AI Use Cases in Insurance by Business Line (Governance Map)

Regulators do not assess “AI” in the abstract. They assess how it is used in a specific operation, and the governance expectations sharpen wherever AI touches a decision a consumer feels. That is why the most useful way to plan an insurance AI program is not by technology but by business line. The same model carries a light burden in one function and the heaviest scrutiny in another, depending on the decision it influences. This guide walks the major lines and what each puts on an examiner’s checklist, with links to the deeper analysis behind each one.

AI use cases in insurance span the full policy lifecycle: underwriting and pricing, claims triage and settlement, fraud detection, and customer and distribution support. What determines the governance burden is not the sophistication of the model but the consequence of the decision. A system that ranks marketing leads and one that recommends a claim denial may share the same architecture and sit at opposite ends of the regulatory-risk scale.

Claims: the hardest case

Claims is where AI most directly shapes an outcome a consumer feels, which makes it the focus of the NAIC evaluation tool’s most demanding exhibit. The line that regulators care about is autonomy. A model that scores a claim for a human to review is one thing. An agentic system that recommends or takes an action, and that a busy adjuster rarely overrides, is another. The documentation bar rises with autonomy, and the question examiners ask is whether you can reproduce, for any given claim, the recommendation the system made, the decision the human reached, and the reason for any difference.

This is why claims AI often carries governance debt. The capability arrived before the paperwork. Insurers can now triage, estimate, and even settle claims faster than they can explain any single decision. The examiner’s question is simple: if a consumer disputes this, can you show the machine’s reasoning, the human’s judgment, and the gap between them? If not, you do not have a claims-AI system. You have a claims-AI exposure.

The pressure is highest where the system is autonomous or near-autonomous. A first notice of loss chatbot that routes documents is low risk. A model that recommends a total loss or a settlement amount is high risk. A model that can issue payment without a person in the loop is the highest risk. Each step up the autonomy ladder requires more documentation, more testing, and more human-review evidence.

Underwriting

Models that sort, rate, and price are the longest-standing use of AI in insurance and the most exposed to unfair-discrimination scrutiny. Accelerated underwriting, telematics, aerial imagery, and external consumer data all feed decisions about who gets covered and at what price, across P&C, life, and health. The recurring regulatory theme is proxy discrimination: a variable that correlates with a protected class without a defensible business justification. Examiners increasingly expect carriers to test for disparate impact, document the analysis, and show there was no less-discriminatory alternative available.

Underwriting AI is a risk multiplier because the decision is made before the consumer has a relationship with the carrier. A denied application or a higher premium has no prior claim history to contextualize it. The consumer’s only recourse is to the regulator. That makes the stakes asymmetric: a small error rate in a large underwriting model can produce a class of people who were treated unfairly, and the regulator knows how to find them.

Life underwriting has its own pattern. Accelerated underwriting models that use prescription history, credit attributes, or third-party data to bypass traditional labs can speed up a sale. They also concentrate risk in the data sources no one in the agency has met. A variable that is predictive of mortality but also correlated with race, income, or disability status will draw scrutiny, especially if the carrier cannot explain why it is necessary or what alternatives were considered.

Health underwriting is constrained by the Affordable Care Act for individual products, but employer-sponsored and short-term products still use risk-selection tools. Medicare Advantage and other segments use prior health data to project cost. The principle is the same: a rating or eligibility variable that is not grounded in the actual risk being insured is a proxy discrimination risk waiting to happen.

Pricing and rating

Pricing and rating sit next to underwriting but deserve their own attention because the decision is made repeatedly, often in real time, and usually without human review. A rate indication or renewal adjustment that passes through an AI model can affect millions of policyholders. The model may be technically accurate in predicting loss, but if its inputs include variables that act as proxies for protected characteristics, the outcome is still unfair.

Telematics is the clearest example. Driving behavior data can be more accurate than traditional rating variables. It can also become a proxy for income, geography, or employment if the way the data is collected or scored embeds those patterns. A carrier that uses telematics without testing for disparate impact is betting that the model is clean. A carrier that tests and documents the analysis is betting on evidence.

Aerial imagery and property-condition data for homeowners pricing raise a similar question. A roof-age model trained on satellite photos may be precise, but if its accuracy varies by neighborhood or housing stock, the pricing pattern can become discriminatory in effect even if no protected class is an input. The governance task is not to eliminate every correlation. It is to know which correlations exist, justify the ones that are actuarially necessary, and mitigate the ones that are not.

Pricing models also have a renewal problem. A model trained on new-business data may drift when applied to renewals, where the customer base and risk profile differ. A regulator that finds a renewal book with a different error rate or a different pattern of outcomes will ask why the model was not retested for the new population. Governance here means monitoring, not just one-time validation.

Health insurance and utilization management

Health insurance is the first flashpoint for AI governance, because prior-authorization, utilization review, and claims-denial systems sit closest to a harm a regulator will act on. A large majority of health insurers already use AI somewhere in these operations, and many rely on third-party components they cannot fully explain. That combination, high-stakes decisions plus outside models, is exactly what draws attention. Health plans face the same governance expectations as other carriers, sharpened by the visibility of the decisions their AI touches.

The UnitedHealth case is the wake-up call. When a large payer’s AI investment collides with governance expectations, the resulting scrutiny affects the whole industry. Regulators do not need to prove a system is biased to create problems for a carrier. They only need to show that the system was not documented, not tested, or not overridable by a clinician.

Health AI also has a unique vendor problem. Many plans use third-party utilization-management tools that apply the same criteria across multiple carriers. A plan that relies on a vendor’s policy without its own clinical review is exposed if the vendor’s criteria become the subject of litigation or regulation. The obligation to govern stays with the plan, even when the tool is licensed.

Fraud detection and operations

Fraud detection and operational triage carry their own monitoring expectations, even though they can look lower-stakes than underwriting or claims. A fraud model that flags a claim for review may influence the outcome enough to matter, depending on how the flag is used downstream. False-positive rates, the consumer impact of a wrongful flag, and the drift of a model over time all come into scope. The governance question is not whether the system catches fraud, but whether you monitor what it does to legitimate claimants along the way.

The same model used in two different workflows can have two different governance profiles. A fraud flag reviewed by a specialized investigator before any action is taken is low risk. The same flag used to delay payment automatically or to trigger a denial is high risk. Governance must follow the consequence, not the model’s name.

Operational triage models, such as work-queue prioritization or resource allocation, are often overlooked because they do not touch a consumer decision directly. But if the triage affects how quickly a claim is handled, or which claims get the most experienced adjuster, it can influence outcomes. The bar is lower, but documentation is still required. A carrier should be able to explain why the model ranks one queue above another and what the downstream effect is.

Marketing and distribution

Marketing and distribution use AI to rank leads, personalize offers, and route prospects to agents. The regulatory risk here is different from underwriting: a marketing model usually does not bind coverage or set a price. It shapes who sees an offer and how it is framed. That can still be a problem if the model steers certain consumers away from products, reinforces existing disparities, or uses data in ways the consumer did not expect.

The governance question for marketing AI is consent and fairness in the funnel. Are the data sources used for targeting permissible? Does the model create a protected-class skew in who is invited to apply? Is the offer presented in a way that could be misleading? These questions sit at the intersection of insurance law, consumer protection law, and emerging state AI rules. For now, most carriers treat marketing AI as a lower-priority governance item. That is reasonable only if the model does not influence coverage or pricing decisions. Once it does, it moves up the risk stack.

Distribution AI also raises producer-side questions. Tools that recommend which agents should handle which leads can affect service quality. If the recommendation engine is opaque, it can concentrate the best leads with a subset of producers and leave others with harder-to-serve customers. The governance task is not to stop the tool but to monitor its effects on service outcomes and producer compliance.

How to use this map inside your own program

The first step is to classify your models by the operation they serve, not by the technique they use. Two models with the same algorithm belong in different governance buckets if one scores marketing leads and the other recommends claim denials. The risk-tiering logic in your AI Systems Program should start with the business-line consequence, then add model complexity and third-party data as secondary factors.

Next, match the governance depth to the risk. Claims and underwriting models need pre-deployment testing, adverse-outcome monitoring, and documented human review. Health utilization models need clinical oversight and vendor diligence. Fraud and marketing models need monitoring and impact analysis, but the cadence and depth can be lighter. A single standard applied to every model is either too expensive for the low-risk ones or too weak for the high-risk ones.

Finally, keep the map alive. Models migrate. A marketing tool can become an underwriting input. A fraud flag can be used to triage claims. A third-party vendor can change its model without telling you. Quarterly reviews of where each model is used, and what it touches, are the only way to prevent governance drift.

---

Every business line answers to the same governance framework; what changes is how hard each one is scrutinized. For the rules and the program behind these use cases, see the AI governance in insurance guide. We expand this guide as we publish deeper analysis on each line, and the InsureAI Wire dispatch tracks the developments weekly.