Preparing AI Client Communication and Advisor Assistant Evidence for a FINRA Cycle Exam
Treat AI like any other communication and supervision channel. Inventory every AI use case, including vendor-embedded features. Map each to Rules 2210, 3110 and 4511 and SEA Rule 17a-4. Define the evidence each obligation needs, and make sure each interaction can be tied to the policy that applied to it. The largest gaps are usually records that link an interaction to its policy, its reviewer and its archive.
Where runtime policy enforcement fits, and where it does not
- 1
From the enforcement layer
Per-interaction and per-tool-call records of the policy applied and the action taken: allow, redact, block or escalate for human approval.
- 2
From outside the enforcement layer
Written supervisory procedures, training records, surveillance and follow-up, archived communications and chat sessions, and retention decisions.
- 3
Model risk documentation
Pre-deployment evaluation, approved-use definitions and ongoing review remain firm-owned records.
- 4
Questions to ask any tool
Does the record show which policy version applied, the action taken, and who approved an escalation? Can it be linked to the archive?
Step 3: Gap review before the request arrives
- Does the inventory include AI features embedded in third-party products?
- Is pre-deployment evaluation documented for each generative AI tool?
- Do written supervisory procedures explicitly cover each AI use case?
- Are chatbot sessions with investors supervised, retained and traceable to an archive location?
- Can you show which policy applied to a given interaction and who reviewed or approved it?
- Is the decision on retaining prompts and drafts documented, with rationale and counsel input?
What an examiner is likely to test
FINRA describes its rules as technology neutral. Regulatory Notice 24-09, issued in June 2024, reminds members that existing rules apply when they use generative AI, whether they build the tools or use third-party technology, including AI features embedded in existing products. The 2026 Annual Regulatory Oversight Report lists two effective practices for generative AI: ensuring AI-assisted customer communications comply with federal securities laws and FINRA rules, and ensuring appropriate supervision and retention of chatbot communications and chat sessions used with investors.
In practice, three rule sets shape the evidence request. Rule 2210 governs content standards and the category of each communication. Rule 3110 requires a reasonably designed supervisory system tailored to the business, with written procedures under 3110(b)(1). Rule 4511 and SEA Rule 17a-4 govern what is preserved, for how long and in what format. This mapping of evidence to rules is a practical inference, not a FINRA checklist, but it gives the exam preparation a workable structure.
One proposal is worth tracking. As of November 2026, Regulatory Notice 26-14 proposes changes to the retail communications rules. It restates that members are responsible for communications whether generated by a human or AI. It would also require evidence that supervisory procedures were implemented and carried out. It is a proposal, and final requirements may differ, but the direction suggests preparing implementation evidence now, not only written procedures.
Step 1: Inventory use cases and assign a Rule 2210 category
Start with a use-case inventory that covers more than tools the firm built. Because the rules apply to embedded third-party features, include AI functions inside CRM, email, meeting and productivity products. An examiner who asks what AI touches client communications expects a complete answer.
For each entry, record the Rule 2210 category. Chatbot output may be correspondence, institutional communications or retail communications depending on its nature, and the category determines which content and review standard applies. Advisor assistants that draft text an advisor then sends raise a separate question: the advisor sends it, but the firm remains responsible for it.
- Use case name, business owner and named accountable approver
- Rule 2210 category and the audience the output reaches
- Whether the capability is built in-house, third-party or embedded in another product
- Who uses it, and whether output reaches clients directly or through an advisor
- Pre-deployment evaluation record and the existing rules it was assessed against
- Written supervisory procedure that covers the use case
Step 2: Define the evidence each obligation requires
Define evidence per obligation, then identify where it lives today. For content standards, the evidence is the approved-use policy, the review or approval record for AI-assisted communications, and the training records for the people who use the tools. For supervision, it is the written procedures that explicitly cover generative AI, the pre-deployment evaluation, and the surveillance and follow-up records showing procedures were carried out. For retention, it is proof that required records are preserved in an easily accessible place, in a format and media that meet SEA Rule 17a-4.
The typical problem is dispersal. Prompt histories sit with the vendor, interaction logs in another system, approvals in email or a ticketing tool, policies in a document repository, and final communications in the archive. Each piece may be sound, but an examiner's question usually crosses them. Can you show, for a given AI-assisted communication, which written procedure applied and who reviewed or approved it? If answering takes manual reconstruction across four systems, that is the gap to close.
Chatbot sessions deserve their own line in the evidence map. The oversight report names them specifically for supervision and retention, so session records are a distinct item from drafts of written communications.
Retention: what is clear and what needs a documented decision
Some points are settled. Rule 4511 requires members to make and preserve required books and records, and records with no specified retention period must be kept at least six years. Records must be in a format and media that comply with SEA Rule 17a-4. The Books and Records Rules cover communications sent and received relating to the business, with examples such as emails, instant messages, text messages, chat messages and interactive blogs. The SEC's 17a-4 amendments took effect January 3, 2023, and firms relying on 17a-4(f) must have filed new undertakings with FINRA.
Other points are not settled. Whether AI prompts, intermediate drafts and iterations are themselves required records is an open question. Industry comment letters have asked FINRA to clarify whether Rule 4511 reaches only final AI-assisted communications or also drafts and prompts, and how review of AI-assisted content should be documented. Those letters are advocacy, not FINRA guidance.
Until the position is clearer, decide the scope with counsel and write down the rationale. An examiner is more likely to accept a documented, reasoned retention decision than an undocumented default. Design records so each AI interaction can be linked to the retention system of record, which keeps either decision workable.
Governance and next steps
Assign accountability per use case, since responsibility for AI-generated communications stays with the firm. Keep training, surveillance and follow-up records current, and track Notice 26-14 and any FINRA response to the recordkeeping questions so procedures can be updated if rules are adopted.
For firms evaluating runtime enforcement, Trussed has run continuously for more than 18 months in a regulated enterprise, at 99.95% uptime on live production traffic, with over 100 models under one policy plane and more than 30 GB of audit evidence generated per day. It supports self-managed deployment in the customer cloud with configurable retention and data residency. Use the evidence questions above as the evaluation criteria for any tool, including this one.
From AI use case to exam-ready evidence
Inventory
Every AI-drafted communication, chatbot and advisor assistant, including vendor-embedded features.
Map
Each use case tied to Rule 2210, Rule 3110 and Rule 4511 obligations.
Evidence
Procedures, interaction records, review records and retention defined per obligation.
Gaps
Policy enforcement, supervision records and archiving tested before a request arrives.
Test your AI evidence chain before an examiner does
Talk through your AI use cases and see how runtime policy enforcement produces per-interaction evidence alongside your supervisory and archiving systems.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.