FINRA Generative AI Testing Evidence: What Examiners May Expect and How Runtime Evidence Complements It

What FINRA has actually said

FINRA states that its rules, which are intended to be technologically neutral, and the securities laws more generally continue to apply when firms use generative AI, as they do for any other technology. That holds whether a firm builds the tool or uses a third party's technology, including embedded features in existing third-party products. Regulatory Notice 24-09 adds that a firm should evaluate GenAI tools before deploying them and ensure it can keep complying with the FINRA rules that apply to the business use of those tools. Which rules apply depends on how the firm deploys the technology.

The 2026 Annual Regulatory Oversight Report, published on December 9, 2025, covers GenAI as a topic. It states that under Rule 3110 a firm must have a reasonably designed supervisory system tailored to its business. It lists testing as a consideration: robust testing of GenAI to understand the capabilities, limitations and performance of the model, with privacy, integrity, reliability and accuracy as areas to consider. It also notes that when GenAI generates or helps create customer communications, those communications must comply with applicable securities laws and FINRA rules.

Rule 3110 itself requires written supervisory procedures. Reviews of correspondence and internal communications must be conducted by a registered principal and evidenced in writing, electronically or on paper. Confirm the current rule text before relying on it, since rule changes can be pending.

What these documents do not do is prescribe an artifact list. Everything below that goes beyond these statements is reasonable inference, and it is labeled that way.

Mapping published expectations to evidence categories

The table groups evidence by what each category can demonstrate and where it stops. The categories are an organizing structure for examination readiness, not a FINRA-defined list.

• Pre-deployment testing: Anchored in the oversight report's testing consideration and the Notice 24-09 call to evaluate before deployment. Shows behavior on tested inputs at a point in time. Does not show what happened afterward.
• Supervisory design: Anchored in Rule 3110 written procedures and principal review evidenced in writing. Shows who is accountable and how GenAI use is reviewed. Does not show model accuracy.
• Runtime control: Inference, not a stated FINRA requirement. Per-interaction records can show what a system was permitted to do and did in production. Does not substitute for testing of capabilities or accuracy.
• Records retention: Applies to all of the above. Retention and access for testing and supervisory records should be defined against applicable books-and-records requirements.

Pre-deployment testing and supervisory design

For testing, the oversight report's four areas (privacy, integrity, reliability and accuracy) are a sensible scope for a test plan. FINRA's 2015 algorithmic trading notice, Regulatory Notice 15-09, concerns a different technology, but it is a useful analogy for what a testing record can contain: independent quality assurance, data integrity and workflow validation, segregated development environments, and records of testing protocols, results and remediation of significant defects. Bank model risk guidance points the same way. SR 11-7 describes test documentation covering the purpose, design and execution of test plans, summary results with commentary, and analysis of informative samples. SR 11-7 has since been revised as SR 26-2. Both are addressed to banking organizations, so using them as a benchmark for a broker-dealer is an inference, not a FINRA expectation.

A practical test file therefore holds the plan, scope, results, commentary and remediation, tied to a specific tool and business use. It should cover vendor-provided and embedded GenAI features as well as in-house builds, because FINRA rules apply to both.

Supervisory design is the second category. Written procedures should name the GenAI uses in scope and describe the principal review of communications that Rule 3110 requires to be reviewed, with that review evidenced in writing. For GenAI that generates or helps create customer communications, document how compliance with securities laws and FINRA rules is checked. Assign accountability for the GenAI supervisory system. Record initial and ongoing vendor due diligence for third-party tools. FINRA lists that as an effective practice for mission-critical vendors generally, not for GenAI specifically.

Where runtime enforcement evidence fits

A passing test report answers a narrow question: how did the system behave on the inputs we tested, before release? An examiner can reasonably follow with a different one: what was this system allowed to do in production, and what did it do? Tests do not answer that. Written principal review covers the communications that are reviewed, and ongoing monitoring records cover what a firm chooses to track. Neither necessarily records each permitted or restricted action.

Runtime enforcement evidence addresses that gap. If policy is enforced in the execution path, each interaction can produce a decision record: allowed, redacted, blocked or escalated for human approval, with the policy that applied. This fits the ongoing monitoring and response-to-problems elements that model risk guidance describes. It does not cover conceptual soundness or accuracy testing, and no FINRA publication states that examiners request it. Treat it as a reasoned complement.

Trussed provides this layer as runtime control. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, applying allow, redact, block, rate limit and escalate-for-human-approval decisions to individual interactions and tool calls. Policy is enforced at runtime, not reviewed after the fact, and the enforcement itself generates continuous audit-ready evidence. Deployment is self-managed in the customer cloud, with configurable retention and data residency, which matters when retention periods must be set against books-and-records requirements.

Trussed is not a pre-deployment testing or red-teaming product, so it does not produce the test file described above. Its deepest experience is in insurance, another regulated financial sector, and it runs 100+ models under one policy plane. The value for a broker-dealer compliance team is the evidence it produces under examiner scrutiny and model risk management review, kept distinct from testing evidence.