Implementation Guide | Financial Services

    Preventing an AML Alert Triage Agent from Closing Alerts Without Human Review

    Do not rely on prompt instructions or after-the-fact log review. Split the agent's case management tools so that closing or dismissing an alert is a separate action, evaluate policy on every tool call in the execution path, and block or escalate closure requests for human approval before they execute. Retain the policy version, requested action, decision, approver, timestamp and outcome as evidence.

    Step 3: Specify the evidence each closure attempt must produce

    • Policy version in force at the time of the call
    • Requested action, target alert and arguments
    • Decision: allow, block or escalate
    • Approver identity and role, where escalated
    • Timestamp of request, decision and execution
    • Outcome: executed, denied or expired

    Why prompts and log review do not prevent unreviewed closures

    An alert triage agent usually reaches the case management system through tools. In many designs the tool that updates alert status can also close or dismiss the alert. A prompt that says "escalate, never close" is advisory. The model reads that instruction in the same natural-language channel as the alert narrative and customer data fields, so it has no hard separation between instructions and data. OWASP describes prompt injection as malicious instructions embedded in input that override intended behavior, with demonstrated outcomes that include unauthorized API calls.

    Injection is only one route to a bad closure. OWASP's Excessive Agency category covers damaging actions taken in response to unexpected, ambiguous or manipulated model output, whether the cause is hallucination, injection, a poorly engineered prompt or a poorly performing model. A misclassified alert that the agent judges low risk leads to the same result as a manipulated one: a potentially suspicious alert is closed with no human decision behind it.

    Post-hoc log review fails differently. It can show that a closure happened, but it cannot stop it, and it does not show that a human decision was required and made. The control has to sit where the agent invokes the closure action, and it has to run before the action executes.

    Step 1: Decompose the tool surface and define the disposition tiers

    Start with an inventory of every tool the triage agent can call, and mark each function that can change alert status or disposition. Remove anything the triage task does not need. OWASP identifies excessive functionality, meaning access to functions not needed for the intended operation, as a root cause, and recommends limiting tools to the minimum necessary.

    Then split the surface into five action classes. This decomposition is a design proposal rather than a regulatory taxonomy, so calibrate the tiers to your institution's own risk profile and model purpose. The revised interagency model risk management guidance (OCC Bulletin 2026-13, issued April 2026 with the Federal Reserve Board and FDIC) takes a risk-based approach and notes that appropriate practice differs between institutions.

    The autonomous tier covers read, annotate, recommend and escalate. The approval tier covers close and dismiss. The agent should not hold both recommend and close authority, which applies the separation of duties principle in NIST's agent identity design tenets.

    • Read: Retrieve alert, account and customer context. Read-only scope.
    • Annotate: Add analysis notes to the case without changing status.
    • Recommend: Propose a disposition and rationale for an analyst to consider.
    • Escalate: Route the alert to the correct human queue or authority.
    • Close or dismiss: Change disposition to a closed state. Requires human approval.

    Step 2: Enforce policy at the tool call, in the execution path

    Once closure is its own tool, place a policy decision point between the agent and that tool. Every call is evaluated before it executes, and the outcome is one of three: allow, block or escalate. Read, annotate, recommend and escalate calls pass if they fall within policy. A close or dismiss call is blocked outright, or held and routed to an authorized human approver, who can approve or deny it. OWASP and NIST both describe approval before execution as the checkpoint that injection alone cannot bypass, and the MCP guidance expects explicit consent before tool invocation and the ability for a human to deny it.

    Four design details determine whether the pattern holds up. Bind each approval to the specific alert, action and arguments, so it cannot be reused for a different closure. Give the agent task-scoped, short-lived credentials with no standing privilege, so the closure function is not reachable by a side path. Pin reviewed tool definitions with hashes and require re-review on change, so scoping is not silently altered. Define risk thresholds that trigger human review and route the request to the correct authority.

    This pattern is the one Trussed implements. It enforces policy in the execution path on individual interactions and tool calls, with the actions allow, redact, block, rate limit and escalate for human approval. Enforcement happens at runtime, not after the fact.

    Governance, tradeoffs and limits

    The revised interagency guidance emphasizes clear policies, roles and responsibilities, plus validation and monitoring, and describes generative and agentic AI as novel and rapidly evolving. As of November 2026, that points to some concrete governance work: name an owner for the closure policy, name who may approve, document how the control is tested, and schedule periodic re-validation of both the control and its policies.

    Testing should include injected instructions placed in alert narratives and customer data fields, and cases where the agent misclassifies risk. The pass condition is not that the agent behaves; it is that the closure call cannot execute without an approval. Anomaly monitoring of agent behavior can supplement this, but it does not replace pre-execution approval.

    The pattern has limits. Human approval is only as good as the review behind it. If approvers see high volumes of requests and approve them reflexively, the control degrades into a formality, so tune thresholds and queue design with that risk in mind. You also need a defined path for when no approver is available: the alert should stay open and visible, never time out into closure. Finally, enforcement at the tool call governs what the agent does through that tool. It does not judge whether the human's decision was correct.

    Questions to ask when evaluating an approach

    Whether you build this pattern or adopt a product, the same questions separate real enforcement from instruction. Can the agent's credentials technically reach a close or dismiss function, or is closure only possible through a separate approval path? Is policy evaluated at each tool call before execution, or only in prompts and later log review? Is each approval bound to a specific alert, action and arguments, and can it be reused?

    Ask also what record shows the policy version, requested action, decision, approver, timestamp and outcome, and how long it is retained and where it resides. Trussed supports self-managed deployment in the customer cloud with configurable retention and data residency, and produces audit-ready evidence from enforcement. A practical next step is to inventory the triage agent's tools this week and identify every function that can change disposition.

    Policy applied at each tool call

    Read

    Agent retrieves alert and customer context. Allowed.

    Annotate

    Agent adds notes to the case. Allowed within policy.

    Recommend

    Agent proposes a disposition. Recorded, not executed.

    Escalate

    Agent routes the alert to an analyst. Allowed.

    Close

    Blocked or held for human approval before execution.

    See enforcement at the tool call

    Review how policy is enforced in the execution path and how approval decisions become audit-ready evidence.

    Request a Demo