Implementation Guide

    Medicaid Managed Care Prior Authorization AI Agents: State Timeframe and Notice Controls

    Federal Medicaid managed care rules set the outer limits for prior authorization decision timeframes and require notice of adverse decisions. States and managed care contracts set the operative values, and they may be stricter. A runtime control layer can enforce those values on individual agent interactions and tool calls: escalate any adverse determination to a human reviewer, block or redact outbound notices that lack required elements, and limit which records and tools an agent can reach. It does not replace clock management, clinical review, credentialing, or appeals.

    Placing the policy decision point

    1. Context in: receipt time, request type, state and contract limit
    2. Policy decision point in front of notice and record-request tools
    3. Adverse determinations routed to human approval before sending
    4. Separate policy sets for drug and non-drug workflows

    Evidence to retain for each decision

    • Request receipt time, request type and the timeframe limit applied
    • Decision time and the time each notice was sent or blocked
    • Policy version evaluated and the action taken: allow, redact, block, rate limit or escalate
    • Identity of the human approver for any adverse determination or reduced authorization
    • Content checks run on outbound notices and their results
    • Records and tools the agent accessed during the interaction

    The requirements the agent has to operate within

    For Medicaid managed care rating periods starting on or after January 1, 2026, 42 CFR 438.210 requires standard authorization decisions and notice within state-established timeframes that may not exceed 7 calendar days after receipt of the request. The cap was 14 days before that. A standard decision may be extended by up to 14 additional calendar days when specified conditions are met. Expedited decisions are required when a provider indicates, or the plan determines, that the standard timeframe could seriously jeopardize the enrollee's life, health or ability to attain, maintain or regain maximum function. Those are due as expeditiously as the condition requires and no later than 72 hours after receipt. Covered outpatient drug requests must be answered within 24 hours.

    These are federal ceilings on what a state may allow. State law may set a shorter timeframe, and a managed care contract may be stricter again. The number an agent must respect is therefore the one in the state contract, not the federal figure.

    Two further obligations matter for agent design. A decision to deny a service authorization request, or to authorize a service in an amount, duration or scope less than requested, must be made by an individual with appropriate expertise in the enrollee's medical, behavioral health or long-term services and supports needs. The plan must also notify the requesting provider and give the enrollee written notice of that decision. As of November 2026, a CMS proposed rule published in April 2026 would extend the 24-hour drug timeframe to all drugs and require a specific denial reason for drugs. It is a proposal, so controls should not treat it as settled law.

    What becomes runtime policy and what stays outside the agent

    Runtime enforcement applies to individual interactions and tool calls. The test for each requirement is whether a single agent action can be allowed or stopped based on information available at that moment. The mapping below is an implementation interpretation, not regulatory text.

    • Denial or reduced authorization needs an expert decision: Runtime: hold any agent-drafted denial or partial approval and escalate for human approval before an outbound action is permitted. Outside the agent: deciding who qualifies as an individual with appropriate expertise and verifying it.
    • Written notice to enrollee and notice to provider: Runtime: block or redact outbound notices missing required elements, using state and contract rules as configuration. Outside the agent: authoring the notice templates and legal review of their content.
    • Specific reason for denial in provider responses: Runtime: output check on provider-facing responses. The 42 CFR 431.80 requirement is written for states, so apply it where the contract extends it to the plan.
    • Standard, expedited and drug timeframes: Runtime: apply a different policy set per request type when receipt time, type and limit are supplied as context. Outside the agent: running the clock and classifying requests.
    • Extensions of up to 14 days: Runtime: block extension notices unless the case system records an approved extension. Outside the agent: deciding whether the conditions are met.

    Controls that remain outside AI scope

    Several obligations cannot be satisfied by controlling a tool call. Clock management, extension decisions, clinical review and appeals belong in the utilization management and case-management systems. A runtime layer can refuse to send a notice, but it cannot know whether a case is truly eligible for an extension unless that decision is recorded elsewhere and passed in as context.

    The expertise requirement shows the boundary clearly. An approval gate ensures a human decision happens before an adverse action. Whether the approver holds the right expertise is a credentialing and staffing question. Document who qualifies, how that is verified, and which controls act in the execution path versus which are process controls. Recording that boundary in compliance documentation prevents the runtime layer from being read as covering more than it does.

    Assign ownership for monitoring changes in state and federal rules, including the status of the April 2026 drug proposal. When a limit changes, the update should be a configuration change reviewed by compliance, not an edit to agent behavior.

    Where a runtime control layer fits, and how to evaluate it

    Trussed is the enforcement and evidence layer for this part of the architecture. It enforces policy in the execution path across models, agents, tools and MCP servers: allow, redact, block, rate limit and escalate for human approval, scoped to individual interactions and tool calls. Enforcement is at runtime, not reviewed after the fact, and it generates continuous audit-ready evidence mapped to HIPAA. It can be self-managed in the customer cloud, with configurable retention and data residency. It does not manage clinical, legal or utilization management processes, and it does not replace them.

    Whatever layer you assess, ask five questions. Can it take the state and contract timeframe and request type as inputs and apply a different policy per interaction? Can it hold an agent-drafted denial until a qualified human approves it, and record the approver? Can it block or redact notices missing required elements, with elements configured per state? What evidence does it retain per decision, and how is it exported to auditors? Can it be deployed in your own cloud and sit in the path of your utilization management and Prior Authorization API workflows? Verify any HIPAA mapping directly with the vendor against your own control requirements.

    From requirement to enforced policy

    Requirement

    Federal baseline plus state and contract timeframes and notice content.

    Context

    Receipt time, request type and applicable limit passed to the policy layer.

    Enforcement

    Allow, redact, block, rate limit or escalate for human approval on each tool call.

    Evidence

    Policy applied, action taken, timestamps and approver retained per decision.

    Map your state and contract requirements to enforceable policy

    Walk through how timeframe, escalation and notice controls apply to agent interactions and tool calls, and what evidence each enforcement action produces.

    See How Evidence Is Produced
    Next step

    How does your AI governance program compare?

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