Implementation Guide | Financial Services

    Enforcing financial assistance screening before AI agents send balances to collections

    Make screening a precondition that is evaluated inside the collections-placement tool call, not a line in the agent's prompt. The call executes only if a policy check, run at the moment of the action against current screening status, returns allow. Missing or unverifiable status blocks the call, ambiguous cases escalate to a human approver, and every decision produces a retained evidence record.

    Where enforcement sits in the tool-call path

    1. 1

      Tool registration

      Route every collections-related tool, including placement, debt sale and credit reporting, through the enforcement point. An unrouted tool is an unenforced path.

    2. 2

      Request capture

      Capture the tool name, arguments, account identifier and calling agent before execution.

    3. 3

      Fresh input retrieval

      Read screening status and account facts from the system of record at decision time, not from a cached or agent-reported value.

    4. 4

      Policy evaluation

      Evaluate the versioned policy and return allow, block or escalate before the call leaves the boundary.

    5. 5

      Result recording

      Record what the tool returned, so the record shows what the agent was permitted to do and what it did.

    Why prompts and after-the-fact review do not hold

    An agent that manages patient or consumer account follow-up often has direct access to a tool that places a balance with a collections vendor or agency. If that tool is callable whenever the agent decides to call it, financial assistance screening depends on the agent choosing to check first. That is a design weakness, not a model defect.

    A prompt instruction is a request to the model, not a control on the action. Federal material on AI agents notes that these systems can take autonomous actions that affect real-world systems and can be hijacked or exploited. An instruction that can be ignored, misread or overridden by injected content is not a reliable basis for a screening requirement.

    Post-hoc review has a different limit. It can find that a balance was placed before screening, but it cannot undo the placement or show that the requirement was enforced when the action happened. The CFPB reports consumer complaints about medical debt collection on debts already covered by hospital financial assistance, so the ability to reconstruct a decision matters. Existing legal authorities also apply to automated systems in the same way as to other practices, so agent-driven collections activity is not outside current obligations.

    Define the precondition and the actions it covers

    Start by writing the requirement as a testable condition: a collections-placement action may proceed only when financial assistance screening is complete, the result permits the action, and any timing conditions are met. Legal and compliance should own the wording and sign off on it, because the conditions come from the institution's financial assistance policy and from applicable state rules, which vary.

    Then define the action set. The CFPB describes extraordinary collection actions as including selling an individual's debt to another party, adverse credit reporting, deferring, denying or requiring payment before care because of previous debts, and anything requiring a legal or judicial process. It also says non-profit hospitals may lose tax status if they fail to evaluate patients for financial assistance before certain collection actions. Whether a given placement with a vendor or agency falls within those categories is a question for counsel. The policy should cover the whole defined set, not only the placement tool.

    Timing rules are state-specific. The CFPB's Regulation V preamble describes California conditions on sale of hospital medical debt, including a finding of ineligibility or no response to financial assistance offers for 180 days, and a California prohibition on collection actions before 150 days after initial billing for patients without coverage or with potentially high medical costs. Treat these as examples of conditions to encode where they apply, and have counsel confirm current status as of November 2026.

    Specify the allow, block and escalate outcomes

    Candidate policy inputs are whether screening is complete, the eligibility result, whether the patient responded to financial assistance offers, time since initial billing, the type of action and the applicable jurisdiction. The outcome logic should be derived from the institution's own policy. The following structure is a starting point for that work.

    Fail closed on uncertainty. If screening status is missing, stale or cannot be verified, the answer is block, with a reason code the agent can pass back to its workflow so it routes the account to screening instead of retrying.

    • Allow: Screening is complete, the result permits the action, and all timing conditions for the jurisdiction and action type are met. The call executes and the decision is recorded.
    • Block: Screening is incomplete, the patient is found eligible, a timing condition is unmet, or status cannot be verified. The call does not execute and a reason code is returned.
    • Escalate for human approval: Facts are ambiguous, such as an incomplete application or a dispute. The call is held until a named approver decides and records a rationale.
    • Approval limits: Define which cases an approver may release and which stay blocked regardless, so escalation does not become a bypass.

    Design the evidence record for each decision

    An examiner or model risk management reviewer will want to see what the agent was permitted to do, what it did, and why the control produced that result. A record that shows only the agent's output does not answer that. The evidence should be generated by the enforcement decision itself, at the time of the action, rather than assembled afterward.

    For each decision, retain a timestamp, the agent and tool identity, the account and policy identifiers, and the policy version in force. Retain the input values the policy evaluated, including screening status and the system they were read from. Retain the outcome, and for escalations the approver identity and rationale. Finally, retain the action result, so a block and an executed placement are equally traceable.

    These fields are a design proposal, not a prescribed standard. Reviewers will also look at how the control is validated and kept current, so keep scenario test results as validation evidence. Tests should include the agent attempting placement before screening, with incomplete data, and after a policy change. Define retention periods and retrieval procedures with compliance, and confirm that a single account's decisions can be produced on request.

    Limits of this control and where Trussed fits

    Enforcement at the tool-call level governs individual interactions and tool calls. It does not provide transaction-level governance across tasks and agents, and it does not trace an entire workflow end to end. It also does not control what a vendor does after a permitted placement. Those downstream steps need their own controls, contractual and operational.

    The screening data itself must be accurate. A policy that evaluates a wrong or stale eligibility flag will enforce the wrong result consistently, so ownership of the screening record matters as much as the policy.

    Trussed enforces policy in the execution path, with allow, block and escalate-for-human-approval outcomes at the level of individual tool calls, and produces audit-ready evidence from that enforcement. Insurance is where Trussed is deepest, and it supports customers in four regulated verticals: insurance, pharma, healthcare and education. The approach described here is a pattern any regulated enterprise can apply, with or without Trussed.

    Screening as a tool-call precondition

    Agent requests placement

    The agent attempts the collections-placement tool call for an account.

    Policy check at execution

    Current screening status and account facts are evaluated against the policy.

    Allow, block or escalate

    The call proceeds, is stopped, or waits for a named human approver.

    Evidence record

    Inputs, policy version, outcome and result are retained for each decision.

    Review your collections-placement tool path

    Walk through where a screening precondition would be enforced in your agents' tool calls and what evidence each decision would produce.

    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.