Implementation Guide | Healthcare

    How to Stop Revenue Cycle AI Agents from Writing Off Patient Balances Without Approval

    Treat a write-off as a consequential tool call and enforce policy where the call is made, not in the prompt. Inventory every agent-reachable function that changes what a patient owes, give the agent an identity that excludes those functions by default, and evaluate each call against amount, reason code, account context and agent identity. The policy returns allow, redact, block or hold for human approval, and the system of record does not change until an approver decides. Each decision is retained as evidence.

    1. 1

      Balance adjustments and write-offs

      Any function that reduces a balance or closes it as uncollectible.

    2. 2

      Refunds

      Distinct financial actions that move money and often carry their own timing obligations.

    3. 3

      Account status changes

      Transitions such as moving an account to closed, bad debt or hold.

    4. 4

      Generic data tools

      Query or update tools that could reach the same tables without a named adjustment function.

    5. 5

      Sub-agent and delegated paths

      Downstream agents or services acting on behalf of the requesting agent.

    6. 6

      1. Intercept

      The gate evaluates the call and matches an escalation rule.

    7. 7

      2. Hold

      The call is suspended and no change reaches the billing system.

    8. 8

      3. Present

      A named approver sees amount, reason, account context, rule and agent identity.

    9. 9

      4. Decide

      Approve, deny or request more information, with the approver's identity recorded.

    10. 10

      5. Release or reject

      An approved call executes as presented. A denied call is returned to the agent with the outcome.

    Why an unapproved write-off is an authorization problem

    When a revenue cycle agent reduces a patient balance it should not, the failure is rarely in the wording of the prompt. The agent held a credential that could reach an adjustment function, and nothing between the agent and that function asked whether this specific call was permitted. The cause might be a model error, manipulated input, or a scope that was broader than the task required. The outcome is the same: the balance changed.

    A prompt instruction is an input to a model. It shapes behavior but does not decide whether a call executes. Federal guidance on AI agent security lists prompt engineering, data or tool restrictions, continuous monitoring and human approvals for consequential actions as separate control categories, and notes that agents can be hijacked. Review after the fact belongs to that list too, but a reviewer who finds a bad adjustment is correcting a change that has already happened. Preventive control has to sit at the point of action.

    Medicare contractor manuals offer a useful analogy, not a rule for private providers. They require documented approval before write-offs or changes to a debtor's amounts, and they describe internal controls as approvals, authorizations, verifications, reconciliation and segregation of duties. In the MSP recovery chapter, the receivable is not changed until written approval arrives. The same ordering applies to an agent: decision first, change second.

    Step 2: Define identity and policy conditions for each call

    Give the requesting agent its own identity and scope adjustment functions out of that scope by default. NIST's proposed design tenets for agent identity describe zero standing privilege, least privilege bounded to a task, scopes that narrow when authority is delegated, and separation of duties so an agent cannot both propose and approve an adjustment. They are a proposal, but they translate directly into tool-call policy.

    Then write conditions the gate can evaluate on a single call. Specific thresholds, reason codes and payer rules must come from your organization's financial policy. The structure below is a starting point. Treat any call that changes an amount owed as approval-required unless a documented rule exempts it, which mirrors the default in Medicare debtor-amount guidance, where documented automatic exceptions also exist.

    • Amount: Allow below a documented threshold, escalate above it, and block anything that exceeds the agent's ceiling.
    • Adjustment reason code: Allow only codes the agent's role is permitted to use. Block missing or unrecognized codes.
    • Account and payer context: Evaluate balance age, full versus partial amount, account status and payer class, the kinds of attributes manual write-off rules already use.
    • Requesting agent identity and role: Check which agent is calling, its delegated scope and whether it is acting for a parent agent.
    • Supporting documentation: Require the justification a human approver would expect before the call can be escalated.
    • Payload content: Redact PHI that the tool does not need before the call proceeds.

    Step 4: Retain evidence an auditor can use

    An auditor asking whether unapproved write-offs were prevented needs more than a record of the ones that happened. The useful evidence shows what the agent was allowed to do, what it attempted, and what the control decided. That means blocked and escalated attempts as well as allowed ones, each tied to an agent identity and a policy version.

    For each gated call, retain the agent identity and scope, the tool and parameters (minimized for PHI), the policy version and rule matched, the decision, the approver identity and time for escalated calls, and the final outcome in the system of record. Also retain the history of agent configuration and policy changes, since the question is often what was in force on a given date. NIST's generative AI profile supports documented decision processes and a retention policy for this kind of history. Reconciliation of balances after the fact remains worthwhile, but it complements preventive approval and does not replace it.

    Trussed enforces policy in the execution path for individual interactions and tool calls, with allow, redact, block, rate limit and escalate for human approval. Enforcement itself generates audit-ready evidence, which is mapped to HIPAA, including audit controls at 164.312(b). Deployment can be self-managed in the customer cloud, with configurable retention and data residency. Trussed runs in production with healthcare organizations, and its record includes 18+ months of continuous production in a regulated enterprise and 99.95% uptime on live production traffic.

    Where interaction-level and tool-call-level enforcement fall short

    Each enforcement point sees a different slice of agent behavior, and neither is a complete answer on its own. Interaction-level enforcement evaluates what goes into and comes out of a model, which is where PHI redaction and obviously malicious instructions are best handled. It does not see the function call that eventually reaches the billing system. Tool-call-level enforcement sees the exact action, parameters and identity, which is where an adjustment is allowed, blocked or escalated.

    Both operate on individual events. A policy that evaluates one call at a time can enforce a ceiling per call, but it does not by itself reason about a business transaction that spans several tasks or agents. Splitting a large adjustment into many small ones is the obvious example. Mitigate it with per-account and per-agent rate limits, aggregate rules evaluated over recent calls where your policy engine supports them, and periodic reconciliation. Be explicit in design reviews about what the gate decides and what it leaves to other controls.

    • Interaction level: strengths: Redacts PHI before it reaches a model, and screens inputs and outputs for manipulation.
    • Interaction level: limits: Does not see the tool call that changes the balance.
    • Tool-call level: strengths: Decides on the exact action using amount, reason code, account context and agent identity, and can hold it for approval.
    • Tool-call level: limits: Evaluates calls individually, so patterns across many calls need rate limits, aggregate rules and reconciliation.
    • Both together: Cover PHI minimization and action authorization, but neither replaces financial reconciliation.

    Tool-call path for a balance adjustment

    Agent request

    The revenue cycle agent proposes a write-off or adjustment with an amount and reason code.

    Policy decision

    The call is evaluated in the execution path: allow, redact, block or escalate.

    Human approval

    Out-of-policy adjustments are held until a named approver decides.

    System of record

    The billing or patient accounting system changes only after an allow or an approval.

    Evidence

    Identity, policy version, decision and outcome are retained for audit.

    Put the approval gate in the execution path

    See how policy enforcement at the tool call routes write-offs to a human approver and produces audit-ready evidence for regulated healthcare teams.

    Explore Agent and Tool-Call Enforcement
    Next step

    How does your AI governance program compare?

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