Key takeaways
- The NAIC model bulletin has been adopted state by state, and audits follow adoption. Insurers are being asked four things: the governance program, the inventory, the risk evaluations, and the evidence.
- The inventory comes first. It is a day of work, and every other artefact depends on knowing what AI is actually in use — including the tools embedded in vendor software.
- The audit’s central evidence question is reconstructability: can the insurer show one decision — a claim, a coverage determination — end to end, with approvals and records?
- Agents acting in claims, underwriting, and customer communications are squarely in scope, and they are usually the highest-impact systems in the inventory.
The one-sentence summary
NAIC AI audits ask insurers to prove what the governance program says is happening actually happened — and the preparation sequence is inventory, program, evidence, in that order.
What the NAIC model actually asks
The Model Bulletin on the Use of Artificial Intelligence Systems by Insurers is deliberately short and deliberately structural. It does not tell insurers how to build AI; it tells them how to be accountable for the AI they use. Four obligations:
- A written AI governance program — who governs AI, how systems are evaluated, what the escalation and accountability paths are.
- A risk-based evaluation of AI systems — proportionate to consumer impact, applied across the insurer’s AI use, including third-party and vendor-embedded systems.
- Defined accountability — named owners for AI systems and their outcomes.
- Documentation retained for regulatory review — the programme and its operation must be demonstrable on request.
The fourth obligation is where audits are won and lost, and it is the one insurers underinvest in: the model does not ask for a document describing controls, it asks for evidence the controls ran.
The four artefacts an audit asks for
1. The AI system inventory
The list of AI systems in use: models, agents, third-party tools, and the AI embedded inside vendor software the insurer already runs. In practice this is the audit’s first question and the insurer’s most common surprise — the vendor-embedded AI is usually the largest unrecorded portion of the inventory. Building it is a day of work with the right queries: IAM identities that belong to non-human actors, procurement records for AI-flagged vendors, and workflow owners naming the systems that touch their processes.
2. The written governance program
The document that answers who governs. It should be written around the inventory, not in the abstract: which systems are high-impact, who owns each, how they are evaluated, how exceptions are handled, and how the program changes as systems change. A program written before the inventory describes an organisation that does not exist; auditors notice the mismatch immediately.
3. Risk evaluations, proportionate to impact
Each system gets an evaluation proportionate to what it affects. The claims triage agent that decides which claims get paid gets a deeper evaluation than the summarisation tool nobody acts on. The evaluation should name the system’s decision surface — what it decides, on what data, with what human oversight — and the controls that govern it.
4. The evidence trail
Records showing the governed systems behaved as the program says. This is where the model’s fourth obligation becomes concrete: for the highest-impact systems, each consequential action should have a correlated, tamper-evident record — what was decided, on what inputs, under which policy version, with which human approval, and what happened. One query reconstructs one decision end to end.
Why agents are the hardest part of the inventory
An insurer’s traditional AI — scoring models, document classifiers — makes predictions that humans act on. Agents act. The claims triage agent does not recommend a payment; it initiates one. The underwriting assistant does not flag a risk; it writes to the system of record. Under the NAIC model, the actions are the scope, and agents generate more of them, in more directions, than any earlier automation.
That makes agents the highest-value part of the inventory to instrument first: attributable identity per agent, policy enforced at the gateway rather than trusted to prompts, approval gates on the consequential decisions, and one record per action. The readiness checklist scores exactly these four domains in 36 checks, and the kill-switch drill produces the stop-control evidence most programs are missing.
The preparation sequence
- Inventory. One day. IAM export, vendor AI flags, workflow-owner interviews. Everything else depends on this list being real.
- Program. Written around the inventory, naming owners for every high-impact system. Two weeks, not a quarter.
- Instrument. The highest-impact agents get identity, policy, approvals, and records first. This is the part with real work, and it is the part the audit actually tests.
- Evidence the controls. Run the stop drill, keep the record. Reconstruct one claims decision end to end and confirm it takes minutes, not months.
- Dry-run the four artefacts. Assemble them as a regulator would ask for them. The gaps that surface in a dry run are cheap; the same gaps surfaced by a regulator are findings.
For the cross-framework picture — where the NAIC model sits among the seven frameworks enterprises must align to — the four recurring controls are the same, which is why an instrumentation investment made once serves every framework the insurer faces. The business case one-pager prices the investment for the board, and what Moring does shows the evidence layer implemented as a product.
US banks face a parallel question from a different examiner: what FFIEC examiners are asking about AI agents covers the banking slice. And if the audit lands on the board’s desk first, the AI governance board briefing gives directors the fiduciary framing without the technical detail.
Frequently asked questions
What is the NAIC AI model bulletin?
The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers sets expectations for insurers using AI: a written AI governance program, risk-based evaluation of AI systems, defined accountability, and documentation retained for regulatory review. States adopt it individually, so applicability depends on where the insurer is domiciled and writes business.
What does a NAIC AI audit ask for?
Four artefacts, in practice: the written AI governance program (who governs, who is accountable, how AI systems are evaluated); the AI system inventory (what AI is in use, including third-party systems and tools embedded in vendor software); the risk evaluations for each system, proportionate to its impact on consumers; and the evidence trail — records showing the governed systems behaved as the program says they did.
How do insurers prepare for an AI audit?
Three moves, in order: build the inventory first — it is one day of work and everything else depends on it; write the governance program around the inventory rather than in the abstract, so the document and the reality match; and instrument the highest-impact systems with attributable, recorded actions so the evidence exists before the regulator asks. Insurers that start with the document usually find at audit time that the document describes a program that never operated.
Do AI agents used in claims fall under the NAIC model?
Yes — agents acting in claims triage, underwriting support, and customer communications are AI systems making decisions that affect consumers, which is the model's core scope. Claims agents are typically among the highest-impact systems in the inventory: they decide what gets paid, denied, or escalated, and the record of each decision is the audit's central evidence question.
What is the biggest NAIC AI audit failure mode?
The document-reality gap: a governance program that exists on paper while the AI systems it describes run without attributable records, named owners, or tested controls. Audits fail on evidence, not on prose — the insurer that can reconstruct one claims decision end to end, with the approvals and the record, is in a different position from the one with a 40-page program and no per-action records.
Governing Autonomous AI Agents in Production
Moring provides the runtime control layer, identity verification, human-in-the-loop approvals, and audit records needed to deploy production-grade AI agents safely.
About the author
Balaji Nagaraj KumarChief AI Officer, Moring AI, Moring AISweden
Leads platform and delivery engineering at Moring AI, turning each engagement into a substrate the customer’s own team operates at handoff. His background spans developer-platform leadership at Volvo Cars and engineering at CERN. He writes about governance for agents that take actions in production systems.
LinkedIn profile