Keeping Human Review of AI Prior Authorization Recommendations Within CMS Decision Timeframes
Treat the CMS decision deadline as a fixed engineering constraint. Stamp every request at receipt, subtract intake, AI processing, notice delivery, and a safety margin to set a maximum review dwell time, and keep AI output in a recommendation state that cannot take effect until a qualified reviewer acts. Define escalation triggers and fallback behavior in advance, and retain per-request evidence of the recommendation, the policy applied, the reviewer, the timestamps, and the outcome.
Work backward from the deadline to a review time budget
- 1
1. Stamp receipt time
Record the receipt timestamp first, since every deadline calculation derives from it.
- 2
2. Classify the request
Assign program, standard or expedited, and drug or non-drug, so the correct deadline logic applies.
- 3
3. Compute the review budget
Deadline minus intake, AI processing, notice delivery, and margin gives the maximum dwell time.
- 4
4. Alert at fractions of the budget
Monitor elapsed time per open request and alert at defined fractions, such as half and three quarters.
- 5
5. Record decision and notice times
The same data supports the average and median decision-time metrics that 42 CFR 422.122 lists.
Evidence to retain for each AI-assisted recommendation
- The AI recommendation and the basis for it
- The policy or criteria applied, and the escalation decision with its trigger
- Reviewer identity, role, and action, including any override of the recommendation
- Timestamps for receipt, escalation, review, decision, and notice
- The outcome, the specific denial reason where applicable, and any extension used
- Fallback events, such as re-routing or deadline alerts, and who responded
The CMS clock is the fixed input
Under the 2024 CMS Interoperability and Prior Authorization final rule, impacted payers must decide standard requests within 7 calendar days and expedited requests within 72 hours, for non-drug items and services. The rule covers Medicare Advantage organizations, state Medicaid and CHIP fee-for-service programs, Medicaid managed care plans, and CHIP managed care entities. An extension of up to 14 days is possible in certain circumstances. As of November 2026, drug requests follow separate clocks. For example, MA plans must answer expedited Part B drug requests within 24 hours and non-expedited requests within 72 hours.
These periods run from the payer's receipt of the request, so every minute in intake, AI processing, and the review queue counts against the deadline. Notice of the decision is also required within the period. For MA, a missed deadline on an organization determination is treated as a denial that can be appealed. A slow review step therefore produces an adverse outcome that no clinician actually decided.
CMS did not directly address AI when implementing these provisions. It encouraged innovation that includes medical professional judgment in coverage decisions and that sits under a governance structure covering transparency, evaluation, and ongoing monitoring. CMS has also noted that before an MA organization issues an adverse decision, a physician or other appropriate healthcare professional with expertise reviews whether the service is medically necessary. The design task is to make that review real and still finish on time.
Define escalation triggers and keep the AI output provisional
The sources do not prescribe escalation triggers, so each organization has to define and document its own. A defensible starting set is any recommended denial, any case with incomplete information, and any case where the model's confidence falls below a threshold you have validated. A stricter design routes every recommendation to a human and uses triggers only to set priority. Either choice should be written down, approved by compliance, and tied to the reviewer role.
The more important control is structural. The AI output should exist in a recommendation state that cannot become a determination until a qualified reviewer acts. If the system can finalize on its own and flag the case afterward, review has become an audit exercise rather than a decision step. Assign the reviewer role deliberately: for adverse medical necessity decisions the expectation points to a physician or another appropriate healthcare professional with expertise.
A review can still be reduced to a rubber stamp. Show the reviewer the recommendation basis and the policy or criteria applied, so the specific denial reason now required for MA and state Medicaid denials (from January 1, 2026, excluding drugs) can be written from that basis. Then measure whether review is meaningful. Reviewer override rates and review durations are useful indicators. They are a design suggestion, not a regulatory requirement.
Design fallback behavior for capacity and deadline pressure
Review capacity will sometimes be exceeded. Decide in advance what happens when queue dwell time approaches its budget. Reasonable responses include priority re-routing to additional qualified reviewers, reordering the queue by time remaining, and paging a supervisor. Segment queues by program and by standard versus expedited so that a backlog of standard requests does not consume the time of an expedited case.
The constraint on fallback is firm: it must not auto-finalize an adverse outcome. Because a missed MA deadline is treated as a denial, and an automated adverse decision would skip clinician review, both failure modes land in the same place. A fallback may escalate, reassign, or invoke an extension where one is available and justified. It should never convert an unreviewed recommendation into a determination.
The tradeoff is staffing against reserve. A larger safety margin shortens the review window and raises reviewer load. A smaller margin leaves no room for notice delivery. Set the margin using observed processing and notice times, and revisit it when volumes or program mix change.
Where runtime policy enforcement fits
There is a real difference between enforcement in the execution path and review after the fact. In the execution path, an AI recommendation can be held or escalated for human approval before it takes effect. After the fact, the same recommendation has already acted and a reviewer is only reconstructing what happened. For a step where a human must decide before an outcome applies, only the first option meets the intent.
Trussed enforces policy in the execution path at the level of individual interactions and tool calls. Its actions are allow, redact, block, rate limit, and escalate for human approval. In this design, a policy can require that an AI interaction or tool call that would act on a recommendation is escalated for human approval, and that enforcement generates audit-ready evidence of the action taken. It is enforced at runtime, not reviewed after the fact.
The scope is specific. Trussed does not provide transaction-level governance or end-to-end tracing of a prior authorization case across systems. The case record, the clinical criteria, and the deadline clock remain in your utilization management system. Runtime enforcement supplies the control and the evidence at the points where AI interactions and tool calls occur.
Review step inside the decision clock
Receipt
The clock starts when the payer receives the request.
AI recommendation
Held as a recommendation, not a determination.
Qualified human review
Reviewer sees the basis and the policy or criteria applied.
Decision and notice
Issued, with a specific reason if denied, inside the deadline.
Map your review step to the decision clock
Talk to an expert about enforcing escalation for human approval in the execution path and about the audit-ready evidence that enforcement produces.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.