Gold carding and AI prior authorization agents: controls and evidence for exemption programs
An AI prior authorization agent that touches a gold carding program needs controls at four points: the exemption status lookup, tool calls to eligibility and claims systems, the approval or routing recommendation, and handoff to clinical review. Each control should be enforced before the action completes, and each should produce a decision-level record showing which policy applied, what the outcome was, and who or what acted. Runtime enforcement can supply that record for individual interactions and tool calls. It cannot prove that the exemption program itself is well designed.
- Agent identity and the timestamp of the interaction or tool call
- The tool or system called and the data scope requested
- The policy name and version applied, including exemption criteria version and assessment period
- The enforcement outcome: allow, redact, block, or escalate for human approval
- For escalations, the reviewer's identity, decision, and time of response
- For any denial communicated downstream, the specific reason, since 42 CFR 431.80 requires one for State Medicaid prior authorization APIs
- Does it enforce allow, redact, block, or escalate in the execution path, before a tool call completes?
- Does each record carry the policy version, outcome, agent identity, and timestamp?
- Can escalation hold an action until a human reviewer approves it?
- Can retention be configured to the payer's legal requirements, and can evidence be exported for regulators and auditors?
- Can it run self-managed in your cloud, and what data residency controls apply to evidence containing PHI?
- Does the vendor state what the evidence does not cover, such as transaction-level governance, end-to-end workflow tracing, and model provenance?
Gold carding is a policy and evidence problem
CMS describes gold carding as programs that relax or reduce prior authorization requirements for providers with a consistent pattern of compliance. Providers are relieved of submitting requests based on adherence to submission requirements, appropriate utilization, or other evidence-driven criteria. The examples CMS has cited are quantitative and periodic. A Medicare home health demonstration allows selective post-payment review or spot check review for agencies with a review affirmation or claim approval rate of 90 percent or greater over 6 months. In July 2025, CMS said it was exploring gold carding under the WISeR model, with exemption possible through a 90 percent provisional affirmation threshold during a periodic assessment. That language was conditional, and a payer's own program may use different thresholds.
For Medicare Advantage, CMS stated in December 2022 proposed rule text that organizations have discretion to implement gold carding programs within each contracted plan. Prior authorization policies for coordinated care plans remain limited by 42 CFR 422.138(b) to confirming diagnoses or other medical criteria, or ensuring medical necessity under specified standards. The consequence is practical: the payer designs the program and must be able to defend it. Once an AI agent performs exemption lookups or routing, the defense depends on showing that the agent applied the criteria in force at the time, stayed within its permitted scope, and escalated what fell outside the rules.
Reviewing outputs afterward does not establish those facts. An output shows what the agent concluded. It does not show what the agent was permitted to call, which exemption record it read, or which policy version applied at that moment. The mapping below is design guidance for closing that gap, not a statement of a legal requirement.
Four control points in the agent's execution path
The first control point is the exemption status lookup. Treat it as a policy-controlled read. The exemption depends on the criteria version and assessment period in force when the decision is made, so the lookup should record the exemption record version, the assessment period, and the criteria applied. Criteria such as threshold, assessment period, and service scope should be expressed in machine-readable policy that mirrors the written program.
The second is tool calls to eligibility and claims systems. The agent's permitted tools should be limited to those systems and to the data fields the task needs. Calls outside that scope are blocked. Responses containing protected health information that the task does not require can be redacted before they reach the agent. Each allow, redact, or block outcome is recorded against the call.
The third is the approval or routing recommendation. Escalation triggers should be defined in advance: an exemption that has lapsed or is under review, a requested service outside the exempted scope, an ambiguous status, or a potential adverse determination. When a trigger fires, the agent should route the case rather than resolve it.
The fourth is handoff to clinical review. CMS cites 42 CFR 422.566(d) as requiring that an adverse medical necessity decision on a prior authorization request be reviewed by a physician or other appropriate health care professional before the MA plan issues it. An agent's exemption logic should therefore never finalize an adverse outcome. A runtime policy that escalates for human approval, and holds the action until a reviewer responds, turns that rule into an enforced step rather than a procedure document.
What interaction-level enforcement covers, and what it does not
CMS did not directly address AI in its 2024 interoperability and prior authorization final rule. It encouraged innovation that is secure, includes medical professional judgment for coverage decisions, and operates under a governance structure that provides transparency, evaluation, and ongoing monitoring. Enforcement at the level of individual interactions and tool calls supports that expectation in one specific way: it produces durable evidence of what the agent was allowed to do and what happened at each step.
Trussed approaches this by enforcing policy in the execution path across models, agents, tools, and MCP servers, with allow, redact, block, rate limit, and escalate for human approval as the available actions. The evidence is generated by enforcement at runtime, not reviewed after the fact, and is mapped to frameworks including HIPAA, NAIC, and NIST. It can be deployed self-managed in the customer's cloud, with configurable retention and data residency. That is one way to meet the requirements above, and the control points apply regardless of the tooling used.
The limits matter for how a compliance leader describes the program. Interaction-level enforcement does not show that the exemption criteria themselves are valid or well calibrated. It does not provide transaction-level governance across tasks and agents, end-to-end tracing of an agentic workflow, or model provenance. State gold carding statutes and record-retention rules need separate legal analysis, and retention periods for interaction-level evidence should be set from that analysis. As of November 2026, the Medicaid prior authorization API requirement in 42 CFR 431.80 applies beginning January 1, 2027, unless an extension or exemption is granted, so teams in scope should test denial-reason handling ahead of that date.
Practical next steps
Start by writing the exemption criteria as policy a system can evaluate: threshold, assessment period, service scope, and the conditions that suspend or revoke an exemption. Then list every tool the agent can call and remove any that the exemption workflow does not need. Define escalation triggers and the reviewer roles that respond to them, and confirm that no path lets the agent finalize an adverse determination.
Next, decide what evidence each control point must retain and for how long, based on your own legal analysis of state and federal requirements. Finally, plan for periodic reassessment. Evidence should show which assessment-cycle data the agent used, so a decision made last quarter can be explained against the criteria that applied last quarter.
Control points in an exemption-aware agent
Exemption lookup
Policy-controlled read that records the criteria version and assessment period applied.
Eligibility and claims tool calls
Permitted tools are scoped; calls outside scope are blocked.
Approval or routing recommendation
Cases outside exemption rules are escalated, not resolved by the agent.
Clinical review handoff
Potential adverse outcomes go to a human reviewer for approval.
Review your exemption controls against the execution path
Walk through how policy enforcement and audit-ready evidence apply to each exemption lookup and tool call in your prior authorization workflow.
Explore Agent and Tool-Call EnforcementHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.