Technical Guide | Financial Services

    Evidence a KYC Review Agent Should Produce for a CIP Rule Examination

    For every interaction and tool call, a KYC review agent should produce a record of the data it accessed, the tool or verification method it invoked, the policy version in force, the outcome (allowed, redacted, blocked or escalated) and the person who approved any exception. These records support examiner requests such as the list of verification methods and the list of accounts with incomplete verification or CIP exceptions. They supplement the institution's own CIP records and do not replace them.

    Criteria for assessing your current environment

    • Can you retrieve, per customer interaction, the data sources accessed and the tools called?
    • Does each action record the policy version in force and its outcome: allowed, redacted, blocked or escalated?
    • Can you list accounts with incomplete verification or CIP exceptions, including the approver's identity?
    • Can you show which identifying-information fields the agent collected or used, against the fields your CIP specifies?
    • Are records protected against alteration and retrievable on examiner request?
    • Are retention settings aligned to the period in the current CIP rule, and is it clear which evidence lives in the agent environment and which in your KYC systems?

    Start from what an examiner asks for

    CIP readiness is an evidence-production problem. The CIP rule for banks, 31 CFR 1020.220, requires a program with risk-based verification procedures that let the bank form a reasonable belief that it knows the true identity of each customer. Those procedures must specify the identifying information the bank obtains, and account-opening procedures must set out what is collected from each customer.

    Examination procedures translate that into concrete requests. Examiners look for a list of the documentary and nondocumentary methods the institution uses to verify identity. They may also ask for a list of accounts opened where verification was not completed, or that were opened with exceptions to the CIP. The same procedures state that making or approving exceptions cannot be allowed by policy, and that isolated, nonsystemic errors may be deemed not to compromise the program's effectiveness.

    If an AI agent assists the review, each of those requests now has an agent-side component. The practical test is whether the agent's records can produce each list on demand, with detail at the level of the individual customer and action. Parallel rules apply to other covered institutions, so confirm the rule that applies to your charter.

    Mapping CIP program elements to evidence artifacts

    The mapping below is a proposed specification, not a regulatory requirement. It pairs each CIP element an AI-assisted review could touch with the artifact that would let you answer an examiner's question about it.

    • Identifying information collected: Per-interaction record and data-access record showing which fields were read or written, compared against the fields your CIP specifies.
    • Verification procedures: Tool-call record naming the documentary or nondocumentary method or source invoked for each customer, including the type of identification reviewed for individuals.
    • Discrepancy handling: Policy-decision record showing the outcome (allowed, redacted, blocked, escalated) and the escalation path taken when information did not match.
    • Exceptions and incomplete verification: Human approval record with approver identity and time, so the list of exception accounts can be produced with the person responsible.
    • Agent acting for another person: Record showing which person was treated as the customer, since the person on whose behalf an account is opened is the customer.
    • Recordkeeping: Retention settings, alteration protection and retrieval path for the agent's records, set to the period in the current CIP rule text.

    Why runtime evidence is more defensible than reconstruction

    Evidence can be produced in two ways. It can be generated at the moment an action is evaluated, or it can be reconstructed afterward from application logs and agent outputs. Reconstruction can show that a tool was called. It generally cannot show which policy was in force at that moment, or whether the action was allowed, redacted, blocked or escalated for approval, unless the system recorded that decision as it happened.

    This is a design argument rather than a regulatory finding, but it matches how examiners probe. A question such as "show me how this customer's discrepancy was resolved" requires the data sources accessed, the tools called, the rule applied and the human who approved the outcome, all tied to one interaction. Records assembled months later from separate systems depend on timestamps lining up and on someone interpreting intent.

    Each action record should therefore carry the policy version applied and the outcome. When policies change, which they will, the record still shows the rule that governed that action rather than the rule in force on the day of the examination. Evidence should also let errors be identified and counted. Because isolated, nonsystemic errors may not compromise a program, an institution needs to show that errors are isolated, and that requires a countable record.

    What stays with the institution

    Agent evidence supplements the CIP. It does not satisfy it. The written program, the verification procedures, the identifying information, the verification results and the retention of CIP records remain the institution's responsibility and stay in its own KYC systems. The agent's records show how the agent behaved inside that program: what it touched, what it was permitted to do, and when a person took over.

    Customer notice follows the same logic. It is delivered through the institution's own channels and procedures, and agent evidence is relevant only if the agent participates in that step.

    Governance questions sit alongside the evidence. Because exceptions cannot be approved by policy, escalation and approval records should show that exceptions were handled outside routine agent authority, and by whom. The model risk function should also decide whether the KYC agent falls within model risk management scope and what validation evidence it expects. The Federal Reserve's SR 11-7 guidance describes effective model risk management and a validation framework, and a later letter, SR 26-2, is titled Revised Guidance on Model Risk Management. As of November 2026, confirm with your model risk team which guidance governs your institution.

    Where policy enforcement in the execution path fits

    Trussed enforces policy in the execution path across models, agents, tools and MCP servers. Policy is applied to individual interactions and tool calls, with outcomes of allow, redact, block, rate limit or escalate for human approval. Because the decision is enforced at runtime, not reviewed after the fact, the audit-ready evidence is generated by the enforcement itself. That is the property the evidence specification above depends on: a policy decision and its outcome recorded for each interaction and tool call.

    Deployment is self-managed in the customer's cloud, with configurable retention and data residency, so an institution can set how long evidence is kept and where it is stored. Trussed is deepest in insurance, another regulated financial sector, and has run for 18+ months in continuous production in a regulated enterprise.

    The limits matter. Trussed's evidence covers individual interactions and tool calls. It does not provide transaction-level governance, end-to-end tracing of agentic workflows, or model and dataset lineage. It also does not establish that CIP obligations are met. The institution's KYC systems and procedures remain the record for the program itself.

    Evidence chain for one KYC agent action

    Interaction

    Per-interaction record tied to the customer review

    Data and tools

    Fields accessed and verification tools called

    Policy decision

    Policy version applied and the allow, redact, block or escalate outcome

    Human approval

    Approver identity for any exception

    Retention

    Retention and residency settings applied to the record

    Test your KYC agent against an examiner's request list

    Walk through how policy enforcement in the execution path produces per-interaction and tool-call evidence, and where your own KYC systems remain the record.

    Assess Your AI Readiness
    Next step

    How does your AI governance program compare?

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