Checklist for Financial Services Compliance

FINRA Rule 3120 Supervisory Control Testing for AI Agents

Rule 3120 requires designated principals to test and verify that a firm's supervisory procedures are reasonably designed, amend them where testing shows a need, and report results to senior management at least annually. For AI agents, that means inventorying every agent and tool, tying each control to a written procedure, defining a test for each control, and retaining evidence that shows what each agent was permitted to do and what it did. This checklist covers each step as of October 2026.

    What Rule 3120 requires and what it leaves to the firm

    Rule 3120(a) requires a member to designate and identify to FINRA one or more principals who establish, maintain and enforce supervisory control policies and procedures. Those policies must test and verify that the firm's supervisory procedures are reasonably designed to achieve compliance with applicable securities laws and FINRA rules. Where testing identifies a need, they must also create additional or amended procedures. The designated principals must report to senior management at least annually: the system of supervisory controls, a summary of test results and significant exceptions, and any new or amended procedures created in response.

    FINRA treats supervisory control policies as distinct from written supervisory procedures. The written procedures document the supervisory system under Rule 3110(b). The control policies test and verify that those procedures are reasonably designed. A firm needs both, so an agent control and the test of that control should be separate artifacts.

    The rule does not prescribe how to test. FINRA says risk-based methodologies and sampling may be used to set the scope of testing. FINRA also states that its rules are technology neutral and apply to generative AI as to any other tool, including third-party tools and embedded features. Its guidance speaks to generative AI and language models rather than agents by name, so the mapping below is a practical interpretation, not FINRA text. It treats each agent action (a tool call, a data access, an output) as a discrete event that can be controlled and tested.

    What runtime enforcement evidence can and cannot show

    When policy is enforced in the execution path, each decision is made at the moment of the interaction or tool call, and the record of that decision is a by-product of enforcement. That differs from a periodic review after the fact, which reconstructs behavior from whatever was captured. Trussed enforces policy in the execution path across models, agents, tools and MCP servers, applying allow, redact, block, rate limit or escalate for human approval to individual interactions and tool calls. It produces continuous audit-ready evidence from that enforcement. For a Rule 3120 test, those records can show what an agent was permitted to do, what decision was applied and when an escalation occurred.

    The limits matter as much as the capability. Enforcement evidence shows that a control operated. It does not show that the underlying procedures are reasonably designed, that the policies cover every agent or tool, or that no unmonitored path exists. Those questions remain with the designated principal and need inventory reconciliation and separately designed tests. Trussed is not a pre-deployment testing product and does not deliver Rule 3120 compliance on its own. It supplies one category of evidence that the firm's own test program then evaluates.

    Three questions help assess any enforcement approach. Does it sit where agent actions occur, including third-party and embedded features? Are policy definition, enforcement and the evidence store separate enough that a reviewer can confirm the policy in force at a given time? Can the firm control retention and data residency to match its own recordkeeping decisions? Trussed supports self-managed deployment in the customer cloud with configurable retention and data residency, and has run continuously in production in a regulated enterprise for more than 18 months.

    Governance and ownership decisions

    Rule 3120 places responsibility with designated principals, and the firm's control policies should document who owns each agent control. In practice that often spans compliance, engineering and model risk. Write the split down, because a test with no clear owner tends to produce an exception with no clear remediation.

    FINRA lists formal review and approval processes, involving business and technology experts, as a practice for generative AI. Apply the same idea as an approval gate for adding an agent, a tool or a data scope, and make the gate itself a testable control. If any AI is used as part of the supervisory system itself, FINRA indicates that the model's integrity, reliability and accuracy may be considered in the firm's procedures, so include it in the test scope.

    Model risk management review will usually ask for the same artifacts: inventory, access scope, policy records and escalation records. Building one evidence set that serves both the Rule 3120 test and that review avoids parallel documentation. Treat the artifact list here as a practitioner suggestion rather than a statement of what FINRA examiners require for agents.

    Practical next steps

    Begin with the inventory and its reconciliation to production, since every later test depends on them. Next, pick two or three agents with the broadest tool or data access and run the full chain once: procedure, control, test, evidence, exception, report. The gaps that surface in that pass, such as a control with no named test or an action with no record, are the first items for the amended procedures the rule expects.

    Then set the cadence. The rule requires reporting at least annually, but agents change as tools and data scopes change, so tie re-testing to the approval gate as well as to the calendar. Keep the annual report factual: what was tested, what failed, what changed.

    From Rule 3120 to agent control tests

    Inventory

    Agents, tools, data scope, owner, origin

    Policy

    Each control tied to a written procedure and a test

    Enforcement

    Controls applied at each interaction and tool call

    Escalation

    Human approval recorded for defined actions

    Evidence

    Records that show what was permitted and what happened

    Reporting

    Annual summary, exceptions and amended procedures

    Review your agent evidence against your test plan

    Talk through how enforcement in the execution path and the evidence it produces could support the tests in this checklist.

    Explore Agent and Tool-Call Enforcement