What should an AI-generated denial explanation establish?

An AI-generated prior authorization denial explanation should identify the actual reason for the coverage decision, link that reason to the applicable coverage criteria and the patient-specific evidence, and make clear what a provider or patient can do next. A fluent narrative is not evidence that the underlying decision was correct. A clinically accurate explanation requires a verifiable decision record and review of the information on which the determination rests. ^[1][3]

This distinction matters when generative AI is used to summarize clinical notes, retrieve policy language, draft letters, or prepare appeal correspondence. Those tasks may reduce administrative work. They may also turn an incomplete or erroneous determination into an apparently authoritative explanation. The critical question is whether each consequential claim can be traced back to the record the responsible reviewer actually evaluated. The parallel problem in financial services is discussed in Explainable AI in Finance; the healthcare evidence and legal tests are materially different.

Prior authorization is only one form of utilization management. Prior authorization occurs before a requested service; concurrent review assesses continuing coverage during care; claims adjudication addresses payment after care. These workflows have different procedural rules. AI-generated text should not obscure which decision has been made. ^[9]

What do CMS requirements actually cover in 2026?

There is no single federal requirement that every private health insurance denial be explained by the same method. The relevant rule depends on the payer, benefit, requested item or service, and type of decision. Two sets of requirements are especially important for AI-assisted coverage review.

CMS-0057-F: specific reasons for prior authorization denials

Beginning in 2026, the CMS Interoperability and Prior Authorization Final Rule requires impacted payers to communicate a specific reason when denying prior authorization for covered medical items and services. Its scope includes Medicare Advantage organizations, relevant Medicaid and CHIP fee-for-service and managed-care programs, and qualified health plan issuers on the federally facilitated exchanges. This specific provision excludes drug prior authorizations. It does not prescribe a particular language model, explanation algorithm, or LLM validation protocol. ^[1]

The same rule shortened standard decision timeframes to no more than seven calendar days and expedited requests to no more than 72 hours for affected payers other than qualified health plans on federally facilitated exchanges, whose timeframes were not changed by this provision. Program-specific extensions and the requirement to respond as quickly as the patient’s condition requires still matter. The new interoperability APIs have a generally later implementation date, beginning in 2027; they should not be confused with the 2026 operational requirements. ^[1][2]

Medicare Advantage: patient-specific determinations

For Medicare Advantage, 42 CFR § 422.101(c) requires medical-necessity determinations to consider the beneficiary’s medical history, physician recommendations, clinical notes, and applicable coverage rules. CMS has clarified that an algorithm may assist but a population-level prediction cannot replace an assessment of the individual patient’s circumstances. In CMS’s post-acute-care example, predicted length of stay alone is insufficient to terminate coverage; the patient’s condition must be reassessed against applicable criteria. ^[3][4]

When the plan expects to issue an adverse medical-necessity determination, 42 CFR § 422.566(d) requires review by a physician or other appropriate healthcare professional with the specified expertise and licensure. Medicare Advantage denial notices must also state specific reasons and provide applicable reconsideration information. These requirements are more precise than a general instruction to keep a human “in the loop.” Medicare Advantage plans are also required to maintain utilization-management committee oversight of policies. ^[5][6][7]

CMS’s separate WISeR demonstration, launched in six states in 2026, applies to selected Original Medicare services and explicitly does not apply to Medicare Advantage. Its use of technology alongside licensed clinical review is relevant to the broader discussion, but its rules should not be substituted for Medicare Advantage obligations. ^[14]

Prediction, coverage determination, and narrative are different acts

Some utilization-management systems classify requests as likely approvable, extract clinical evidence, or estimate an expected length of stay. A separate workflow applies coverage criteria and issues an approval, request for information, or denial. A language model may then draft the explanation. These components can share data, but they do not provide equivalent evidence. Research on AI use in utilization management distinguishes predictive models from generative systems that compose text, and documents the risks of treating automated output as an adequate substitute for oversight. ^[9][10]

A common design error is to allow a text generator to infer why a request was denied from the model’s predicted outcome. A classifier’s recommendation is not necessarily the final basis for the decision. A policy exception, additional clinical document, human correction, or different coverage criterion may change the outcome. In such cases, the generated reason must follow the actual authorized determination, not whichever intermediary score is easiest to access.

The 2026 Health Affairs analysis of utilization review highlights the practical problem: AI may help both payers and providers with documentation and review, yet automation bias and inadequate understanding of system limitations can undermine oversight. ^[9]

How can an AI-written denial rationale fail?

Generative AI can produce grammatically correct, clinically plausible text without verifying that each asserted fact was present in the case or that each cited rule controlled the decision. NIST’s Generative AI Profile identifies fabricated information and reasoning, including false citations, as a material risk; healthcare-specific reviews describe incorrect patient-context summaries and unsupported medical statements. ^[12][13]

Missing evidence

A clinician’s updated note or a functional assessment is omitted, making the record appear less supportive than it is.

Invented clinical detail

The denial text asserts improvement, independent mobility, or a completed therapy goal that was never documented.

Wrong policy basis

The explanation cites an outdated guideline or an internal rule that does not apply to the beneficiary, service, or geographic area.

Decision substitution

A forecast of average recovery duration is presented as if it conclusively determines an individual patient’s eligibility.

Review mismatch

The reviewer approved a different rationale, but the generated notice introduces an extra ground that no authorized reviewer considered.

Unusable specificity

The notice says only that care was “not medically necessary,” without identifying the operative facts or criteria.

Not every failure originates in the LLM. An incorrect extract, incomplete claim file, misconfigured policy library, or unsupported decision can produce a technically faithful explanation of a faulty determination. Explanation validation must therefore examine both the prose and the underlying process. A 2026 systematic review located only 16 eligible studies of AI in real prior-authorization workflows and found limited external validation, subgroup reporting, and patient-centered outcome evidence. This research does not establish that all AI-enabled denials are inaccurate; it demonstrates why strong assurances based solely on technical benchmark scores would be premature. ^[11]

How should an organization verify the explanation?

Verification should begin at the final coverage determination and work backward through the source record. A groundedness check against supplied documents is useful, but insufficient: the documents themselves may be incomplete, the model may cite the wrong criterion, or the narrative may misrepresent what the actual reviewer decided.

1. Fix the decision identity. Record the request, member/encounter reference, service, decision type, date, authorized decision, and relevant versioned system output. Separate preauthorization from continued-stay review and claims payment.
2. Identify the legally applicable criteria. Preserve the exact benefit provision, coverage rule, or permitted internal criterion used for the determination, including its effective date and jurisdiction.
3. Trace the patient-specific facts. Link each material clinical assertion to a dated record excerpt, recommendation, assessment, or other admissible source. Mark missing or conflicting evidence as unresolved rather than filling gaps by inference.
4. Compare explanation with the decision record. Check whether the prose accurately states the actual grounds, omits a material countervailing fact, invents a clinical conclusion, or substitutes a predictive score for a coverage criterion.
5. Require qualified review where applicable. Show what the reviewer examined, whether they could change the outcome, how discrepancies were handled, and whether the issued reason matches the approved determination.
6. Preserve and test the final notice. Archive the exact communicated wording, delivery route, appeal/reconsideration information, system versions, identified exceptions, and correction history. Sample both initial and appealed decisions.

These are proposed operational verification practices, not a new CMS-prescribed checklist. Their purpose is to make statutory and program-specific requirements demonstrable in an AI-assisted workflow. A useful distinction is between factual support (is a claim in the record?), decision fidelity (was this a real ground of the determination?), and communicative adequacy (is the explanation sufficiently specific and understandable?).

Worked example: a post-acute care denial that the record does not support

Consider a fictional Medicare Advantage beneficiary receiving post-acute rehabilitation following a stroke. A utilization-management tool forecasts that patients with broadly similar characteristics generally recover enough function to leave the facility within a particular period. During concurrent review, a draft coverage rationale is prepared recommending termination of the current level of care.

The beneficiary’s latest assessment, however, documents continued assistance required for transfers, an unresolved swallowing concern, and a treating clinician’s recommendation for additional supervised rehabilitation. A language model produces this draft:

This is not a claim that the patient necessarily qualifies for continued care. It demonstrates why a length-of-stay estimate cannot itself establish medical necessity or the reason for a denial. CMS addressed essentially this category of misuse in its February 2024 guidance: population-based forecasts may assist, but cannot alone justify ending post-acute services under Medicare Advantage rules. ^[3][4]

A separate issue arises when an insurer drafts a response to a provider’s new prior authorization request: the same traceability principles apply, but the governing process and notice requirements must be those for prospective authorization, not the termination-of-ongoing-care procedure illustrated here.

What does meaningful human clinical review involve?

Under Medicare Advantage’s organization-determination rules, a proposed partially or fully adverse medical-necessity decision must be reviewed by an appropriate licensed healthcare professional with relevant expertise before issuance. The rule does not mean every administrative workflow in every insurance market requires the same type of physician sign-off. ^[5]

For AI-supported denials, the reviewer needs more than a binary “approve” button. They need access to the source facts, actual coverage criteria, prior authorization history where relevant, clinician recommendations, conflicting assessments, and any system-generated recommendation. They must be able to identify missing information and change or withhold the proposed determination when warranted. When the approved outcome changes, the generated explanation must be regenerated or corrected against the final record.

Where a beneficiary appeals a Medicare Advantage adverse medical-necessity determination, the reconsideration rules prescribe additional review separation and clinical expertise. Review should not merely repeat the first denial letter; the reviewer must consider the case under the applicable reconsideration process. ^[15][16]

Evidence from 2026 research on utilization review raises questions about automation bias and the limits of superficial human oversight. An organization should therefore evaluate not only whether a reviewer was assigned but also whether that reviewer had the information, time, competence, and authority to exercise judgment. ^[9][10]

What belongs in an auditable explanation record?

Organizations can avoid building a second documentation system by connecting the source record, coverage determination, generated draft, and issued communication. The following is a practical record design for auditors, clinical governance teams, and utilization-management leaders—not a representation that CMS requires these exact field names.

Case identity and scope

Member/claim identifier, plan, service, request category, decision stage, submission and response timestamps.

Patient-specific evidence

Versioned clinical records, treating-provider recommendations, missing documentation requests, contradictory findings.

Controlling criteria

Source and effective version of Medicare coverage criteria, permitted internal policies, and relevant benefit terms.

Decision provenance

Final determination, responsible reviewer, reviewer qualifications where required, and recorded decision rationale.

AI participation

Which model extracted data, scored risk, drafted text, or assisted review; its relevant version and inputs.

Explanation validation

Claims checked, unsupported statements rejected, corrections made, reviewer acceptance, and unresolved exceptions.

Communication and appeal

Exact issued notice, date, delivery channel, recipient, appeal/reconsideration information, later amendments.

Such records can support analysis of whether a disputed explanation is inaccurate, whether the original coverage decision was flawed, or both. Those are related but separate findings. HHS-OIG’s 2022 review reported that 13% of denied Medicare Advantage prior authorization requests in its sample met Medicare coverage rules; this was not an estimate of AI-generated denials. It illustrates why examination of the underlying decision—not merely the language of the notice—is indispensable. ^[8]

Record access also needs privacy safeguards. If an external AI service needs access to protected health information to perform a service for a HIPAA covered entity, the vendor may be a business associate requiring an appropriate agreement; providing software without access to protected information is not, by itself, sufficient to make a vendor a business associate. ^[17]

How should payers and providers measure performance?

Counting the number of letters generated or the speed of drafting does not show whether a denial is clinically justified. Evaluation needs to include explanation quality, reviewer performance, and the patient’s ability to navigate the decision. The 2026 systematic review of AI in prior authorization found that available studies focused heavily on technical metrics; real-world outcomes and subgroup analyses remained comparatively thin. ^[11]

Unsupported-claim rate

Share of sampled generated notices containing a material clinical assertion without source evidence.

Decision-to-notice consistency

Share of issued explanations accurately reflecting the approved coverage grounds and relevant criteria.

Clinical override effectiveness

Cases in which qualified reviewers identify, correct, or reject unsupported machine recommendations or text.

Documentation-gap detection

Frequency with which workflows flag missing records before a determination rather than treating absence as an adverse fact.

Appeal and reversal analysis

Patterns in reconsiderations, overturned denials, recurrent rationale defects, and potentially delayed access to care.

Time and subgroup outcomes

Decision timeliness, patient/provider burden, and differences in outcomes across relevant populations and service types.

Sampling should include both automated and manually reviewed cases. A low overall error rate may conceal problems concentrated in a particular procedure, provider setting, diagnosis group, or population. Where the plan uses delegated utilization-management vendors, the underlying evidence and criteria should remain accessible to the accountable organization; CMS also began collecting Medicare Advantage internal coverage criteria information in 2026. ^[18]

The operating question is whether automation makes the process demonstrably more accurate, timely, and contestable. Until the evaluation supports that claim, efficiency gains and assurance of decision quality should be reported separately.