Risk-Based Review of AI-Generated Correspondence Under FINRA Rule 3110(b)(4): Sampling vs Full Capture

What Rule 3110(b)(4) requires and how it reaches AI output

Rule 3110(b)(4) requires supervisory procedures for reviewing incoming and outgoing written, including electronic, correspondence and internal communications relating to the member's investment banking or securities business. For correspondence, the procedures must identify and handle customer complaints, instructions, funds and securities, and communications whose subject matter requires review under FINRA rules and federal securities laws. Internal communications must be reviewed to identify those of a subject matter requiring review. The review must be conducted by a registered principal and evidenced in writing, electronically or on paper.

Supplementary Material .06 adds that, by employing risk-based principles, a member decides how far additional policies and procedures are needed for correspondence outside the listed subject matters, given its business and structure. FINRA's Rule 2210 FAQ says the procedures must suit the member's business, size, structure and customers.

For AI, FINRA's position is that firms are responsible for their communications whether a human or AI generated them, and that its rules are technology-neutral. Regulatory Notice 24-09 makes the same point for Rule 3110 generally: a reasonably designed supervisory system tailored to the firm's business. The FAQ adds that chatbot output may be correspondence, retail communications or institutional communications, depending on its nature and the number of recipients. Classification therefore comes before any capture or review decision, because the applicable standard changes with distribution.

Capture and review are separate design choices

Capture is the act of recording the AI interaction. Review is the act of a registered principal examining some or all of it and recording the outcome. The rule text governs the second. Retention is a separate question: as FINRA describes Exchange Act Rule 17a-4(b)(4), firms preserve originals of communications received and copies of communications sent that relate to the broker-dealer's business as such. Regulatory Notice 25-07 says AI-generated chatbot communications and transcripts raise novel recordkeeping questions. Whether prompts and intermediate drafts are business communications is a determination the firm should make with counsel and document.

Treating the two choices separately removes a common false trade-off. A firm can capture an interaction in full yet review only a defined subset. A firm can also decide that a low-risk type needs only a retained output and no routine review. What it cannot do is leave the reasoning undocumented. FINRA's 2026 oversight report lists prompt and output logs, model-version tracking and human-in-the-loop review as things a reasonably designed supervisory system may include. These are examples, not mandates, and they map to the capture and review decisions above.

Sampling, targeted review and full capture compared

No option is correct in the abstract. Each trades risk coverage, evidence quality and operational burden differently, and the right choice can differ across communication types within one firm. Sampling relies on the sampling design being defensible, which matters more when AI output is high in volume and variable in wording.

• Sampling: Risk: unreviewed items may contain exceptions, and variable AI output makes the sample harder to defend. Evidence: shows the method and the reviewed items, but not coverage of the rest. Burden: lowest, scaling with the sample rate.
• Targeted review: Risk: depends on how well triggers match real exposure, such as topic, recipient type or prior exceptions. Evidence: strong for the flagged population if trigger logic and escalations are recorded. Burden: moderate, and the triggers need maintenance.
• Full capture, with review scoped by risk: Risk: lowest gap in what can be examined later, since every interaction is recorded. Evidence: complete record to retrieve and test. Burden: highest for storage, retrieval and retention, and review volume if all items are reviewed.
• Full capture and full review: Risk: widest coverage. Evidence: strongest, with a reviewer action for each item. Burden: heaviest, and may be reserved for high-risk types such as autonomous client-facing output.

Where runtime policy enforcement fits, and where it does not

Rule 3110(b)(4) requires review by a registered principal, evidenced in writing. An automated control sits alongside that review and does not replace it. What enforcement in the AI execution path can do is act on each individual interaction before output reaches a recipient: allow it, redact content, block it, or escalate it for human approval. Each decision is recorded as it happens. That supports a documented risk-based approach in two ways. It applies the firm's written policy for a communication type consistently at runtime, rather than after the fact. It also produces records of the policy decision for each interaction, which are the kind of evidence FINRA's oversight report points to, such as prompt and output logs and model-version tracking. Whether a specific record satisfies an examiner is for the firm to assess.

The limits are plain. Enforcement does not classify communications for the firm, set the sampling rate, perform principal review or settle the recordkeeping analysis. Trussed enforces policy in the execution path of individual interactions and tool calls and generates audit-ready evidence from that enforcement. It is not a supervision, archiving or surveillance system, and it does not trace end-to-end agentic workflows. Firms using AI inside their own review process should also apply the technology governance Notice 24-09 describes.

FINRA Regulatory Notice 26-14 proposes changes to Rule 2210 communications review standards and, as of November 2026, is a proposal only. Firms should track it, since it discusses the challenge of applying principal pre-use approval to AI-generated retail communications.