GRC Platform vs Runtime Proxy for NAIC AI Model Bulletin Evidence

Two kinds of bulletin evidence

The NAIC AI Model Bulletin, adopted in December 2023 as principles-based guidance, reminds insurers that decisions or actions affecting consumers that are made or supported by AI must comply with all applicable insurance laws. It expects insurers to develop and maintain a written program for the responsible use of AI. State adoption happens through individual state bulletins, and as of August 2026 the NAIC's status map lists at least 20 state or district bulletins, so multi-state carriers should work from the version in each state where they operate.

The information an examiner may request falls into two groups. The first describes program and design: the written AI Systems program, inventories and descriptions of AI systems, third-party due diligence and vendor contracts. The second describes operation: validation, testing and auditing documentation, including drift evaluation, and the techniques, measurements and thresholds applied. This page treats the first group as attested evidence and the second as needing observation of system behavior. That split is an analytic framing, not regulatory text, but it is a practical way to assign each request to a system that can credibly answer it.

Attested evidence answers "what was the AI allowed to do?" Someone with authority states that a policy exists, a control was designed, and a review was completed. Enforced runtime evidence answers "what did it do?" The system in the execution path records the decision it made on a specific interaction.

What each approach produces, by evidence type

The table is organized by what a regulator may ask for, not by product category. Capabilities described for GRC platforms are the typical scope of the category. Confirm specifics with each vendor.

• Program documentation: GRC platform: strong. Holds the written AI Systems program, policies, ownership and review history. Runtime proxy: not its role. It does not manage the program or assign accountability.
• Control design: GRC platform: strong. Maps controls to bulletin expectations and records risk assessments. Runtime proxy: partial. The policies it enforces are themselves a precise statement of how a control is configured.
• Control operation: GRC platform: depends on attestations and periodic testing entered by people. Runtime proxy: strong for enforced controls, because each decision is recorded as it happens.
• Per-interaction records: GRC platform: does not observe individual interactions or tool calls. Runtime proxy: strong. Records allow, redact, block, rate limit and escalation outcomes per interaction.
• Third-party oversight: GRC platform: strong for due diligence, questionnaires and contract records. Runtime proxy: shows which interactions involved external models or tools and which policy applied, but does not hold contracts or assessments.

Gaps in relying on either alone

A GRC platform on its own can show that a control exists and that an owner attested to it. It cannot show that the control operated on a given interaction. If an examiner asks for evidence that a restriction held in practice, an attestation is a statement about intent and process, and the platform has no independent observation behind it. The gap widens as AI use spreads into agents, tools and coding assistants, where behavior changes faster than quarterly attestation cycles.

A runtime proxy on its own has the opposite gap. It can produce a precise record of enforcement, but a record of enforcement is not a program. It does not state which systems are in the inventory, who owns each one, how risk was assessed, or what was agreed with a vendor. It covers only traffic that passes through it, so systems outside the execution path produce no evidence. An examiner reading runtime records without the written program would see decisions without the framework that explains them.

Both gaps are about the kind of question being asked. Treat each system as authoritative for its own questions and avoid asking either to stand in for the other.

How the two fit together in an examination

In practice, the GRC platform holds the narrative and the runtime proxy holds the proof. The written program states that sensitive data must not reach an external model. The control in the GRC platform references the enforced policy. During an inquiry, the compliance team presents the program record and then produces runtime evidence showing how that policy was applied across interactions in a defined period.

The handoff needs design. Decide which runtime evidence is referenced from control records, who reviews exceptions such as blocked or escalated interactions, and how often. Assign third-party oversight evidence, validation and testing documentation, and runtime evidence to separate named owners. Each has a different cadence and a different audience.

Regulatory references should be dated. For carriers that also operate in Colorado, Senate Bill 26-189, signed in May 2026, repeals and reenacts the earlier SB 24-205 provisions with new requirements on automated decision-making technology in consequential decisions. As of October 2026, review Colorado exposure against SB 26-189 rather than the earlier bill. The same evidence structure, program records plus per-interaction records, applies whichever state requirements govern.

Where Trussed fits

Trussed is not a GRC platform. It does not manage the AI Systems program, ownership, attestations or vendor questionnaires. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants: allow, redact, block, rate limit, or escalate for human approval, at the level of individual interactions and tool calls. Enforcement is at runtime, not reviewed after the fact, and that enforcement generates continuous audit-ready evidence mapped to NAIC and other frameworks.

Deployment is self-managed in the customer's cloud, with configurable retention and data residency, which addresses where evidence lives and how long it is kept. Trussed operates in production for customers in regulated verticals including insurance, and generates 30+ GB of audit evidence per day. Insurers keep their GRC platform as the system of record for the program and use runtime evidence as the operating record behind it. Trussed does not currently provide transaction-level governance, model lineage or end-to-end tracing of agentic workflows, so those questions need other sources.