Insurance AI Model Risk Management Checklist | Moring
A practical insurance model risk checklist covers inventory, risk tiering, validation, independent review, production monitoring, evidence retention, third-party assurance, and board reporting. Each item maps to expectations in the NAIC model bulletin, the EU AI Act for European exposure, and DORA where it applies.
Short answer
A practical insurance model risk checklist covers inventory, risk tiering, validation, independent review, production monitoring, evidence retention, third-party assurance, and board reporting. Each item maps to expectations in the NAIC model bulletin, the EU AI Act for European exposure, and DORA where it applies.
How it works
An inventory comes first on any insurance model risk checklist, because nothing else can be governed if it is not known. The inventory should cover traditional actuarial models, machine learning underwriting and pricing models, claims triage and fraud scoring systems, generative AI assistants, and third-party AI embedded in vendor software. Governance guidance for insurers stresses that carriers are accountable for vendor models too, so purchased AI belongs on the same list with the same ownership fields.
Risk tiering is the second item. Assign each model a tier based on the consequence of its decisions: models feeding rate filings or claim payments sit at the top, advisory copilots in the middle, and internal search or summarization at the bottom. Tiering determines the depth of validation, the frequency of monitoring, and the level of approval required. This mirrors the EU AI Act's logic of proportionate obligations by risk category and keeps scarce validation capacity pointed at the models that can hurt the carrier most.
Validation and independent review form the third block. Each high-tier model needs documented conceptual soundness, data lineage, performance testing against alternatives, and challenge from a function independent of the model's builders. Regulatory commentary on AI model risk is explicit that periodic validation is no longer sufficient for adaptive models, so the checklist should require monitoring thresholds for drift, performance decay, and fairness metrics, with defined actions when thresholds breach.
The fourth block is evidence and accountability. Retain validation reports, approval records, monitoring output, and incident reviews in a form an examiner can reconstruct, an expectation consistent with the NAIC model bulletin's emphasis on governance and documentation for insurer AI use. Where DORA applies, extend evidence to ICT third-party register entries and resilience testing for systems supporting critical insurance functions.
Third-party assurance and human oversight complete the checklist. For every purchased model, require contractual audit rights, performance disclosures, and notification of material model changes. For every consequential decision stream, define where a human approves, reviews on a sample basis, or overrides, and record those interventions as part of the model's ongoing evidence.
A checklist only works if it closes. Assign each item an owner and a review cadence, report status to the board or its risk committee on a fixed schedule, and treat checklist gaps as model risk findings with remediation dates. Carriers that operationalize the list this way convert a static document into the running control system that both supervisors and boards expect.
Frequently asked questions
What should be first on an insurance model risk checklist?
Inventory comes first, because ungoverned models cannot be governed at all. The list should include actuarial models, machine learning pricing and underwriting models, claims triage and fraud systems, generative AI assistants, and third-party AI embedded in vendor software. Since carriers are accountable for vendor models too, purchased AI belongs on the same inventory under the carrier's model risk standards with the same ownership and evidence requirements.
How should an insurer tier models in its risk checklist?
Tier by decision consequence. Models feeding rate filings or claim payments occupy the top tier and require the deepest validation, most frequent monitoring, and senior approval. Advisory copilots sit in a middle tier, and internal summarization tools at the bottom. This mirrors the EU AI Act's proportionate, risk-based structure and concentrates scarce validation capacity on the insurance models most able to harm the carrier.
What monitoring items belong on the checklist?
Require monitoring thresholds for drift, performance decay, and fairness metrics for every adaptive model, with defined actions when thresholds breach. Supervisory commentary on AI model risk is clear that periodic validation alone no longer suffices. In insurance, seasonal portfolio shifts can miscalibrate a claims or pricing model within a quarter, so continuous monitoring with documented escalation is the realistic baseline.
How does evidence retention fit the insurance model risk checklist?
Retain validation reports, approval records, monitoring output, and incident reviews in a form an examiner can reconstruct. The NAIC model bulletin emphasizes governance and documentation for insurer AI use, and where DORA applies the evidence set extends to ICT third-party register entries and resilience testing. Assign every checklist item an owner and a review cadence so the model risk checklist stays live rather than decorative.