Using LLMs to Draft Rate Filing Narratives for State Regulators

Whether and when LLM drafting is permissible

Rate filing narratives include actuarial memoranda, rationale summaries and responses to objections. They are exactly where fluent but wrong text does harm. NAIC states that LLM tools can generate information that sounds accurate but is incorrect, and that AI-generated information should be reviewed carefully, especially for important decisions. That is a reason to constrain the use, not to prohibit it.

A workable position has three conditions. First, drafting is treated as a defined AI use case with its own risk assessment, inside the insurer's written AI program. Second, the workflow limits which models and tools can be used and what data can reach them. Third, a human approves the final text, and the record shows who approved what, and when.

Nothing here removes the insurer's responsibility. The filed narrative is the insurer's statement to the regulator, whether a person or a model produced the first draft.

What the regulatory texts expect, and what is practice

The NAIC AI Model Bulletin was approved by NAIC members on June 11, 2024. It is a guiding document, not a model law or regulation, intended to foster uniform regulator expectations. It reminds insurers that decisions or actions affecting consumers that are made or supported by AI must comply with all applicable insurance laws and regulations. It sets out an expectation that insurers create and implement a written AIS Program commensurate with a risk assessment. NAIC also advises that the bulletin describes the type of information a Department may request in an investigation or examination, and that regulators may ask insurers to explain how AI tools are used in pricing. Adoption and application vary by state, so confirm how each target state treats the bulletin.

NYDFS Insurance Circular Letter No. 7, issued July 11, 2024, is titled for AI use in underwriting. It expects AIS and ECDIS to be used in compliance with all applicable federal and state laws and regulations, and it describes auditing that assesses the effectiveness of the insurer's AIS and ECDIS risk management framework. Colorado SB 24-205 addresses consumer protections for AI, including developer duties to disclose known or foreseeable algorithmic discrimination risks. As of October 2026, whether and how each of these reaches a drafted rate filing narrative is a legal interpretation for counsel.

Keep the control matrix honest about the difference between expectation and practice. The table below shows one way to label them.

• Written AIS Program covering the use case: NAIC expectation. Treat narrative drafting as its own use case with a risk assessment.
• Compliance with insurance law for AI-supported work: NAIC and NYDFS expectation. The filed narrative must meet the same standards as any other filing.
• Ability to explain AI use and support an examination: NAIC and NYDFS expectation. The insurer should be able to show what was used and how it was governed.
• Oversight of third-party models and vendors: Part of the insurer's accountability under the AIS Program. Place vendor tools inside its scope.
• Runtime redaction, blocking and escalation: Good practice. It is not a stated requirement, but it produces the evidence the expectations imply.
• Named human sign-off on filed text: Good practice grounded in NAIC's call for careful review. The sign-off standard is set in internal policy.

Evaluating whether runtime enforcement fits your filing process

Post-hoc review of prompts and outputs can find problems, but it cannot show that a rule was applied to a given interaction at the time. Enforcement in the execution path can, because the allow, redact, block or escalate decision and its record are the same event. That is the argument for enforcing policy at runtime, and it matters most when an exam request arrives long after the filing.

To test fit, ask whether the approach can show, for each filed narrative, which models and tools were permitted and which were used. Ask what happens to non-public information in a prompt, and whether that action is recorded. Ask whether enforcement happens before a model or tool call executes. Ask how an approval is linked to the specific interactions behind the final text, and whether evidence can be exported to answer an exam request under your written AIS Program with retention you control.

Scope matters. Interaction-level and tool-call-level enforcement does not by itself establish model lineage or trace an entire workflow, and it does not replace the written AIS Program or human accountability. Trussed enforces policy in the execution path across models, agents, tools and MCP servers, and produces audit-ready evidence mapped to NAIC and NYDFS expectations. It deploys self-managed in the customer cloud, with configurable retention and data residency. A pilot on one filing type is a sensible way to test any approach against these questions.