Implementation Guide

    Approve-only prior authorization agents: how to block denial actions at runtime

    An approve-only prior authorization agent is enforced by a policy layer in the execution path, outside the model and prompt. The layer allows approve and recommend-approve actions, blocks any denial tool call, denial status write or denial output, and routes the blocked request to a human reviewer. Each decision produces evidence showing which policy was in effect and what the agent attempted.

    Where denial actions can be intercepted

    1. 1

      Tool calls

      A named tool such as a deny-request function or a set-status function. Policy can block the call by tool name and by arguments, for example a status parameter set to a denial value.

    2. 2

      API and FHIR endpoint calls

      Direct calls to a payer Prior Authorization API or similar endpoint. A denial status write is a distinct action from the agent's text and should be classified as a non-approve outcome.

    3. 3

      MCP server actions

      Tools exposed through an MCP server are callable by the agent like any other. The same allowlist and block rules should apply to them.

    4. 4

      Generated outputs

      Text that states or drafts a denial determination or denial reason. An output condition can block content that would reach a provider or be written into a record.

    Evaluating whether a runtime control fits your architecture

    • Does it decide before a tool, API or MCP call executes, or only review afterward?
    • Can it block a specific denial tool call or status write while permitting approve actions?
    • Is it separate from the model, prompt and agent framework you already use?
    • Does it cover tool calls, API and MCP actions, and generated outputs?
    • Does it route blocked actions to a named human reviewer and help manage escalation volume?
    • Does it record which policy was in effect, and can it run in your cloud with the retention and residency you need?

    Approve-only is an enforcement problem, not a prompt problem

    Many teams want a prior authorization agent limited to approving or recommending approval, with any denial reserved for a human reviewer. The usual first step is a system prompt that says the agent must never deny. That is an instruction, and an instruction describes what you want the model to attempt. It does not control what the model is able to do.

    NIST guidance on agent identity makes the same distinction: the logic that determines what an agent attempts should be architecturally separated from the systems that determine what it is allowed to do. For approve-only, that means the decision to permit or refuse an action cannot live inside the prompt or the model. It has to live in a layer the agent cannot rewrite or talk its way around.

    After-the-fact review has a different weakness. A reviewer who finds a denial status in a queue the next day is looking at an action that has already happened. If the status was transmitted to a downstream system, review documents the problem but does not prevent it. Enforcement at runtime decides before the action executes, which is the only point at which approve-only can actually hold.

    The regulatory context raises the stakes for denial handling. As of November 2026, 42 CFR 422.122 for Medicare Advantage organizations and 42 CFR 431.80 for State Medicaid agencies require a specific reason in the response to the provider whenever a prior authorization request is denied, beginning January 1, 2026. Drug coverage requests are excluded from that paragraph. Both rules also require a Prior Authorization API beginning January 1, 2027, and for States the API communicates whether a request is approved or denied, unless an extension or exemption is granted. A denial is therefore a distinct, consequential output, and an agent that can produce one outside its intended role creates exposure.

    Defining approve-only as enforceable policy

    Approve-only becomes enforceable when it is written as explicit rules that a runtime layer can evaluate on every interaction and tool call. A workable policy has five elements. The elements below are a design approach rather than a standard, so adapt the vocabulary to your own systems.

    Start by classifying outcomes. Approve, deny and request-more-information outcomes appear in prior authorization interface descriptions, and your policy should treat everything other than approve as a non-approve outcome by default. A default-deny posture for unlisted actions is safer than trying to enumerate every denial path.

    • Action allowlist: Name the actions the agent may take, such as approve or recommend approval, and permit nothing else by default.
    • Denial-tool blocking: Block any denial tool, denial status write or endpoint call that sets a non-approve outcome, regardless of how the agent reached it.
    • Output conditions: Block generated content that states a denial determination or a denial reason, since a denial must carry a specific reason and that content should come only from a human reviewer.
    • Escalation rule: When an action is blocked, escalate for human approval to a named reviewer group so the request still reaches a decision-maker.
    • Scoped, short-lived credentials: Issue the agent only the access needed for the task, for a short period, so authority can be revoked quickly if behavior drifts.

    Escalation design and operational tradeoffs

    Blocking a denial without routing the case anywhere creates a stalled request. The blocked outcome should become an escalation to a human reviewer with the context needed to decide, and the agent's own record should show that it did not complete the denial. Keep any human-made denial separate from agent actions so the record shows who decided.

    Escalation volume needs deliberate management. NIST notes the risk of consent fatigue: if reviewers receive so many escalations that they approve habitually, human review loses its value. Tune the policy so that escalations represent genuine non-approve outcomes rather than noise, and document who sets the risk thresholds and who receives escalations.

    There are tradeoffs to accept. Enforcement in the path adds a decision step to each governed interaction, so latency and availability of the enforcement layer matter. A strict allowlist will block legitimate new tools until they are reviewed, which is usually the intended behavior but needs an owner. And per-interaction enforcement governs individual interactions and tool calls. It does not by itself give you a view of an entire multi-step workflow, so do not rely on it for that purpose.

    This guide does not assume a legal rule that a clinician must review every denial. Approve-only is an organizational control you choose to implement, and your compliance team should determine which review obligations apply to you.

    Evidence to retain for each decision

    A regulator or auditor will want to see two things: what the agent was permitted to do, and what it actually did. Enforcement that produces evidence at the moment of decision answers both. Reconstructing the answer later from application logs is slower and less reliable.

    Trussed enforces policy in the execution path for models, agents, tools and MCP servers, with allow, redact, block, rate limit and escalate for human approval decisions at the level of individual interactions and tool calls. Evidence is generated by that enforcement rather than assembled afterward, and it is mapped to frameworks including HIPAA, NAIC, NYDFS, the EU AI Act and NIST.

    Whatever tooling you use, the fields below are a sensible minimum for each blocked or escalated action. They are a recommendation to validate with your audit team.

    • Policy identity and version: Which approve-only policy was in effect at the time of the decision.
    • Action attempted: The tool, endpoint or output, with the relevant arguments such as the requested status.
    • Decision: Allow, block or escalate, with the rule that produced it.
    • Timestamp and agent identity: When the attempt occurred and which agent and credential made it.
    • Escalation target and reviewer outcome: Who received the case and what the human reviewer decided.
    • Retention and residency: Records kept for the period and in the location your obligations require.

    Approve-only enforcement in the execution path

    Agent attempts an action

    A tool call, API or MCP server action, or generated output.

    Policy decides before it runs

    Allow approve actions. Block denial actions. Escalate for human approval.

    Human reviewer receives the case

    Blocked non-approve outcomes are routed, not dropped.

    Evidence is retained

    Policy in effect, action attempted, decision and outcome.

    Test approve-only against your own agent architecture

    Walk through your agent's tools, API calls and outputs with us and see how allow, block and escalate decisions are enforced and how audit-ready evidence is produced.

    Request a Demo