Implementation Guide | Financial Services

    Logging every document a KYC review agent read before recommending approval

    A defensible KYC agent document audit trail records each document read and tool call as evidence produced at the point of enforcement. Each entry identifies the document and version, the requesting agent, the policy applied, the decision (allow, redact, block or escalate), a timestamp, and the recommendation the read informed. Capture inside the execution path is more complete than reconstructing activity from scattered application logs. The evidence store also needs its own retention, redaction and residency controls.

    Testing whether the evidence would hold up to scrutiny

    • Can you list every document the agent read for this recommendation, in order, with the policy applied and the decision?
    • Does the record show blocked and escalated requests, not only successful reads?
    • Is capture in the execution path, and which reads could it miss?
    • Does the record hold descriptors or content, who can access it, and how is it encrypted?
    • Can retention and residency be set separately for evidence containing customer identifiers?
    • Can the link from each read to the recommendation be reproduced for model risk review?

    What a defensible record of document access contains

    When an agent recommends approval in a KYC review, an examiner or an internal model risk team can ask three questions. Which documents did the agent read? In what order? Under what policy? A record that answers all three for a specific recommendation is the target.

    Existing customer identification program (CIP) record-keeping rules give a useful baseline for the fields. As of November 2026, the final CIP rule for banks requires a description of any document relied on to verify identity, rather than a copy. The description covers the document type, any identification number, the place of issuance, and issue and expiration dates where they exist. FINRA's notice for broker-dealers takes the same approach and adds that, for non-documentary verification, records must describe the methods used and their results.

    These rules concern identity verification records, not AI agents, and they do not say how an agent should log a read. They do show what regulators already expect a firm to retain about the documents behind a decision. An agent record that carries at least those descriptors, plus the agent-specific fields below, gives reviewers familiar reference points. FDIC guidance also encourages reviewing more than one document, so the record should show multi-document review explicitly, not only the final document relied on.

    Evidence fields to capture at each document read or tool call

    The fields below are a design recommendation. The first group follows the CIP record elements. The second group is what an agent adds, because a reviewer must connect a read to a decision and a policy.

    Capture the same structure for non-document tool calls, such as a sanctions or database lookup, recording the method invoked and the result returned. This mirrors the non-documentary verification record and prevents the trail from covering only the documents.

    • Document descriptor: Document identifier and version, type, identification number where present, place of issuance, and issue and expiry dates. Store the descriptor even if content is not retained.
    • Requesting agent and session: Which agent, under which identity and review case, made the request.
    • Policy applied: The policy and version evaluated, including whether the document type is on the list of documents the firm's procedures permit.
    • Enforcement decision: Allow, redact, block or escalate for human approval, with the reviewer outcome if escalated.
    • Sequence and timestamp: A timestamp and ordinal position so the read order can be replayed.
    • Recommendation reference: A key linking the read to the specific recommendation it informed, including reads that were blocked.

    Where capture has to happen in the execution path

    A KYC agent touches several systems on the way to a recommendation. It retrieves from a document store, calls tools or MCP servers, and invokes a model with the content it received. Each system keeps its own logs, with its own identifiers and clocks. Reconstructing a review means joining those logs after the fact, and the join usually fails at the point that matters most: tying a specific read to the specific recommendation it informed.

    Capture is more complete when it happens where the request is enforced. If policy is evaluated in the execution path, every tool call and retrieval passes through one point that sees the request, applies the policy, and records the outcome before the content reaches the model. The evidence is a by-product of enforcement, not a reconstruction. Reads that were blocked or redacted appear in the same record as reads that were allowed, which after-the-fact assembly often loses.

    This has limits. Capture at the level of individual interactions and tool calls shows what was requested, what the policy decided, and what was released. It does not show whether the underlying CIP or CDD program is adequate. It also does not describe wider workflow context beyond those interactions. Reads that bypass the enforcement point, for example a tool the agent can reach by another route, will not appear, so the agent's reachable paths need to be constrained as well as observed.

    Recording policy decisions alongside each access

    A list of documents read tells a reviewer what happened. It does not tell them whether it was allowed. The policy decision should sit in the same entry as the read, so the record answers the question of whether the agent was limited to permitted sources.

    Four outcomes are worth distinguishing. Allow means the document was released as requested. Redact means it was released with specified fields removed, and the entry should note which classes of field were removed. Block means the request was refused, which is itself evidence that a boundary held. Escalate means a human approver was required before release, and the entry should capture the approver's decision.

    Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with allow, redact, block, rate limit and escalate for human approval. The audit-ready evidence is generated by that enforcement rather than assembled later. Whichever tooling a firm uses, the test is the same: can the decision be shown next to the read, with the policy version that produced it?

    Retention, redaction and data residency for the evidence itself

    The record of document access can be as sensitive as the documents. An entry that stores an identification number or a document image is customer information, and it is easy to build an evidence store that is less controlled than the source system.

    Start with scope. The final CIP rule asks for descriptions rather than copies, while FDIC guidance notes that a bank may keep copies and that risk-based procedures may warrant it in some situations. A reasonable design stores descriptors and identifiers by default, and retains content only for risk tiers where the firm has documented why. That is a design choice, not a regulatory requirement, and the rationale should be written down.

    Then apply controls to the store. The FTC Safeguards Rule treats an agent as an authorized user, calls for encryption of customer information in transit and at rest or approved compensating controls, and expects monitoring and documentation of security events. Whether that rule applies to a given institution is a separate question, but it is a sensible baseline for an evidence store holding identification numbers. Set retention periods for evidence containing customer identifiers separately from other evidence, and keep data residency consistent with where the source documents may lawfully sit.

    Trussed supports self-managed deployment in the customer's cloud, with configurable retention and data residency. This keeps the evidence within the firm's own boundary and lets the firm set those controls itself.

    One approval recommendation, reconstructed from its evidence

    Document read

    Identifier, version and type of each document the agent accessed, in order.

    Policy applied

    The rule that governed the read and the decision it produced.

    Decision

    Allow, redact, block or escalate for human approval, recorded with the read.

    Recommendation link

    A reference tying every read to the approval it informed.

    Check your KYC agent evidence against a real approval

    Walk through how enforcement in the execution path produces a per-read record, and test it against the questions above.

    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.