What Training Data Powers Insurance Underwriting and Claims AI?

Why insurance AI needs its own training data

A general-purpose model can draft a plausible claims-denial letter or summarize a policy in plain English. It cannot underwrite a risk, adjudicate a claim, or price a policy, because those decisions carry financial and legal consequences that require grounding in real loss history, not fluent-sounding text. Insurers have poured resources into AI precisely because the underlying processes, evaluating risk, paying claims, catching fraud, are data-dense and repetitive at scale. McKinsey estimates generative AI alone could unlock $50 billion to $70 billion in US insurance industry revenue, concentrated in claims, underwriting, and customer operations. ^[4]

That value only materializes if the models are trained on the right data. Insurance AI breaks down into five core applications, and each one is bottlenecked by a distinct, mostly proprietary data type:

• Underwriting risk models. Score applicants and price policies using application data, historical loss experience, and external consumer data sources.
• Automated claims processing and fraud detection. Triage incoming claims, flag anomalies, and route suspicious files to Special Investigations Units (SIU).
• Damage assessment from photos and video. Estimate repair cost and severity for auto and property claims directly from images, cutting adjuster inspection time.
• Policy Q&A chatbots. Answer coverage, billing, and claims-status questions using policy language and prior support interactions.
• Actuarial pricing models. Set rates and reserves from aggregated loss, exposure, and mortality/morbidity data.

The data types that actually power insurance AI

Not all insurance data is equally useful for training. The data that moves model performance the most is the data with the richest labels and the clearest link to a real-world outcome.

• Historical claims records with outcomes. The claim narrative, adjuster notes, coverage determination, and final payout or denial. This is the single highest-value asset in the vertical, covered in depth below.
• Policy and underwriting data. Application answers, rating factors, endorsements, and the resulting bound policy terms, the ground truth for underwriting models.
• Telematics and IoT sensor data. GPS, acceleration, braking, and cornering data from connected vehicles, plus smart-home sensor feeds for property risk. ^[4]
• Property and vehicle imagery and video. Pre-loss and post-loss photos, drone and satellite imagery, and repair-estimate line items paired with images, the training signal for computer-vision damage assessment.
• Call transcripts and claims intake notes. First-notice-of-loss calls, adjuster-claimant conversations, and chat logs that train intake automation and policy chatbots.
• Fraud-labeled datasets. Claims with a confirmed SIU finding, either substantiated fraud or a cleared investigation, which is the supervised signal fraud models actually need.

Telematics illustrates why metadata and linkage matter as much as raw sensor volume. A feed of GPS pings and hard-braking events is only useful for underwriting or claims work once it's tied to a policy, a driver, and ideally a loss event. Usage-based programs that combine that linkage with driver feedback have been associated with roughly a 50% reduction in crash risk, which is exactly the kind of outcome signal that makes telematics-linked claims more valuable than an unlinked sensor dump. ^[4]

Why labeled claims-outcome data is the scarce, high-value asset

Raw claim volume is not scarce. Carriers process millions of claims a year. What's scarce is the labeled outcome: a claim record where a human adjuster's final coverage decision, settlement amount, and (for a fraud model) confirmed disposition are all attached to the underlying claim file in structured, usable form. Three things make this hard to get at scale.

• 1The label requires human adjudication. An adjuster's coverage determination is a judgment call informed by policy language, precedent, and negotiation, not a fact that exists in the world before the claim is handled. There is no way to synthesize it after the fact with the same fidelity.
• 2Confirmed fraud is rarer than suspected fraud. Most claims flagged for investigation are cleared or closed without a definitive finding. A dataset with genuinely adjudicated fraud outcomes, positive and negative, is far smaller than a dataset of merely flagged claims, and the imbalance is exactly what a supervised fraud model needs handled correctly. The stakes justify the effort: insurance fraud is estimated to cost US consumers $308.6 billion a year across all lines, with the FBI estimating it adds $4,000 to $7,000 in increased premiums per family over a 10-year period. ^[8]
• 3Retention schedules destroy the asset before it's ever licensed. Many carriers and MGAs purge or archive closed claim files on a fixed schedule. A startup MGA or insurtech winding down often holds the single window in which that outcome-labeled history is still intact and exportable.

The US state regulatory landscape for insurance AI

Insurance is regulated state by state under the McCarran-Ferguson Act framework, and AI oversight has followed the same pattern: no single federal rulebook, but a fast-converging patchwork of state bulletins, regulations, and bills. Any buyer training on claims, underwriting, or policy data has to clear this landscape, not just federal privacy law.

The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted by the National Association of Insurance Commissioners in December 2023, is the baseline. It requires insurers to maintain an AI governance framework, document how AI systems are developed and validated, and be ready to produce that documentation to regulators on request. ^[4] As of March 2025, 24 states plus Washington D.C. had adopted the bulletin with little to no material changes, roughly half of US jurisdictions, and adoption has continued to spread since. ^[8]

Several states have gone further with binding, insurance-specific rules rather than a principle-based bulletin:

• Colorado (SB21-169 and Regulation 10-1-1). Requires covered insurers to inventory every model that uses external consumer data or algorithms, run quantitative bias testing, and file annual compliance attestations. Adopted for life insurers first, the governance and testing framework was extended to health and private passenger auto insurance in August 2025. ^[6]
• New York (DFS Circular Letter No. 7, 2024). Requires insurers using AI systems or external consumer data in underwriting and pricing to demonstrate a "clear, empirical, statistically significant, rational, and not unfairly discriminatory relationship" between every model variable and the insured's actual risk. ^[4]
• Illinois (HB 35, pending). Would bar health insurers from issuing an adverse coverage determination that results solely from an AI system, and require meaningful human review by someone with authority to override the model. The bill passed the Illinois House and was pending in the Senate as of mid-2025, not yet enacted. ^[6]

"Unfair discrimination": the central compliance concept, plus privacy law

Outside insurance, "discrimination" usually means treating a protected class worse than others, and the compliance question is whether a model does that. Inside insurance regulation, "unfair discrimination" is a defined legal term of art that predates AI by decades and covers more ground than civil-rights law does. It has two distinct branches, and every state AI rule discussed above is an AI-specific application of one or the other.

• Protected-class discrimination. Distinctions based on race, religion, national origin, and similar protected characteristics are barred outright, regardless of whether there's an actuarial justification for them.
• Actuarial (class-plan) discrimination. Separately, and this is the part general anti-discrimination law does not cover, insurers may not charge people in the same actuarial risk class different rates without a sound actuarial basis for the difference. A pricing variable that correlates with risk for a spurious or unexplainable reason can be "unfairly discriminatory" even with zero protected-class involvement.

New York's DFS made this second branch explicit for AI: the empirical-relationship standard requires insurers to show the statistical link between a model input and actual risk before they may use it, precisely to catch variables that predict outcomes for reasons regulators can't validate as actuarially sound. ^[2] Colorado's quantitative testing regime exists to detect the same problem at the model level, not just the variable level. ^[4]

This matters directly for anyone assembling or buying a training dataset, because the actuarial branch is the one AI makes newly dangerous. A model trained on millions of rows can surface a variable, a ZIP code, a device type, a call-time pattern, that correlates with loss experience for a reason no actuary can explain and that happens to track a protected class closely enough to function as a proxy. Regulators call this proxy discrimination, and it's why Colorado's testing rule and New York's empirical-relationship standard both push insurers to test a model's outputs against protected-class membership even when the model never sees a protected-class field directly. ^[2][4] A dataset seller who can document which fields were included, excluded, and tested for exactly this problem is offering a materially more valuable asset than one who can't.

For any dataset that includes EU or UK personal data, whether claims from an international insurtech or telematics data from EU-based drivers, the GDPR's lawful-basis requirements apply on top of state insurance rules. Consent, contractual necessity, or legitimate interest must justify the original collection and any transfer for AI training, and that basis has to be documented before the data changes hands. ^[2] A growing number of US states also layer their own consumer-privacy statutes on top of insurance-specific rules, particularly around telematics and biometric-adjacent data such as driving behavior.

Data type, AI use case, and the regulatory issue that governs it

Worked example: evaluating an auto-claims dataset for AI training

Consider a regional auto-insurance MGA winding down operations. It holds 640,000 closed auto claims spanning six years across 14 states, with telematics feeds linked to 180,000 of them, first-notice-of-loss photos attached to 210,000, and SIU-confirmed fraud dispositions on 4,100 claims. Here's how a buyer-facing evaluation actually runs.

• 1Check outcome-label completeness first. Of the 640,000 claims, how many carry a final coverage determination and settlement amount in structured fields, not just free-text adjuster notes? If only 60% do, the trainable claims-outcome subset is closer to 380,000, not 640,000, and that's the number that sets value.
• 2Score the telematics linkage separately. The 180,000 claims with a matched telematics feed support usage-based pricing and driving-risk models that the other 460,000 can't. That linked subset is a distinct, higher-value asset, not a rounding error on the whole book.
• 3Assess imagery consistency. Are the 210,000 photo sets standardized (angle, resolution, metadata) or an inconsistent mix from different adjuster tools over six years? Inconsistent imagery needs normalization before it trains a reliable damage-assessment model, which affects price and time-to-use.
• 4Weigh the fraud-labeled slice on its own terms. 4,100 confirmed dispositions out of 640,000 claims is a realistic, useful base rate for a fraud model, small in absolute count but exactly the kind of adjudicated signal that's hardest to source elsewhere.
• 5Map the jurisdiction mix against the regulatory landscape. Claims from Colorado and New York carry the added burden, and the added credibility, of having been generated under quantitative-testing and empirical-relationship regimes. ^[4][6] Claims from states with no AI-specific insurance rule yet are simpler to clear but offer a buyer less evidence the underlying rating practices were already validated.
• 6Confirm PII/PHI scrubbing and consent documentation before any sample leaves the building. Claimant names, VINs, addresses, and any bodily-injury medical detail need redaction or tokenization, and the wind-down entity needs to confirm its original policy language and privacy notices permitted this kind of downstream use.

How this data gets priced

Insurance claims data follows the same seven value drivers that govern any vertical dataset: volume, quality and annotation depth, domain scarcity, metadata richness, recency, licensing terms, and legal cleanliness. Two of those drivers do disproportionate work here. Annotation depth, specifically whether the outcome, settlement, and (for fraud) disposition fields are structured rather than buried in free text, separates a commodity data dump from a trainable asset. And legal cleanliness, meaning documented consent for downstream use plus PII/PHI handling appropriate to each state's rules, is a gate rather than a bonus: buyers with real compliance obligations will not touch outcome-labeled claims data without it, regardless of price.

Licensing terms move price the same way they do in every other vertical. A non-exclusive license lets a seller monetize the same claims history to multiple buyers, while an exclusive license removes the data from the market entirely for the buyer's competitive benefit and commands a substantial premium over the non-exclusive baseline. For a wind-down entity, the practical takeaway is the same one that applies across healthcare, legal, and finance data: structure and labeling, not raw row count, determine which buyers show up and what they're willing to pay.

The claims files insurers are about to delete are the asset the market wants

Insurance AI is advancing fast on the applications side, underwriting, claims automation, fraud detection, damage assessment, and chatbots are all shipping in production today. What's advancing more slowly is the supply of the one input that actually differentiates these models: outcome-labeled claims history, telematics-linked risk data, and confirmed fraud dispositions. That data is proprietary by nature, shrinking as retention schedules purge old files, and increasingly governed by a real, binding regulatory framework rather than a vague promise of "responsible AI."

For a carrier, MGA, or insurtech winding down or pivoting, that combination, genuine scarcity plus rising buyer demand, is exactly what turns a closed claims book from a compliance liability into a sellable asset, provided the provenance, consent, and jurisdictional exposure are documented before a buyer ever sees a sample.