Only One Document on File for the NAIC's Prior Authorization Meeting Mentions AI

As of the morning of August 31, four documents stood on the NAIC’s meeting page for the Regulatory Framework (B) Task Force’s prior authorization session. The session was set for 2:00 p.m. Eastern under Maryland Commissioner Marie Grant, carrying the discussion over from the task force’s June 29 meeting. Artificial intelligence appears in exactly one of the four: the deck filed by three of the NAIC’s consumer representatives, which names AI five times across four of its eighteen pages, twice in the title of a report it cites.

Carl Schmid of the HIV and Hepatitis Policy Institute, Wayne Turner of the National Health Law Program, and the consumer advocate Harry Ting call the presentation “Consumers’ Perspective: The Need for Greater Transparency and Accountability for Prior Authorization.” One slide, on clinical criteria for prior authorization and other utilization management, opens with the line “State regulators can require insurers to:” and lists “Ensure that plans and AI tools properly apply criteria” second among its bullets. A later slide headed “Is prior authorization appropriate?” carries “Pre-launch and ongoing testing of AI tools used for PA” directly above “Meaningful human in the loop,” a bullet that points the reader to ProPublica reporting.

The other three documents do not use the term. The agenda, the AHIP and Blue Cross Blue Shield Association material on the industry’s prior authorization initiative, and the American Hospital Association’s slides return nothing for artificial intelligence, algorithm, or machine learning. The AHA deck does carry the word automation once, in the line “Electronic Standard: Automation or status quo?” That line sits in a discussion of electronic prior authorization transaction standards, which on our reading is not a reference to automated decision-making or to AI.

One limit on that count matters, because the filed record and the meeting are different things. The agenda lists the American Medical Association presenting alongside the consumer representatives and the AHA, and no AMA material was on the page that morning. The provable statement is therefore narrow. It describes what was on file that morning, and says nothing about what anyone argued in the session itself.

One recommendation is aimed at the NAIC itself. On its recommendations slide the deck asks the NAIC to “Partner with H Committee on best practices on AI and prior authorization,” alongside five other asks. The rest of that slide calls on the NAIC to establish a prior authorization working group, conduct individual state market conduct exams, and increase data transparency, for example through state MCAS data. It also asks the NAIC to partner with the D Committee to develop a new working group and market standards, and to monitor federal actions and industry voluntary commitments.

The H Committee in that ask is the Innovation, Cybersecurity, and Technology (H) Committee. Its Big Data and Artificial Intelligence (H) Working Group is revising the AI Risk Evaluation Supplement, with two exposure periods planned for this fall. Our inference, and not a step either the task force or the working group has announced, is that a route between them would put prior-authorization-specific expectations inside an instrument examiners are already piloting under a new name.

None of this carries force, and the deck does not pretend otherwise. It is an advocacy document. A bullet on its second slide reads “Voluntary measures are helpful but fall short and are no substitute for enforceable consumer protections.” A later slide sets out its objections, among them no enforcement or penalties, no standard reporting requirements, and many commitments already sitting in current law. The resources page cites “Artificial Intelligence in Health Insurance: The Use and Regulation of AI in Utilization Management,” a November 2024 National Health Law Program report the consumer representatives prepared themselves. The presenters are citing their own earlier work. The report is not new.

The testing ask would change what a health plan running AI across prior authorization, claims adjudication, and risk adjustment has to keep on hand. The phrase “pre-launch and ongoing” names two things: a pre-deployment validation package, and the production monitoring log and retesting trigger trail that run after it. Holding the first does not produce the second. State law reaches the human end of the same decision. Minnesota bars automated processing alone from carrying an adverse determination without clinician review, effective January 1, 2027. What the consumer representatives put on file asks about the other end. Whether anyone acts on it, the filed record does not say.