Implementation Guide

    Stopping Medicare Advantage AI from Applying Internal Coverage Criteria Beyond CMS Rules

    Medicare Advantage organizations must apply Traditional Medicare coverage criteria and may use internal criteria only in defined circumstances, and only if those criteria are publicly accessible and evidence-based. Policy documents do not stop an AI tool from applying other criteria at the moment of a decision. Compliance teams need to constrain what each AI interaction and tool call may use and do, by enforcing policy in the execution path (allow, redact, block, escalate for human approval), and retain evidence of what was permitted and what happened.

    Where internal criteria can enter an AI decision

    1. 1

      System and prompt instructions

      Instructions written by developers or vendors can embed thresholds, rules of thumb or step requirements that no NCD, LCD, statute or regulation specifies.

    2. 2

      Retrieved documents

      Retrieval indexes may hold vendor guideline libraries or internal policies. Third-party products such as InterQual or MCG are a documented route to internal criteria, and may be used only if the plan meets 422.101(b), (c) and 422.566(d), including the evidence-base requirement.

    3. 3

      Tool and API calls

      A rules engine inside a utilization management platform can return criteria the model then treats as authoritative, without anyone reviewing the rule.

    4. 4

      Model outputs

      An output that proposes a denial or limitation may rest on criteria the model learned or inferred rather than on any approved source.

    Evidence to retain for each decision

    • The approved-criteria register, with version history and the authority for each entry
    • Public postings of internal criteria, with their evidence summaries and rationale
    • Per-interaction and per-tool-call enforcement records: policy version, decision (allow, redact, block, escalate) and identifiers of the sources and tools involved
    • Escalation records, including the reviewer's identity and qualifications, and the decision made
    • The criteria cited in each denial notice, including any internal criteria, which must be clearly stated
    • Records of utilization management committee review of AI-applied criteria and policy rules

    What the CMS rules require of AI-assisted coverage decisions

    Under 42 CFR 422.101, Medicare Advantage (MA) organizations must cover all services covered by Part A and Part B. Coverage criteria come from Traditional Medicare statutes, regulations, national coverage determinations (NCDs) and local coverage determinations (LCDs). The regulation states that an MA organization may not deny coverage for basic benefits based on criteria not specified in paragraphs (b) or (c) of that section.

    The CY2024 final rule, published in April 2023, added a narrow allowance for internal coverage criteria. They are permitted only when coverage criteria are not fully established, which the rule defines in three circumstances: (A) additional, unspecified criteria are needed to interpret or supplement general provisions so that medical necessity is determined consistently; (B) an NCD or LCD explicitly allows coverage beyond its listed indications; or (C) no applicable statute, regulation, NCD or LCD sets coverage criteria.

    Even then, internal criteria must be publicly accessible and based on current evidence in widely used treatment guidelines or clinical literature. The organization must publish the criteria in use, a summary of the evidence considered and the rationale for adopting them. CMS has also said internal criteria cannot add new, unrelated criteria, without supplementary or interpretive value, to an item or service that already has existing but not fully established coverage policies.

    On AI, CMS stated in the CY2026 policy rule discussion (published December 2024) that these requirements already apply when MA plans use AI or automated systems, and that the MA organization is responsible for ensuring AI tools comply with the internal coverage criteria rules. Contracts with model or tool suppliers do not move that accountability. This guidance reflects the position as of November 2026.

    Applying runtime controls in the execution path

    A policy document describes what should happen. A runtime control determines what can happen. The practical approach is to place enforcement in the execution path between the agent and its models, retrieval sources and tools, and to evaluate each interaction and tool call against a defined policy before it proceeds.

    The foundation is an approved-criteria register that maps each permitted source to its authority: statute, regulation, NCD, LCD, or a published internal policy that fits one of the three circumstances above. The register is what policy rules reference. Without it, there is nothing concrete to allow or block against.

    With a register in place, the actions are straightforward. Allow what maps to a register entry. Redact passages from a retrieved document that carry unapproved criteria while leaving approved content available. Block a document, tool response or instruction that is not on the register. Rate limit tool calls that behave outside their defined purpose. Escalate for human approval any output that proposes a denial or limitation.

    Escalation matters for a regulatory reason. A denial based on medical necessity must be reviewed by a physician or other appropriate health care professional with expertise appropriate to the service, so AI-proposed adverse determinations should route to a qualified reviewer rather than finalize automatically. Tool calls should also be bound to a purpose. Prior authorization may only confirm the presence of a diagnosis or other medical criteria that are the basis for coverage, and medical necessity determinations must draw on the enrollee's medical history, physician recommendations and medical notes.

    Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with allow, redact, block, rate limit and escalate for human approval applied to individual interactions and tool calls. The control is enforced at runtime, not reviewed after the fact. The approach described here does not depend on any one product, but it does require enforcement at the interaction and tool-call level.

    Rollout sequence and governance

    Start with an inventory of AI touchpoints in utilization management and coverage workflows: prompts, retrieval indexes, tools and outputs. Classify every criterion source as an NCD, LCD, statute, regulation or published internal criterion. Anything that cannot be classified is a candidate for blocking.

    Next, map each internal criterion to 422.101(b)(6)(i)(A), (B) or (C) and confirm it is published with its evidence summary and rationale. Validate any third-party criteria library against the evidence-base and public-accessibility requirements before exposing it to an AI tool. Then define escalation routing so that adverse medical-necessity determinations reach a qualified reviewer, and confirm that denial notices can cite the criteria actually used.

    From contract year 2024, MA plans must have Utilization Management Committees that review all utilization management policies at least annually. Placing AI-applied criteria and enforcement rules within that review scope is a sensible way to keep the register current, though the regulation does not require it.

    One item to monitor, not to act on: H.R.9734 in the 119th Congress would require MA plans to report prior authorization denials based on AI output. As of November 2026 it has been introduced in the House only and is not law.

    Where Trussed is used, enforcement generates audit-ready evidence continuously, with self-managed deployment in the customer cloud and configurable retention and data residency. Teams should confirm that retention settings meet their own record-keeping obligations.

    Where policy is enforced in an AI-assisted coverage workflow

    Prompts

    Instructions are checked for criteria outside the approved register.

    Retrieval sources

    Documents not on the register are redacted or blocked before reaching the model.

    Tool calls

    Each call is allowed, blocked or rate limited against its defined purpose.

    Model outputs

    Proposed denials are escalated for human approval.

    Constrain what AI can use, and keep the evidence

    Discuss how runtime policy enforcement and audit-ready evidence can apply to your Medicare Advantage utilization management workflows.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.