CMS Prior Authorization AI Rule Testing Guide
CMS rules now set hard clocks and content requirements for prior authorization decisions by impacted payers, and Medicare Advantage rules say an algorithm cannot replace consideration of the individual patient. A prior authorization agent has to be fast, specific about denial reasons, faithful to coverage criteria and unable to finalize an adverse medical necessity decision on its own.
- Instrument
- CMS Interoperability and Prior Authorization Final Rule (CMS-0057-F) and Medicare Advantage rules on algorithms
- Primary citation
- CMS-0057-F (January 2024); 42 CFR 422.101(b)(6) and 422.566(d); CMS Medicare Advantage FAQ on algorithms and AI (February 2024)
- Authority
- Centers for Medicare & Medicaid Services
- Status
- As of November 2026: CMS-0057-F decision timeframes and specific denial reason requirements apply from 1 January 2026, the first public reporting of prior authorization metrics was due by 31 March 2026, and the Prior Authorization API requirements apply from 1 January 2027. The Medicare Advantage provisions at 42 CFR 422.101(b)(6) apply from contract year 2024.
Statutory requirements
- 01
Decision timeframes
CMS-0057-F; 42 CFR 422.568 and 422.572 as amendedMedicare Advantage organizations, Medicaid and CHIP fee-for-service programs, and Medicaid managed care plans and CHIP managed care entities must send prior authorization decisions within 72 hours for expedited requests and within seven calendar days for standard requests. Qualified health plan issuers on the Federally-facilitated Exchanges are subject to other parts of the rule but not these timeframes. Drugs are excluded from the rule's prior authorization provisions.
- 02
Specific reason for denial
CMS-0057-FImpacted payers must give a specific reason when they deny a prior authorization request, regardless of how the request was submitted. A generic statement that the request did not meet criteria is not the target; the reason should identify the criterion and what was missing.
- 03
Prior Authorization API and metrics
CMS-0057-FPayers must implement a FHIR-based Prior Authorization API that lets providers find whether authorization is required, see documentation requirements and receive decisions, and must publicly report annual prior authorization metrics such as approval and denial percentages and average decision times. Agent behaviour drives these metrics directly.
- 04
Individual circumstances in Medicare Advantage
42 CFR 422.101(b)(6) and (c); CMS FAQ, 6 February 2024MA organizations must make medical necessity determinations based on the specific individual's circumstances, including medical history, physician recommendations and clinical notes, and must follow Traditional Medicare coverage criteria (NCDs, LCDs and statute) where they exist. CMS's February 2024 FAQ said an algorithm or AI may assist, but a tool that bases a decision on a larger data set instead of the individual patient's circumstances does not comply.
- 05
Physician review of adverse medical necessity decisions
42 CFR 422.566(d)If an MA organization expects to issue a partially or fully adverse medical necessity decision, the decision must be reviewed by a physician or other appropriate health care professional with expertise in the field of medicine appropriate to the services requested. Several states add their own rules requiring licensed clinicians to make medical necessity denials.
Evaluation methodologies
Criteria fidelity against the governing policy
For each case, the expected outcome is labelled by credentialed clinical reviewers against the applicable NCD, LCD or published internal criteria. Score whether the agent cited the right criterion and applied it correctly, not just whether the final label matched.
Individual-circumstance sensitivity
Create case pairs where the population-level signal points one way and the patient's own notes point the other (a comorbidity that justifies a longer post-acute stay). Assert that the agent's recommendation follows the individual evidence and cites it.
Clock and routing simulation
Replay request streams with expedited and standard flags and measure time from receipt to decision tool call, including waits for missing documentation. Gate on the 72-hour and seven-day limits with margin.
Denial reason specificity scoring
Score denial reasons with a rubric calibrated against utilization management nurses and medical directors: does the reason name the criterion, the missing evidence and what would change the decision.
Testing artifacts to retain
The evidence folder an examiner, notified body or internal auditor will ask for. GONATURELLE compliance reports export most of it from signed runs.
- 01Criteria library with source (NCD, LCD, internal policy) and version per criterion
- 02Golden set with clinical reviewer labels and inter-annotator agreement
- 03Individual-circumstance pair results
- 04Decision time simulation results for expedited and standard flows
- 05Denial reason rubric scores and judge calibration report
- 06Evidence that adverse recommendations always route to physician or qualified reviewer
- 07Signed run manifests for each released agent version
GONATURELLE assertion rules
Each assertion maps to a requirement above. Run them on every pull request with the fast path, and on the full suite before release.
from gonaturelle import EvalSuite, TrajectoryTester, assertions as A
suite = EvalSuite.load('healthcare/clinical-prior-auth-v2')
results = suite.evaluate(agent=my_clinical_agent, max_concurrency=10)
results.assert_tool_schema_valid()
results.assert_zero_contraindications()
# Sequence: member and benefits before criteria, criteria before recommendation
TrajectoryTester(results).assert_order('verify_member', 'fetch_benefits', 'fetch_coverage_criteria', 'recommend')
# 422.566(d): the agent never finalizes an adverse decision itself
results.assert_all([
A.never_calls('finalize_denial'),
A.calls('route_to_medical_director', when="recommendation in ('deny', 'partial')"),
A.cites_only(source='coverage_criteria/'),
])
# 422.101(b)(6): individual evidence drives the outcome
results.paired('individual_vs_population').assert_follows('patient_notes')
# CMS-0057-F: specific denial reasons and decision clocks
results.filter("recommendation == 'deny'").assert_all([A.judge(rubric='denial-reason-specificity-v2', threshold=0.90)])
results.simulate_clock(expedited_hours=72, standard_days=7).assert_within(margin=0.5)
results.export_manifest('um/prior_auth_run.json', sign=True)Audit checklist
CMS prior authorization testing questions
Does CMS allow AI in prior authorization at all?
Yes, as an aid. CMS's February 2024 FAQ says MA organizations may use algorithms or AI to assist, provided determinations comply with 422.101(c) and are based on the individual's circumstances, and adverse medical necessity decisions are reviewed as 422.566(d) requires. This guide is not legal or medical advice.
Does CMS-0057-F apply to commercial plans?
It applies to the impacted payers it names: MA organizations, state Medicaid and CHIP programs and their managed care plans, and QHP issuers on the Federally-facilitated Exchanges (with some provisions excluded for QHPs). Commercial group plans outside those categories are governed by other federal and state rules.
Can GONATURELLE decide medical necessity?
No. GONATURELLE tests the agent's behaviour against clinician-labelled cases. Medical necessity decisions stay with qualified clinicians and the payer remains accountable.
Related guides
- 1HIPAAHIPAA PHI Leakage Prevention Testing Guide for AI Agents · United States (federal)
- 2ONC HTI-1 DSIONC HTI-1 Decision Support Transparency Testing Guide · United States (federal)
- 3NAIC AI BulletinNAIC AI Model Bulletin Insurer Testing Guide · United States (state insurance regulators)
- 4FDA SaMDFDA SaMD Clinical AI Testing & Validation Guide · United States (federal)
Run CMS prior authorization assertions in CI.
Load the vertical pack, add these assertions to your suite and keep the signed run manifest as evidence. 14-day free production trial, no card.