Implementation Guide | Financial Services

    Stopping a KYC Review Agent from Approving the Onboarding Case It Prepared

    A KYC review agent should never be able to approve a case it prepared. Give the preparer and the approver separate identities, remove approve and release tools from the preparer's scope, and evaluate every approval-related tool call in the execution path. Block the call, or escalate it for human approval, when the requester is the preparer. Prompt instructions alone are not enough, because approval is a tool call the agent can still make.

    A runtime policy pattern that blocks and escalates

    1. 1

      1. Prepare under a distinct identity

      The preparer agent runs with its own identity and task-bounded scopes, and the case record stores that identity as preparer of record.

    2. 2

      2. Intercept the approval call

      Any call to an approval-related tool is evaluated in the execution path before it executes.

    3. 3

      3. Evaluate identity and scope

      Policy checks the requesting identity, the tool, and whether the requester matches the preparer of record.

    4. 4

      4. Block self-approval

      If the requester is the preparer, or lacks approval scope, the call is blocked and the attempt is recorded.

    5. 5

      5. Escalate for human approval

      Calls that require sign-off, including red-flag cases, route to a named reviewer role under written escalation procedures.

    6. 6

      6. Allow only distinct approvers

      The approval proceeds only when the approver is a separate, accountable identity, and the decision is recorded.

    Why agent self-approval is a control failure

    Maker-checker exists so that the party who prepares an onboarding decision is not the party who approves it. NIST SP 800-53 Rev. 5 describes separation of duties as dividing mission or business functions among different individuals or roles. It says the control addresses the potential for abuse of authorized privileges and reduces the risk of malevolent activity without collusion. It also says separation of duties is enforced through account management, access control mechanisms and identity management.

    That last point matters for agents. If a single agent identity holds both the case-assembly tools and the approve or release tool, the split does not exist at the system level, whatever the process document says. The agent does not need to be malicious for this to fail. A misread instruction, an unexpected tool selection or a prompt injection can all lead to the same approval call.

    Supervisory attention to this pattern is already visible. FINRA's 2024 AML oversight report lists auto-approving customer accounts despite red flags as a finding, and it describes written escalation procedures as an effective practice. FINRA's report applies to FINRA members, so confirm how it maps to your own regulators. The underlying expectation is portable: an approval that bypasses independent review is hard to defend.

    Identify the approval-related tool calls and identities

    Start by documenting which KYC duties require separation. NIST SP 800-53 Rev. 5 asks organizations to identify and document the duties of individuals requiring separation and to define system access authorizations that support it. For an onboarding agent, that means preparation on one side and approval or release on the other.

    Then inventory the actions, not just the systems. Anything that finalizes or changes a decision belongs on the approval side of the line. Remove these from the preparer agent's credentials or tool list before relying on any prompt instruction.

    • Approve case: Marks the onboarding case as approved.
    • Release or activate account: Moves the customer from pending to active.
    • Override risk rating: Changes the rating the preparer or a rule produced.
    • Clear screening alert: Dispositions a screening hit as resolved.
    • Shared service accounts: One credential used by both preparer and approver paths makes every action unattributable.
    • Combined tool servers: An MCP server or tool bundle exposing both read and approve actions can hand the preparer approval capability by default.

    Evidence an examiner or model risk function can inspect

    The control is only as useful as the record it leaves. The aim is to show what the agent was allowed to do and what it did, including the attempts that were stopped. A useful per-decision record includes the agent identity, the tool called, the policy evaluated, the decision (allow, block or escalate), the escalation target, the human approver and a timestamp.

    This matters for model risk management as well as audit. FINRA Regulatory Notice 24-09 says a firm using Gen AI tools within its supervisory system should have policies and procedures covering technology governance, including model risk management and the reliability and accuracy of the model. A retrievable record of blocked and escalated approvals gives that function something concrete to test.

    Trussed enforces policy on individual interactions and tool calls in the execution path, with allow, redact, block, rate limit and escalate for human approval as outcomes. The audit-ready evidence is generated by that enforcement and mapped to frameworks including NYDFS, the EU AI Act and NIST. Whatever tooling you use, check that evidence is produced at the moment of enforcement, with configurable retention, rather than reconstructed afterward.

    Limits of tool-call-level enforcement

    A tool-call policy evaluates one call at a time, using the attributes available with that call. It is strong where the question is local: is this identity permitted to invoke this tool, and does it match the preparer of record the call carries. It is weaker where the question spans systems. NIST notes that separation-of-duties violations can span systems and application domains, so policy should consider the whole system. A conflicting privilege held through another credential or application may be invisible to a check on a single call.

    There is also a dependency on the preparer-of-record attribute. Enforcement at the level of interactions and tool calls does not supply case-level or cross-agent transaction identity. If the preparer and approver are linked only in a case management system, that link must be passed into the call, or the policy cannot use it.

    Compensate with practical measures. Review credentials across systems, not only the agent's tool list. Test the control by having the preparer agent attempt the approval tool directly, and record whether each attempt is blocked or escalated. Test that red-flag cases cannot be released without independent review. Keep escalation paths written, with named reviewer roles, so an escalated approval reaches an accountable person.

    Preparer and approver split, enforced per tool call

    Preparer identity

    The agent assembles identification, beneficial ownership, screening results and risk rating under its own scoped identity.

    Approval tool call

    Any attempt to approve, release or override is intercepted before it reaches the system of record.

    Runtime policy

    Policy compares the requester with the preparer of record and decides to allow, block or escalate.

    Evidence record

    Each decision is retained with identity, tool, policy, outcome and approver.

    See how approval calls are enforced and recorded

    Talk through how policy enforcement on tool calls and audit-ready evidence could apply to your onboarding agent design.

    Request a Demo
    Next step

    How does your AI governance program compare?

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