Key takeaways
- FINRA Rule 3110 Supplementary Material .06 allows risk-based review of correspondence and internal communications.
- Risk-based review requires training, documentation of training, and surveillance and follow-up.
- Evidence of review must identify the reviewer, the communication and the date, and merely opening it is not review.
- A principal may delegate first-pass review but remains ultimately responsible.
- You can only sample AI use that your approved tools actually log.
Does FINRA require you to review every AI prompt?
No. FINRA has no AI-specific review rule, and Rule 3110 lets a firm use risk-based principles to decide how much correspondence and internal communication to review. What it does require is a reasonably designed system, evidence that reviews happened, and follow-up on what they find.
FINRA Rule 3110 requires each member to have "a system to supervise the activities of each associated person that is reasonably designed to achieve compliance," and paragraph (b)(4) requires procedures to review incoming and outgoing written correspondence and internal communications relating to the firm's investment banking or securities business. Regulatory Notice 24-09 reminded members that these technology-neutral rules apply when they use generative AI, whether they build the tool or license it from a third party.
If a registered representative pastes a client email into an AI assistant and asks it to draft a reply, the prompt, the output and the sent message can all touch obligations you already supervise. A monthly sample is a practical way to show that supervision reaches AI use without reviewing thousands of routine prompts.
What does Rule 3110 say about risk-based review?
Supplementary Material .06 lets a firm decide, using risk-based principles, how far its review procedures go beyond the subject matters the rule requires. If procedures do not require every piece of correspondence to be reviewed before use, they must include training on the procedures, documentation of that training, and "surveillance and follow-up to ensure that such procedures are implemented and followed."
A monthly AI prompt sample is that surveillance and follow-up. It works alongside a written AI use policy and training, and it should be described in your written supervisory procedures with the same precision as your email review.
What does FINRA AI guidance expect firms to capture?
FINRA's 2026 Annual Regulatory Oversight Report lists, among practices firms may consider, "storing prompt and output logs for accountability and troubleshooting," tracking "which model version was used and when," and "ongoing monitoring of prompts, responses and outputs." The same report says GenAI use "can implicate rules regarding supervision, communications, recordkeeping and fair dealing."
You cannot sample what you do not capture. Before a monthly review is possible, firm-approved AI tools need prompt monitoring that logs prompts and outputs to a store the compliance team can search. Where AI-assisted messages relate to the firm's business as such, SEC Rule 17a-4(b)(4) requires broker-dealers to preserve communications received and sent for at least three years. Whether a given prompt log is a required record is a question for your compliance counsel, but the sample cannot run without the log.
Our earlier piece on supervising GenAI under FINRA Rule 3110 covers what to log in more detail, and how Rule 2210 applies to AI-generated communications covers content that reaches the public.
How do you build a monthly sample?
Combine a risk-based targeted pull with a smaller random pull, so the review catches known risks and also tests what you did not think to look for.
- Define the population. All prompts and outputs from firm-approved AI tools for the calendar month, by user. Note any tool that does not log, because that is a gap in itself.
- Targeted pull. Run lexicon searches for terms that signal regulatory risk: guarantee, promise, risk-free, recommend, performance, outside business, complaint, wire, password and account numbers. Include every hit for representatives on heightened supervision.
- Random pull. Add a fixed random sample per user or per branch so that every user has some chance of review each month. Keep the method the same month to month so results are comparable.
- Customer data check. Flag prompts containing client names, account numbers or Social Security numbers going to any tool not approved for that data. FINRA's 2026 report lists, as an effective practice, vendor contract language that prohibits firm or customer sensitive information from being ingested into a third-party vendor's open-source GenAI tool. The sample checks whether your own staff keep to the same line.
- Size it to your risk. Rule 3110 sets no sample size. Document why your size is reasonable for your headcount and AI use, and increase it if exceptions rise.
Who should do the review, and how should it be evidenced?
A registered principal must own the review, but first-pass work can be delegated. Supplementary Material .08 allows a supervisor or principal to "delegate certain functions to persons who need not be registered," while remaining "ultimately responsible for the performance of all necessary supervisory reviews."
Evidence is where most programs fail. Supplementary Material .07 requires the evidence of review to be "chronicled either electronically or on paper," to identify the reviewer, the communication reviewed and the date, and to record actions taken on significant regulatory issues. It also states that "merely opening a communication is not sufficient review." For each sampled item, record the reviewer, date, disposition (no issue, coaching, escalation) and any follow-up.
What should happen when the sample finds a problem?
Every finding needs a disposition, an owner and a closing date, and patterns need to change the procedures. A sample that finds problems and changes nothing is weak evidence of supervision.
Typical dispositions include coaching the user, escalating to the CCO, opening a complaint or communications review under existing procedures, and blocking a tool or a data type. Track repeat issues by user and by type. If prompts that promise performance keep appearing, add training and expand the lexicon. If client data keeps reaching an unapproved tool, change the technical controls.
Summarize the month's results in one page: population size, sample size, method, findings by category and actions. Those monthly summaries become inputs to the annual testing and supervisory controls report under FINRA Rule 3120.
What should you do this quarter?
Get the capture, the written procedure and the first two months of evidence in place before your next annual supervisory controls review.
- List every AI tool in use, approved or not, and confirm which ones log prompts and outputs to a store compliance can search.
- Write the sampling method into your written supervisory procedures: population, lexicon, random sample size, reviewer, evidence fields and escalation path.
- Assign the reviewing principal and any delegates, and document the delegation.
- Run the first monthly sample, record every disposition and write the one-page summary.
- Review the lexicon and sample size after three months and record why you kept or changed them.
How can s90 help?
s90's prompt monitoring captures prompts and outputs from approved AI tools, applies a firm-defined lexicon and flags sensitive client data, then delivers a monthly sample with fields ready for principal review. The output is formatted as exam evidence, and s90 commits to producing it within 24 hours of an examiner request. If you would like to see how a monthly sample would look for your firm, a short walkthrough is a good place to start.
