The regulatory map every insurer must know

Layer 1: the EU AI act

The EU AI Act (Regulation (EU) 2024/1689), the first comprehensive AI law from a major regulator, entered into force in August 2024 with obligations phasing in through 2027. It classifies AI systems by risk tier: unacceptable (banned), high-risk (heavily regulated), limited-risk (transparency duties), and minimal-risk (largely unregulated).

Insurance pricing and underwriting for life and health insurance are explicitly named as high-risk in Annex III. Property and casualty pricing isn't automatically swept in the same annex language, but claims automation and fraud-detection tools can still qualify if they meet the "significant risk to health, safety or fundamental rights" threshold.

High-risk status under the EU AI Act triggers concrete obligations:

• A risk management system maintained across the model's lifecycle, not just at launch.
• Data governanceData governanceData governance is the set of policies, roles, and processes that ensure data is accurate, secure, well-defined, and used responsibly across an organization.Voir la définition complète →: training data must be relevant, representative, and checked for bias.
• Technical documentation and automatic logging for traceability.
• Human oversight built into the deployment, not bolted on.
• Conformity assessment before market entry, similar in spirit to CE marking for physical products.

Penalties are steep: up to 35 million euros or 7% of global annual turnover for the most serious violations (banned-practice breaches), whichever is higher. Full text and phase-in timeline are on the European Commission's AI Act page.

Layer 2: NAIC model bulletins in the US

The US has no federal AI Act. Instead, the NAIC (National Association of Insurance Commissioners, the standard-setting body that coordinates the 50 state insurance regulators) issues model bulletins: template guidance that states can adopt, adapt, or ignore.

The key document is the NAIC's 2023 Model Bulletin on the Use of Artificial Intelligence Systems by Insurers. It doesn't create new law. It clarifies that existing unfair trade practices laws and rate/form filing requirements already apply to AI-driven decisions. As of early 2026, over 20 states have adopted some version of this bulletin, per NAIC's tracking page.

The bulletin expects insurers to:

• Maintain a written AI governance program, with board or senior-management accountability.
• Document a model's intended use, inputs, and limitations before deployment.
• Run bias testing on outcomes, not just on inputs, since a model can be facially neutral and still produce disparate impact.
• Be ready to explain adverse decisions (a denied claim, a rate increase) to a regulator on request.

This is principles-based, not prescriptive. There's no fixed technical checklist like the EU's conformity assessment. It relies on insurers self-attesting and regulators auditing after the fact.

Layer 3: State insurance commissioner rules

Here's where it gets genuinely fragmented. Insurance is regulated state-by-state in the US (there's no federal insurance regulator equivalent to a banking prudential regulator). Individual commissioners can go further than the NAIC bulletin.

Colorado is the sharpest example. Under its Algorithms and Predictive Models law (SB21-169) and Colorado Division of Insurance regulations, life insurers using external consumer data and algorithms must actively test for unfair discrimination against protected classes, using a specific quantitative testing methodology the state itself published guidance on (see the Colorado DOI AI/ML governance rules).

California's Department of Insurance separately enforces its own unfair claims practices rules against algorithmic claims denials, and has pursued action where automated claims tools were used to deny volumes of claims without individualized review.

New York's DFS (Department of Financial Services) issued its own circular letter on the use of external data and AI in underwriting, requiring insurers to justify that any variable correlated with a protected class is not a proxy for it.

So a single US deployment can face: NAIC-bulletin-style general governance duties, Colorado's quantitative bias testing mandate, New York's proxy-discrimination justification, and California's claims-specific scrutiny, all for the same model, depending on where the policyholder lives.

What this means for a cross-jurisdictional model

Say you build one claims-triage model and want to run it in Germany, Colorado, and Texas.

| Layer | Governs | Core obligation | Enforcement style |

|---|---|---|---|

| EU AI Act | Germany deployment | Conformity assessment, logging, human oversight | Pre-market, documented |

| NAIC bulletin | Baseline US expectation | Governance program, bias testing | Principles-based, post-hoc audit |

| Colorado SB21-169 | Colorado policyholders | Quantitative discrimination testing | Prescriptive, state-specific |

| Texas | Texas policyholders | General unfair-practices law, lighter AI-specific rule | Reactive, complaint-driven |

Practical consequence: you cannot build one compliance file and copy-paste it. You need a base governance layer (documentation, human oversight, bias monitoring) that satisfies the strictest regime, plus jurisdiction-specific modules on top: EU conformity assessment paperwork for Germany, Colorado's specific testing protocol for Colorado risks, general reasonableness documentation for Texas.

A simplified way teams track this in practice:

Simple as this looks, it's the backbone of a real governance system: a per-model, per-jurisdiction obligations registry that compliance and model risk teams both query before any release.

The guardrails that work across all three regimes

Regardless of jurisdiction, four checks show up in every serious insurance AI governance framework:

1. Pre-deployment bias testing on outcomes. Test the model's actual decisions (approve/deny, price tiers) across protected-class proxies (ZIP code, credit-based insurance scores), not just whether protected attributes were fed in as inputs.

2. Explainability at the individual decision level. Regulators in all three regimes expect you to explain *why* a specific policyholder got a specific price or denial, not just aggregate model accuracy.

3. Human-in-the-loop for adverse actions. A fully automated claim denial with no human review is a red flag everywhere, EU AI Act human oversight rules, NAIC governance expectations, and state unfair-practices doctrines all converge here.

4. Continuous monitoring, not one-time validation. Models drift. Data changes. All three regimes expect ongoing monitoring post-launch, not a single certificate at go-live.

Key Takeaways

• The EU AI Act treats life and health insurance pricing as high-risk by name, triggering pre-market conformity assessment, logging, and mandatory human oversight, with penalties up to 7% of global turnover.
• The US has no federal AI Act; the NAIC's model bulletin sets principles-based expectations (governance program, bias testing) that individual states adopt unevenly, over 20 states as of early 2026.
• US state rules can be far more prescriptive than the NAIC baseline: Colorado mandates quantitative discrimination testing for life insurance algorithms, while other states rely on general unfair-practices enforcement.
• A model deployed across jurisdictions needs a base governance layer built to the strictest applicable standard, plus jurisdiction-specific compliance modules, not a single uniform compliance file.
• Four guardrails hold up across every regime: outcome-level bias testing, individual-decision explainability, human review of adverse actions, and continuous post-launch monitoring.