LLM Summarization of Recorded Statements in Insurance Claims: Controls and Evidence Required

What regulators expect of LLM use in claims

The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers was adopted in December 2023 as model guidance. States adopt it or act separately. In 2025 NAIC working-group materials, 24 states had adopted the bulletin or pursued legislation, and four more had related activity. The NAIC says the bulletin reminds insurers that decisions or actions made or supported by AI must comply with all applicable insurance laws, sets governance expectations, and describes the information a Department may request in an investigation or examination.

The phrase "made or supported by" matters here. A summary that an adjuster reads before a coverage, liability or fraud decision is AI supporting that decision. Calling the tool assistive does not take it out of scope. It does mean the governance program should name the use case, the accountable owner and the evidence the carrier can produce on request.

NYDFS Insurance Circular Letter No. 7, issued July 11, 2024, is titled for AI systems in underwriting, and its fairness principle is framed around underwriting and pricing. It does not directly address claims, so counsel should confirm how far a carrier applies it there. Its expectation of a corporate governance framework with oversight of AI use, and its reference to the internal audit function under 11 NYCRR § 89.16, are still a sound basis for placing summarization inside internal audit scope.

Colorado SB 24-205 sets developer duties to disclose known or reasonably foreseeable algorithmic-discrimination risks to the attorney general and to known deployers within 90 days. A carrier that deploys a vendor model should expect to receive and act on that kind of risk information. How the statute applies to a specific carrier's claims use, and its current status as of October 2026, should be confirmed with counsel.

Failure modes specific to recorded statements

Summaries of conversational audio can fail in several ways. The model may state facts the speaker never gave. It may omit a material detail, such as a hedge, a correction or an admission. It may assign a statement to the wrong speaker, which is especially serious when claimant, witness and insured give conflicting accounts. Errors in the transcript can pass through as if they were fact. Summaries can also frame a speaker differently depending on accent, dialect or manner of speaking, so bias testing belongs in the evaluation.

A further risk is the input itself. Transcripts are untrusted free text, and NIST AI 600-1 notes that generative AI expands the attack surface and is exposed to attacks such as prompt injection. A transcript could contain text that tries to steer the model. Treating transcript content as untrusted data, and limiting what tools a summarization workflow can call, reduces that exposure.

Mitigations are practical. Use a fixed summary format that attributes each statement and cites its position in the transcript. Sample summaries against the audio on a schedule. Test before deployment on the carrier's own statements, covering omissions, misattribution and differences across speaker groups, and keep evaluating after release. No general accuracy figure substitutes for that testing.

Data handling and human accountability

Recorded statements routinely contain personal and health information. Depending on the carrier's role and the content, HIPAA, state insurance privacy laws and GDPR may apply, and counsel should set the rules for each. The design principle is the same in every case: send the model the minimum it needs. Redact identifiers and health details before the model call where the summary does not require them, keep processing inside approved boundaries, and set retention and data residency deliberately. Self-managed deployment in the carrier's own cloud is one way to keep prompts and evidence within those boundaries.

Human accountability should scale with consequence. A summary used as an orientation aid for an adjuster who then listens to the recording carries less risk than one that feeds a coverage position, a denial or a special investigation referral. Define approval thresholds by action, and require a recorded human approval before a summary is used in the second category.

One speaker at a 2026 NAIC working-group meeting argued that human review of adjuster-note summaries reduces the need for stronger governance, because the adjuster can check the underlying entries. That was one person's view about short note sets, not a regulatory position, and it does not clearly carry over to long audio transcripts. Do not use it to justify fewer controls.

Enforcing policy at runtime and producing evidence

Policy that is reviewed after the fact can show that a violation occurred. Policy enforced in the execution path can prevent it. For summarization, that means each request is evaluated before it reaches the model: allowed, redacted, blocked, or escalated for human approval. Output and any tool calls are evaluated the same way.

Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with actions to allow, redact, block, rate limit or escalate for human approval. Its scope is individual interactions and tool calls. It generates audit-ready evidence from that enforcement, mapped to HIPAA, NAIC, NYDFS, EU AI Act and NIST requirements, and it can be deployed self-managed in the customer's cloud with configurable retention and data residency. It has run continuously in production in a regulated enterprise for more than 18 months, and it serves customers in insurance, pharma, healthcare and education.

The boundary should be stated plainly. Per-interaction evidence does not by itself trace a whole workflow from audio to final decision. The carrier connects those steps through its claim file, using a shared case identifier across the recording, transcript, evidence record and reviewer approval.

For each interaction, the evidence should show the input, the policy applied and its decision, any redactions, the output, and the human approval where one was required. That record is what lets a carrier answer an NAIC-style information request about what the AI was allowed to do and what it did.