保险业人工智能

Decisions affecting policyholders must account for accuracy, unfair discrimination, explainability, data provenance, and applicable state requirements. A model’s predictive score is not itself a complete underwriting decision.

主要要点

• Define decision context and jurisdiction.
• Test outcomes and disparities, not only model score.
• Maintain governance, explanations, and appeal records.

Define the line of business, decision, and information available at the time. Claims images, telematics, credit-related data, and third-party scores can have different permissions and error patterns. Check whether a feature is a legitimate measure of risk or a proxy for protected or irrelevant characteristics. The NAIC Model Bulletin says decisions supported by AI must comply with applicable insurance laws and regulations, including unfair-trade and unfair-discrimination rules. It also expects governance and information that regulators may request. Treat the bulletin as a framework to organize a current, jurisdiction-specific review. Evaluate error rates and outcomes by relevant groups and claim conditions. Monitor appeals, overrides, complaints, and changes in the data source. A lower fraud-payment rate may reflect more wrongful denials rather than better detection. Keep records of model versions, vendor data, reasons, human review, and corrective action. Provide a path for a policyholder to ask questions and challenge an outcome where required.

1. Imagine a pricing model uses a feature highly correlated with neighborhood boundaries.
2. Test whether the feature adds legitimate risk information and how outcomes differ across affected groups.
3. Remove or govern the feature if it creates an unjustified disparity, then re-evaluate the complete pricing workflow.

The hypothetical review shows why feature usefulness and fairness need separate analysis.