Implementation Guide | Financial Services

    How to Prove to Examiners Which Transactions an AML Investigation Agent Reviewed

    Capture evidence at the point of execution, not from the agent's own summary. For each alert or case, record every tool call, every data retrieval with the identifiers of the records returned, every policy decision (allow, redact, block, escalate) and every human approval, and keep those records in durable, tamper-evident storage. Then reconcile them to case-management and transaction systems. Tool-call evidence covers individual interactions. It does not by itself give transaction-level governance across tasks and agents.

    Where evidence can be captured in the execution path

    1. 1

      Prompts and instructions

      The task given to the agent, tied to the alert or case ID, so scope can be compared with what followed.

    2. 2

      Tool calls

      Each call the agent makes to a query tool, search function or case system, with parameters and timestamp.

    3. 3

      Data retrievals

      The identifiers of the records returned, so a claim about a specific transaction can be checked against a specific retrieval.

    4. 4

      Policy decisions

      The outcome of each enforcement decision (allow, redact, block, escalate) and the policy that produced it.

    5. 5

      Human escalations and approvals

      Who was asked, who approved or rejected, and when, linked to the action that was held.

    6. 6

      Outputs

      The narrative, recommendation or case note the agent produced, linked to the retrievals that preceded it.

    Requirements for durable, audit-ready evidence

    • Each record carries the alert or case ID, agent identity, tool, timestamp and returned record identifiers.
    • Policy decisions (allow, redact, block, escalate) are stored with the interaction they governed.
    • Human approvals identify the approver, the decision and the action held.
    • Storage is tamper-evident, with access controls and independent review of integrity.
    • Retention is set by the institution, at least as long as the SAR supporting documentation period, and retrieval is practical during an examination.
    • Storage location meets the institution's data residency requirements.

    What examiners are likely to expect

    Examination procedures describe an examiner who tests rather than listens. OCC community bank procedures include sampling suspicious activity alerts, discussing the investigation process with staff, and reviewing the decision-making process for SAR filings. Examiners may also review auditor reports, scope and supporting workpapers. FFIEC procedures describe transaction testing of suspicious activity monitoring systems as intended to determine whether policies, procedures and processes are adequate and effectively implemented.

    When an AI agent supports alert review, the same testing applies. An examiner who samples an alert will want to know what the agent was permitted to do, what it actually accessed, what it produced, and who approved any consequential action. Model risk management guidance such as SR 11-7 frames this in terms of sound implementation and use, effective validation, and governance, policies and controls. For an investigation agent, those categories translate into records that can be inspected.

    SAR supporting documentation adds a retention dimension. OCC guidance on SAR supporting documentation states that a filer must keep a copy of the SAR and the original or business record equivalent of supporting documentation for five years from the filing date, as of November 2026. Whether agent records fall within that definition is a question to settle with counsel. Planning for at least that retention period is a reasonable default.

    Before designing anything, define what "reviewed" means. Records the agent retrieved and records the analyst relied on are different sets. Your definition should be explicit, and the evidence should be able to test it.

    Why summaries and application logs fall short

    An agent's narrative states what it believes it reviewed. It is generated text, and it can omit retrievals, overstate coverage or describe records it never accessed. It is useful as a work product but weak as proof of scope.

    Application logs written after the fact have a different weakness. They are usually designed for troubleshooting, so they may record that a query ran without recording which records came back, which policy applied, or whether a person approved the step. They may also sit in several systems with different clocks and retention settings. The result is a gap between what the agent claims and what it retrieved, which is the gap an examiner will probe.

    Evidence generated by enforcement at the tool-call boundary closes much of that gap because the record is created as the action is decided and executed. The agent's summary can then be tested against it. If the narrative cites a transaction that no retrieval returned, or a retrieval returned records the narrative never mentions, the discrepancy is visible rather than hidden.

    Enforcing policy in the execution path

    Trussed is the runtime control and evidence layer for this part of the design. It enforces policy in the execution path across models, agents, tools and MCP servers, with outcomes of allow, redact, block, rate limit or escalate for human approval. Enforcement happens at runtime, not after the fact, and the enforcement itself generates audit-ready evidence of individual interactions and tool calls. Trussed can be deployed self-managed in the customer's cloud, with configurable retention and data residency.

    For an AML use case, this means a query tool can be limited to the data an investigation task requires, sensitive fields can be redacted, out-of-scope requests can be blocked, and consequential actions can be held for analyst approval. Each outcome becomes part of the record. Trussed has run continuously in production in a regulated enterprise for more than 18 months and generates over 30 GB of audit evidence per day. Its customers are in insurance, pharma, healthcare and education, and it does not claim financial-services deployments. The evidence-design principles in this guide apply regardless of the tooling you choose.

    Limits of tool-call evidence and how to reconcile

    Tool-call evidence is scoped to individual interactions and calls. It shows what an agent asked for and what was returned. It does not establish business transaction identity, and it does not provide transaction-level governance across tasks and agents. A single investigation may involve several tasks or several agents, and a transaction may appear in more than one retrieval. Linking those into one complete account of how a transaction was handled is work for your case-management and core transaction systems, which remain the systems of record.

    Treat reconciliation as a standing control, not a one-time exercise. It supports the independent testing and transaction testing examiners expect, and it gives you an early view of gaps before an examiner finds them.

    • Compare agent-recorded retrievals with case-management notes for a sample of closed alerts.
    • Compare retrieval records with query history in the transaction systems to confirm the same records were returned.
    • Test the agent's stated scope of review against enforcement records, and log every discrepancy.
    • Document scope, procedures, testing and findings so workpapers can be made available to examiners.
    • Assign an owner outside the agent's operating team to review evidence integrity and completeness.
    • Record known gaps openly, such as cross-task linkage, along with the manual or system step that covers each.

    From examiner question to evidence

    Examiner asks

    Which transactions did the agent review for this sampled alert?

    Enforcement record

    Tool calls, retrievals and returned record identifiers captured in the execution path.

    Policy decision

    Allow, redact, block or escalate, stored with the interaction.

    Reconciliation

    Agent evidence compared with case-management notes and transaction system queries.

    Design the evidence before the examiner asks for it

    See how enforcement in the execution path produces audit-ready evidence of agent interactions and tool calls, and where your own reconciliation still needs to start.

    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.