How to Stop a KYC Agent from Accepting Expired or Unverified Identity Documents
To prevent a KYC agent from accepting expired identity documents, remove document acceptance from the model's judgment. Compute expiry deterministically against a recorded evaluation date, require a completed authenticity status from the verification provider, and enforce both as preconditions on the tool call that advances onboarding. Map every edge case to allow, block or escalate, and record the evidence each decision produces.
Why a KYC agent accepts documents it should not
The failure is usually structural rather than a model defect. An agent in an onboarding workflow reads a document, calls tools, and decides whether to proceed. If the model's reading of the document, or a tool result it receives, is treated as sufficient, then acceptance depends on judgment that can be wrong, incomplete or manipulated.
Several patterns follow from this design. The model may extract an expiry date correctly but fail to compare it to the current date. It may treat a vendor response as a pass when the authenticity check is pending, partial or absent. It may resolve conflicting data between the document and other inputs by choosing the more convenient value. It may proceed when a field such as the expiry date is missing, because nothing told it that absence is a failure. These are failure modes inferred from how such workflows are built, and each one is closed by a control outside the model.
A prompt instruction such as "never accept expired documents" is a request, not a control. It cannot guarantee behavior, and it leaves no proof of what the agent was permitted to do. Examiners and model risk reviewers ask exactly that question for each document, so the control has to sit where the agent acts.
What the rules imply for document validity
As of November 2026, U.S. customer identification program rules require risk-based procedures that enable a reasonable belief about each customer's true identity. For firms that rely on documents, the program must set out the documents it will use, and the example given for an individual is an unexpired government-issued identification that evidences nationality or residence and bears a photograph or similar safeguard. Expiry is therefore a condition of acceptability for those documents, not an optional signal.
Agency guidance on verifiable digital credentials takes the same line. An unexpired government-issued credential such as a mobile driver's license can qualify, and indications of fraud must be weighed in deciding whether a reasonable belief can be formed. Programs must also have procedures for when that belief cannot be formed. Under the EU AML Regulation, customer due diligence includes identifying and verifying the customer's identity.
None of these texts prescribes how an automated agent should be controlled. They set the conditions the control must enforce: the permitted document list, currency, verification, and a defined path when the result is uncertain.
Separate genuineness from currency
NIST identity-proofing guidance treats validation as two distinct questions. Is the evidence genuine, and is the data on it valid and current? An expiry check answers only the second. A verification provider's authenticity result answers the first. Neither substitutes for the other, so the design should require both as independent conditions.
The deterministic layer is ordinary code, not a model call. It takes the extracted expiry date and compares it to a defined evaluation date, then records both values and the result. A missing or unparseable date returns a distinct state rather than a pass. The vendor layer supplies the authenticity status. Define in advance which statuses count as complete and passing, and treat every other state, including pending, partial and error responses, as not verified.
The model can still read the document and extract fields. It should not decide whether the extracted values are acceptable. Its output becomes an input to the deterministic checks, and the check results become inputs to the policy condition.
Enforce acceptance conditions on the tool call
The decisive control is a precondition on the tool call that advances onboarding, such as the call that marks identity as verified or submits the application for approval. The condition is evaluated in the execution path, so neither the model nor a tool result can satisfy it by assertion. NIST guidance on agent identity makes a related point: without runtime verification, an agent operating under a valid token cannot be interrupted mid-task.
Specify the policy against the document types listed in your customer identification program, and evaluate it on structured fields rather than free text.
- Permitted document type: The document type is on the program's approved list for this customer segment.
- Currency result: The deterministic expiry check ran, returned unexpired, and recorded its evaluation date.
- Authenticity status: The provider returned a completed, passing status for this specific document.
- Consistency: Document data does not conflict with other captured inputs, and no fraud indicator is set.
- Agent identity: The call carries a distinct agent identity and the initiating entity for the task.
- Outcome: If any condition fails or is missing, the call is blocked or escalated, never silently passed.
Mapping edge cases to allow, block or escalate
The guidance requires defined procedures when a reasonable belief cannot be formed and supports human review at organization-defined risk thresholds. It does not assign outcomes to specific edge cases, so the mapping below is an internal policy decision. These are reasonable starting points to adapt to your risk assessment.
- Unexpired, authenticity complete and passing, no conflicts: Allow. Record the inputs and the policy version applied.
- Expired on the evaluation date: Block. Return a reason code and request a current document.
- Expiry date missing or unreadable: Block by default. Escalate if your program allows a defined alternative document or method.
- Verification pending, incomplete or errored: Block the advancing call until a completed status exists.
- Conflicting data between document and other inputs: Escalate for human approval. The agent does not choose between values.
- Fraud indicator on the document or credential: Escalate or block per risk threshold, with the indicator recorded.
Evidence each decision must produce, and how to keep controls honest
Customer identification guidance expects records of the document relied on: its type, any identification number, place of issuance, and the issuance and expiration dates where they exist, along with the methods and results of the verification measures. Retain these for each document decision. Beyond those required fields, add the fields a model risk reviewer will ask for: the evaluation date used, the provider's status value, the policy conditions that applied, the outcome, the agent identity, the initiating entity, and any reviewer decision and rationale on escalations. Retention periods depend on the rules that apply to you. An EU AML Regulation recital refers to keeping due diligence information for at least 5 years.
This is where runtime enforcement earns its place. Trussed enforces policy in the execution path on individual interactions and tool calls, with outcomes of allow, block or escalate for human approval, and generates audit-ready evidence from that enforcement rather than from after-the-fact review. Trussed does not authenticate documents or perform KYC. The verification provider and your deterministic checks do that, and the policy confirms their results are present and passing before the agent can act. Deployment is self-managed in your cloud, with configurable retention and data residency.
Finally, test the controls continuously. NIST notes that validation checks can fail with no outward sign until an attack occurs, and that they must reject invalid input as reliably as they accept valid input. Run known-invalid cases, such as expired, unverified, missing-date and conflicting-data documents, as regression tests and confirm each is blocked or escalated.
Layered controls on document acceptance
Currency check
Deterministic comparison of the extracted expiry date to a recorded evaluation date.
Authenticity status
Completed verification result from the document verification provider.
Tool-call condition
Policy evaluated in the execution path before the agent can advance onboarding.
Decision evidence
Allow, block or escalate, recorded with the inputs that produced it.
Enforce document acceptance where the agent acts
See how policy enforcement on agent and tool-call interactions produces audit-ready evidence for each decision.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.