GenAI in Regulated Communications: Supervision and Recordkeeping for AI-Assisted Client Messaging

The moment a language model drafts something that ends up in front of a client, the output stops being a productivity gain and becomes a supervised communication and a record the firm is required to keep. Most firms wire in the AI long before they wire in either of those.

The pitch for generative AI in a broker-dealer or adviser is that it drafts faster: the client email, the account summary, the response to a routine information request, the first pass at a market note. The part that gets skipped is that none of the existing obligations move out of the way to make room for it. A message drafted by a model and sent to a client is exactly as much a communication with the public, and exactly as much a book and record, as one a human typed. The technology changed. The rulebook did not.

The rule set does not care that a machine wrote it

FINRA said this plainly in Regulatory Notice 24-09, published in June 2024. Its rules are technology-neutral and continue to apply when a member firm uses generative AI and large language models. The notice created no new requirements and no new interpretations; it reminded firms that the obligations they already carry attach to the output regardless of what produced it. Two of those obligations do the heavy lifting here: Rule 3110 on supervision and Rule 2210 on communications with the public.

The notice also flagged, under the supervision heading, that a firm deploying AI in its systems has to address technology governance, model risk management, data integrity, and the reliability and accuracy of what the tool produces. That is the governance frame. The operational problem sits one level down, in the plumbing, and it is where firms actually get caught.

Where the record actually is

The books-and-records regime for a broker-dealer runs through SEC Rules 17a-3 and 17a-4 and FINRA Rule 4511. Between them they say, in effect: make the record, keep it for the required period, and keep it in a form a supervisor can retrieve and rely on. Client communications are squarely in scope. That has been true for email and for chat for years, and it is true for a message a model helped write.

The gap opens because of where AI-assisted drafting happens. The model output is generated in a copilot side panel, a browser tab, a vendor chat window, or an internal tool bolted onto the CRM. The registered representative reads it, edits it, and pastes the result into the channel the client actually receives it on. The sent message lands in the archive because the email or messaging platform is already captured. The interaction that produced it, the prompt, the raw draft, the edits, frequently lands nowhere. If a supervisor later asks how a particular client note was arrived at, or whether a model produced a claim that should never have gone out, the firm can show what was sent but cannot reconstruct how it came to be sent. That is a supervisory blind spot the firm built itself.

Supervision attaches to the output, not the tool

Rule 2210 sorts communications by audience and sets the review obligation accordingly. A retail communication, meaning one distributed or made available to more than 25 retail investors within any 30 calendar-day period, must be approved by an appropriately qualified registered principal before the earlier of its use or filing. Correspondence, meaning 25 or fewer retail investors in that window, does not need pre-use approval but remains subject to supervision and review under Rule 3110. Institutional communications need written review procedures rather than sign-off on each item.

Drop AI drafting into that structure and the question is not whether review applies. It is whether the firm can still tell which tier a given message belongs to, and whether the principal approving a retail piece knows a model wrote it and can see what it wrote before the edits. A model that drafts a template one rep then sends to 400 clients has produced a retail communication that needs principal approval before use, regardless of how quickly it was generated. Volume is the thing AI is good at, and volume is precisely what pushes correspondence over the line into the pre-approval tier. The content standards, no misleading claims, no unwarranted promises, fair and balanced, apply identically whoever or whatever drafted the text.

What the archive is now expected to do

The SEC amended Rule 17a-4 in 2022, effective January 2023 with a broker-dealer compliance date of 3 May 2023. The headline change was to allow an audit-trail alternative to the long-standing write-once-read-many (WORM) requirement. A firm can now preserve electronic records on a system that maintains a complete, time-stamped audit trail, recording every modification and deletion and the identity of who made it, such that an original record can be recreated if it is altered. WORM is no longer the only compliant path.

That amendment matters directly for AI-assisted content, because the audit-trail model is exactly the shape of the evidence a supervisor will want: not just the final message, but who touched it, when, and what changed. The same discipline that makes an access audit trail defensible, capturing actor, timestamp and mutation rather than just end state, is what turns an AI drafting workflow from an unrecorded side channel into something you can stand behind under inspection.

Wiring capture and review into the flow

The fix is not a policy document that says representatives must use AI responsibly. It is capture and review built into the path the message travels, so that using the model cannot route around supervision. In practice that means a small number of concrete decisions:

• Decide which AI drafting tools are permitted for client-facing content, and block the rest. An unapproved copilot pasting into a personal channel is the same governance failure as any other off-channel communication.
• Capture the AI-assisted draft into the same archive as the sent message, tagged as machine-assisted, so the retention clock and the supervisory review cover it too.
• Route retail-tier output through principal approval before use, and make the approver able to see what the model produced, not only the polished version.
• Record who approved what, against the audit-trail standard 17a-4 now permits, so the evidence of supervision is a by-product of the workflow rather than a reconstruction after the fact.

The instinct to enforce this at the point of use, rather than trust everyone to remember, is the same instinct behind a control that actually runs in the pipeline instead of living in a wiki nobody reads. A gate that can be skipped will be skipped, quietly, by the person under deadline pressure. And as AI drafting spreads from a pilot on one desk to routine use across business units, the governance question shifts from whether it is allowed to who is accountable for the volume each unit generates, which is the same problem as attributing AI usage back to the unit that owns it.

Firms tend to treat AI-assisted messaging as a tooling upgrade and reach for supervision and recordkeeping later, if at all. The regulator’s position is the reverse: the obligations were already there, and the firm chose to run a communication channel that does not capture what it produces. The model did not break the supervisory system. Deploying it without wiring it in did.