the model ranked them, and what governance keeps the fix durable.

The 17-point loss ratio swing

State Farm General stopped accepting new California property applications on May 27, 2023, citing construction costs, catastrophe exposure, and "a challenging reinsurance market." In March 2024 it announced non-renewal of about 30,000 California policies starting July 3, 2024. At the parent level, State Farm's Q3 2023 homeowners loss ratio hit 99.4%, up 26 points year over year, driven in part by Maui. (California-only results sit in separate CDI filings.)

Year-over-year loss ratio variance in a stable model and peril mix should sit in single digits. A swing of 17 points, roughly $68M on a $400M book, forces the treaty underwriter to decompose it: a peril shift that shows up across peers, a model misspecified at the last filing, or a book that changed composition underneath the rate. Only the first can be defended with industry data. Under Prop 103's historically backward-looking rate process, the corrective filing gets judged against a baseline that already includes the bad year. A reinsurer can compute the swing from public filings without opening your model documentation, and it moves before the regulator does. State Farm had the scale to withdraw in stages; a regional carrier usually does not.

Lemonade's AI Jim: public statements as disclosure events

On May 24, 2021, Lemonade posted a thread saying its AI analyzed claimant videos for "non-verbal cues." Within about two days it retracted the wording, calling it "a bad choice of words" for facial recognition used to flag duplicate identities, stating that flagged claims go to human investigators and that it has never let AI auto-reject claims (CNN coverage). That retraction belongs in any retelling.

A treaty desk would read the episode as a gap between what was said publicly and what the placement file contained, which alone justifies follow-up questions at renewal. A blog retraction does not cure the disclosure file. No published reinsurance dispute has turned on an AI tweet, but the fix is cheap: a short list of treaty-sensitive AI topics (claims, pricing, and fraud models, and anything that moves the loss curve) that route through reinsurance review before publication, one approved description of what your AI does shared by PR and the placement team, and a correction protocol that notifies the broker within 24 hours and the panel in writing within 72, logged in the disclosure register.

How specialty lines differ: cyber, professional, marine

Personal-lines AI playbooks assume a stable risk surface. Specialty lines do not have one.

Cyber models drift in weeks. The At-Bay 2026 InsurSec Report showed Akira ransomware claims up 364% between Q3 and Q4 2025, with one crew driving more than 40% of ransomware claims on At-Bay's book. VPN-initiated compromise rose from 38% of incidents to 73%, and SonicWall appliances appeared in 27% of ransomware claims and 86% of Akira attacks. A quarterly review cadence leaves a cyber model pricing a stale threat surface for months. Cyber needs continuous evaluation against threat-intelligence feeds and documented mid-cycle retraining triggers, with evidence they fired. The Colorado AI Act's annual impact assessment and the bulletin's monitoring expectation are both too slow on their own.

Professional liability resists feature-level explainability. D&O and E&O models encode context (board history, disclosure narrative, litigation profile) as embeddings. SHAP and LIME work on tabular features and break down there. The useful audit trail runs from source documents to decision: the spans the model relied on for each submission. Brokers now ask how AI-augmented submissions are documented in the file. Our AI explainability guide covers the methods.

Marine geospatial models fail abruptly. Satellite-trained transit models give confident, narrowly wrong answers on inputs just outside the training data: a reconfigured port, an ice-opened route, a new piracy hotspot. Aggregate metrics like AUC will not surface a $30M single-cargo error in time; out-of-distribution scenario testing and human review of predictions outside a defined spatial or temporal envelope will. Marine treaty negotiations in 2025 and 2026 have started asking whether imagery-derived models drive acceptance, pricing, or aggregate management.

Surety deserves a mention: models trained on pre-2022 contractor financials now face different supply-chain, labor, and completion risk, and bank-style model risk frameworks fit better than the personal-lines bulletin response.

Across specialty, three questions keep appearing in treaty submissions: the model inventory by line with retraining cadence and triggers; the drift detection process and the longest detection-to-response gap last treaty year; and per-decision documentation where AI materially shaped acceptance, pricing, or limits.

Where ESG reporting overlaps

Rating agencies and investors increasingly ask questions spanning ESG and AI governance, and the work is largely shared. TCFD, the EU CSRD, and the NAIC Climate Risk Disclosure Survey demand documented, auditable methodology, the same discipline as model cards and decision audit trails. The "S" in ESG and AI fairness testing rely on the same disparate impact analysis.

The overlap is sharpest in underwriting. A model pricing coastal flood risk must be scientifically sound (environmental commitment) and non-discriminatory (social commitment and fairness requirement); property-level features that correlate with community demographics can pass one test and fail the other. AI-driven investment screening that misclassifies a high emitter fails both programs. Running the two on separate documentation and committees produces conflicting findings on the same system; shared evidence gives a rating analyst one consistent story.

What to have ready for your reinsurance renewal

No state department asks for most of what follows. Your reinsurer and your rating analyst will, and the time to build it is before the meeting where the question becomes contested.

Two habits tie it together. First, write governance materials in two registers from one evidence base: a regulator-facing map to the bulletin, and a reinsurer-facing record of model performance, bias testing, external validation, and change history over time. Second, treat every placement as a governance event and send the lead reinsurer a model-change letter at each renewal. ORSA exercises usually model treaty exhaustion; few model rescission, which arrives without a triggering loss and needs a different recovery path.

All of this depends on records most carriers do not keep yet: a continuous, audit-ready log of every AI interaction, governance mapped to NAIC and state requirements, and reporting a board, a rating analyst, or a treaty underwriter can read without a data scientist in the room. That is the work Swept's governance offering is built to keep running between renewals. If you want a quick read on where your evidence stands today, the NAIC readiness scorecard takes a few minutes, or talk to us about what your panel is likely to ask at the next placement.