Implementation Guide | Financial Services

    How to Prevent Tipping Off in AML Investigations in Customer Service, Including AI-Assisted Tools

    Preventing tipping off is a control-design problem, not a training problem. Use four layers: minimize the investigation-linked data that agents and AI tools can see, give agents approved neutral responses, escalate defined triggers to a designated AML or compliance team, and enforce disclosure policy at runtime on AI-assisted interactions. The test is whether a reply would reveal the existence of a SAR, not whether it mentions a restriction. Retain evidence that each layer operated.

    Evidence to retain for examiners

    • Policy version applied to each AI-assisted interaction, with change history
    • Blocked, redacted and rewritten outputs, with the rule that triggered the action
    • Escalation records, including the designated team that received the case and the outcome
    • Human approval records for drafts routed for review
    • Test results measuring leakage and false positives before and after enforcement
    • Documented link between the policy and the current joint statement, with review dates

    What the control must actually prevent

    The Bank Secrecy Act prohibits disclosing a SAR, or any information that would reveal its existence, including to the customer who is the subject of the report. That prohibition reaches a bank's officers, directors, employees and agents, which covers customer-facing staff.

    The joint statement issued on 2 September 2026 by the OCC, Federal Reserve, FDIC, FinCEN and NCUA narrows the practical target. As of November 2026, it says the BSA and its implementing regulations do not prohibit a bank from communicating with a customer about potentially fraudulent or other suspicious transactions on the account, or from notifying the customer of an intention to close the account for such activity, so long as the communication does not reveal the existence of a SAR. Communications involving the underlying facts, transactions and documents on which a SAR is based are also not prohibited. The agencies say banks should consider customer communication case by case and take precautions where information could reveal a SAR.

    This matters for control design. A rule that blocks every mention of a restriction or review would be over-broad and would degrade legitimate customer communication. The rule you implement should separate two things: permitted statements about transactions or account actions, and statements that confirm, imply or hint at a SAR or an investigation that would reveal one. Have legal and compliance own that boundary and document how it maps to the joint statement.

    Map where investigation information can reach an interaction

    Start with an inventory, before writing any policy. For each system that holds or derives investigation-linked data, record whether an agent can see it and whether an AI tool can read it. AI-assisted tooling adds pathways that traditional access reviews may not cover, because a model can summarize or paraphrase content the agent never opened. Whether drafts, summaries or retrieval results can surface investigation content has to be tested in your own stack.

    Common pathways to check:

    • CRM records and free-text case notes that mention reviews, referrals or filings
    • Restriction and transaction-hold reason codes displayed in the agent desktop
    • SAR indicators or investigation status flags held in or synced to customer profiles
    • AI-generated summaries of account history, notes or prior interactions
    • Retrieval indexes that include case management exports, memos or internal guidance
    • AI-drafted replies that infer a reason for a restriction from whatever context they receive

    Define the layered control model

    No single layer is sufficient, and each has limits you should test. Together they cover what agents and tools can see, say and generate.

    • 1. Access minimization: Remove or mask investigation-linked fields and reason codes from agent and AI views, and keep SAR existence restricted to need-to-know roles. Limit: masking depends on correct data classification, and free-text notes are hard to classify.
    • 2. Approved neutral responses: Draft wording for restrictions, holds and planned closures with legal and compliance review against the joint statement. Limit: scripts depend on agents following them and cannot constrain an AI draft by themselves.
    • 3. Escalation to a designated team: Define triggers, such as a customer asking why an account is restricted, that move the interaction to the AML or compliance team. This aligns with the FFIEC expectation of a clear escalation process from initial detection to disposition. Limit: triggers must be specific enough to avoid constant escalation.
    • 4. Runtime output enforcement: Check AI-assisted interactions as they happen and act before the customer sees the result. Limit: output checks produce false negatives and false positives, so measure both.

    Where runtime policy enforcement fits

    Training and static scripts do not technically prevent a disclosure, and after-the-fact quality review finds one only after the customer has heard it. Runtime enforcement addresses that gap by placing policy in the execution path of AI-assisted tooling. Checks can sit at the input, retrieval, tool-call and output stages. Possible actions are to allow neutral wording, redact investigation-linked fields, block drafts that reveal or imply an investigation or SAR, and route a draft for human approval.

    Trussed enforces policy at runtime in this way: in the execution path of models, agents, tools and MCP servers, with allow, redact, block and escalate-for-human-approval actions applied to individual interactions and tool calls. Scope is the boundary to be clear about. Trussed does not replace case management, transaction monitoring or core banking access controls. Those systems remain the source of truth for investigations and for who may open a record. Runtime enforcement is a control on what AI-assisted tools retrieve and generate in a given interaction. Trussed is deepest in insurance, and does not claim financial-services customers or deployments.

    When you evaluate any runtime control, ask whether it can handle the permitted case. A draft that says a transaction is under review for suspected fraud may be allowed under the joint statement, while a draft that refers to a regulatory filing is not. The policy needs to express that distinction, and the false-positive rate on permitted communications needs to be measured, not assumed.

    Test, govern and keep the policy current

    Bring AI-assisted customer communication tools into your existing model risk management and independent testing process. Risk-based, reasonably designed controls with independent testing are what make an AML program defensible, and that applies to these tools as much as to monitoring systems.

    Test adversarially. Prompt the tools with questions such as why an account is restricted or whether the customer is under review, and measure how often investigation-linked content leaks before and after enforcement. Cover the scenarios the joint statement names: potentially fraudulent transactions, payment fraud and account closure. Review the policy whenever the agencies update their guidance.

    • Inventory pathways first, including reason codes and free-text notes
    • Version neutral responses and escalation rules so each can be tied to an interaction
    • Run adversarial prompts against agent-assist, summarization and retrieval tools
    • Track false negatives and false positives as ongoing metrics
    • Assign a named owner for the boundary between permitted and prohibited wording

    Four layers between investigation data and the customer

    Minimize access

    Mask or remove investigation-linked fields from agent and AI views.

    Neutral responses

    Approved wording for restrictions, holds and account closure.

    Escalate

    Route defined triggers to the designated AML or compliance team.

    Enforce at runtime

    Allow, redact, block or escalate AI drafts before the customer sees them.

    Show examiners the controls were enforced

    See how policy enforced in the execution path of AI-assisted tools produces audit-ready evidence for regulated enterprises.

    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.