Insurance AI Model Risk Management Compliance | Moring
Model risk compliance for insurers means demonstrating, with evidence, that AI models in underwriting, pricing, and claims are inventoried, validated, monitored, and approved per the NAIC model bulletin, the EU AI Act for European exposure, and DORA where applicable. Compliance is now continuous, not an annual attestation.
Short answer
Model risk compliance for insurers means demonstrating, with evidence, that AI models in underwriting, pricing, and claims are inventoried, validated, monitored, and approved per the NAIC model bulletin, the EU AI Act for European exposure, and DORA where applicable. Compliance is now continuous, not an annual attestation.
How it works
Compliance in insurance model risk management has shifted from aspiration to accountability. Supervisors are driving new standards for AI model risk as advanced models move into mission-critical functions, and insurers increasingly need to show working controls rather than written policies. The practical test an examiner or supervisor applies is simple: show the inventory, show the validation, show the monitoring, and show who approved each model for the use it is in.
The NAIC is the anchor in the United States. Its model bulletin on insurer use of artificial intelligence, adopted in December 2023, sets expectations for governance, risk management, and documentation across underwriting, pricing, and claims AI. States reference it when examining carriers, so a compliance program should produce exactly the artifacts it implies: a board-approved AI policy, a model inventory with owners, validation records, third-party diligence, and evidence of ongoing monitoring.
European exposure adds the EU AI Act and, for insurers and reinsurers authorized under Solvency II, DORA. The AI Act, Regulation (EU) 2024/1689, imposes risk-tiered obligations, with high-risk systems carrying duties for data governance, human oversight, logging, and accuracy. DORA frames AI within ICT risk management, adding register entries for third-party providers and resilience testing. Carriers operating across jurisdictions should maintain one control set mapped to all three expectations rather than separate programs.
Building the compliance program follows a familiar sequence. First, inventory every model, including AI embedded in vendor software, because accountability follows the carrier regardless of who built the model. Second, tier models by decision consequence and set validation depth accordingly. Third, implement monitoring for drift, performance decay, and fairness metrics with defined breach responses. Fourth, retain evidence in a form that can be reconstructed on demand. Fifth, report status to the board or its risk committee on a fixed cadence.
Common compliance failures are organizational rather than technical. Carriers discover vendor AI missing from the inventory, validation reports that predate material model changes, monitoring dashboards no one is accountable for acting on, and evidence scattered across teams so reconstruction takes weeks. Guidance for financial institutions on AI model risk consistently flags these same gaps; closing them is less about new tooling than about decision rights and ownership discipline.
Treat compliance as a byproduct of a working model risk system, not a parallel exercise. When inventory, tiering, validation, monitoring, and approval are genuinely operating, the regulatory artifacts fall out automatically, and the conversation with supervisors shifts from demonstrating effort to demonstrating outcomes.
Frequently asked questions
What does the NAIC expect from insurers on AI model risk?
The NAIC model bulletin on insurer use of artificial intelligence, adopted in December 2023, expects governance, risk management, and documentation for AI used across underwriting, pricing, and claims. Practically, carriers should maintain a board-approved AI policy, a model inventory with named owners, validation records, third-party diligence, and evidence of ongoing monitoring, because states reference these expectations when examining insurance carriers.
How do the EU AI Act and DORA affect insurer model risk compliance?
The EU AI Act, Regulation (EU) 2024/1689, applies risk-tiered obligations, with high-risk systems subject to duties on data governance, human oversight, logging, and accuracy. DORA applies to insurers and reinsurers authorized under Solvency II and frames AI within ICT risk management, adding third-party register entries and resilience testing. A single internal model risk control set mapped to both, plus the NAIC bulletin, avoids parallel compliance programs.
What are the most common model risk compliance failures in insurance?
The recurring failures are organizational: vendor AI missing from the inventory, validation reports that predate material model changes, monitoring dashboards without a named owner acting on them, and evidence scattered across teams so examiner reconstruction takes weeks. Closing these gaps depends less on new tooling than on explicit decision rights and disciplined model risk ownership across actuarial, data science, risk, and compliance.
How can an insurer make model risk compliance continuous?
Make compliance a byproduct of the operating control system. When inventory, tiering, validation, monitoring, and approval genuinely run, the artifacts supervisors want are produced automatically by the model risk system and stay current. Add a fixed board reporting cadence, define breach responses for monitoring thresholds, and review the control set whenever regulation changes, so compliance reflects live model behavior rather than an annual attestation.