FINRA Email Retention and AI Drafting: What Firms Must Do

The question behind FINRA email retention and AI drafting is usually asked backwards. Firms ask which special rule covers a machine-written email. There isn't one. FINRA said so in Regulatory Notice 24-09: its rules "are intended to be technology neutral" and continue to apply when member firms use generative AI, and Rule 2210's content standards apply "whether member firms' communications are generated by a human or technology tool."

So classification decides your obligations, and classification turns on audience and content, not authorship. That is good news and a trap at once, because AI changes how easily one paragraph reaches forty people.

This is regulatory orientation with links to the rule text, not compliance advice. Your firm's written supervisory procedures govern, and your CCO makes the call.

The short answer#

A FINRA-regulated firm can use AI to draft email. What must happen around it comes from four places, and none of them are new.

• Supervision — FINRA Rule 3110(b)(4). Correspondence is reviewed under the firm's procedures, and that review "must be evidenced in writing, either electronically or on paper."
• Communications standards — FINRA Rule 2210, which applies regardless of who or what wrote it. If the message is a retail communication, an appropriately qualified registered principal must approve it "before the earlier of its use or filing with FINRA's Advertising Regulation Department."
• Records of the communication — SEA Rule 17a-4(b)(4) for the sent and received copies, plus FINRA Rule 2210(b)(4) for the retail-communication file.
• The catch-all — FINRA Rule 4511, requiring members to "make and preserve books and records as required under the FINRA rules, the Exchange Act and the applicable Exchange Act rules," in "a format and media that complies with SEA Rule 17a-4."

Why the rules do not care that a machine wrote it#

Nothing in Rule 2210 conditions an obligation on authorship. The definitions turn on distribution: correspondence is "any written (including electronic) communication that is distributed or made available to 25 or fewer retail investors within any 30 calendar-day period," and a retail communication is the same thing distributed to more than 25.

Rule 3110's review obligation is equally indifferent — a draft from a model, a paralegal or a template library arrives at the same supervisory gate.

Where AI does raise a distinct obligation is when the firm points it at supervision itself. Notice 24-09 is explicit: if a firm uses these tools "as part of its supervisory system — for the review of electronic correspondence, for instance — its policies and procedures should address technology governance, including model risk management, data privacy and integrity, reliability and accuracy of the AI model." The same expectation applies "whether member firms are directly developing Gen AI tools for their proprietary use or when leveraging the technology of a third party."

That is the real division. AI writing a draft a person then approves is an existing workflow with a faster first step. AI deciding what gets reviewed is a new supervisory system, and it needs governance.

The line that actually decides it: 25 retail investors in 30 days#

This is the part most guidance skips, and it is the part AI genuinely changes. Classification depends on how many retail investors receive the communication in a rolling 30-calendar-day window. Twenty-five or fewer, it is correspondence, supervised and reviewed under Rule 3110. Twenty-six or more, it is a retail communication, and a registered principal must approve it before first use.

A human writing individually to clients crosses that line slowly, because forty tailored emails are expensive. A drafting assistant makes the second, tenth and fortieth nearly free. Volume is the variable the technology moves, and volume is what the definition keys on.

So a firm can manufacture a retail communication without anyone deciding to create one. Nobody wrote an ad. Somebody reused a good paragraph, and the count did the rest.

What must be retained, and for how long#

Two details in that last column matter more than they look. Rule 3110.07 states that "merely opening a communication is not sufficient review" — so an approval click that writes nothing fails the test even though a person did, in fact, look.

And since the May 3, 2023 compliance date for the SEC's 2022 amendments, electronic records no longer have to sit on write-once-read-many media. Rule 17a-4(f)(2)(i) offers a choice: maintain "a complete time-stamped audit trail" of every modification and deletion, or "preserve the records exclusively in a non-rewriteable, non-erasable format." The audit-trail alternative fits a draft-then-approve workflow, because that workflow is a sequence of changes with names and timestamps on them.

Criteria that actually matter when choosing a workflow#

Feature lists do not decide this. Six questions do, worth asking of any tool before it touches client mail.

• Does the sent copy land in the channel the firm already captures? If mail leaves through the firm's own connected account, the archive sees it. Anything else is an off-channel problem.
• Can the tool be prevented from sending without a person? Approval-before-send has to be a setting the firm controls, not a habit the user maintains.
• Does the approval write a retrievable artifact naming the reviewer, the item, the date and the action? That is the Rule 3110.07 test, and most drafting tools fail it silently.
• Does anything count recipients per communication across the firm? Almost nothing does this automatically, so it becomes a written procedure someone owns.
• What does the vendor do with the mail? Training use, storage location and record production are governance questions under Notice 24-09, not procurement niceties.
• Does adding the tool change how the archive captures anything? Rewritten headers or threads, or sending via the vendor's own infrastructure, can break capture unnoticed until an exam.

Five workflows, scored against those criteria#

Only the first row keeps every obligation on artifacts the firm already retains. Each of the others moves one piece outside the perimeter — the draft, the reviewer's name, the recipient count, the send itself — and an examiner can ask for any of them.

A worked example: one paragraph, two classifications#

1. 1

   Monday — the draft

   An advisor asks an assistant to explain a bond-fund reallocation in plain language, edits it, and sends it to 14 clients. Fourteen retail investors in the window. This is correspondence: supervised and reviewed under Rule 3110(b), with the sent copies preserved under SEA Rule 17a-4(b)(4).

2. 2

   Wednesday — the reuse

   A colleague likes the paragraph and sends it to 19 of their own clients. Same firm, same 30-day window, 33 retail investors. It has just become a retail communication.

3. 3

   The obligation that appeared without anyone creating it

   Rule 2210(b)(1)(A) required a registered principal to approve it before the earlier of its use or filing — due before Wednesday's send. Nobody sought one, because to both advisors this was still a client email.

4. 4

   What the file now has to contain

   Under Rule 2210(b)(4): a copy, the dates of first and last use, the name of the approving principal, the date of approval, and the source of any yield figure quoted.

5. 5

   The part that has nothing to do with AI

   All of this would be true had a human written the paragraph from scratch. The model did not create the obligation — it removed the friction that used to keep reuse below the threshold.

6. 6

   The control that would have caught it

   A rolling recipient count per reusable communication, checked before the second send rather than at the next exam. That is a supervisory procedure, and no drafting tool maintains it for you.

Red flags#

• Any configuration that can send client-facing mail without a person approving it, deployed before the written supervisory procedure covering it exists.
• Approval that leaves no artifact. If the record cannot name the reviewer, the item, the date and the action taken, Rule 3110.07 is not satisfied by the fact that someone looked.
• "Six years" used as the retention answer for client email — that is Rule 4511(b)'s fallback, not email's period.
• Drafting in a consumer chat tool and pasting into firm email. The sent copy is captured; the client data that reached an unassessed vendor is not.
• A vendor that will not state in writing whether customer mail is used to train models, or cannot describe how records are produced to regulators on request.
• Template reuse with no recipient counting anywhere in the firm — the likeliest way a retail communication goes out unapproved.
• An AI summary standing in for the supervisory review itself, without the model risk, data integrity and accuracy governance Notice 24-09 expects.

What we'd pick, and who should pick something else#

On the criteria above, the workflow that holds up is the boring one: draft inside the mail account the firm already journals, require human approval before every send, and make that approval produce a record. Every obligation in this post then lands on an artifact the firm already retains — and the audit-trail alternative in Rule 17a-4(f) is built for exactly that kind of change history.

AI Emaily is built in that shape, and we build it — so weigh this accordingly and check the mechanics in our docs rather than in this paragraph. Copilot mode holds every send for human approval, agent actions are written to an audit log you can query, and mail leaves through your own connected Gmail, Outlook or IMAP account, so the sent copy reaches the archive by the path it always did. Drafting voice comes from a Personal Context brain you write and per-client profiles you set, not from training on your mailbox — which is the answer a vendor questionnaire actually asks for.

Now the part that decides it for many readers: we are not an archive. AI Emaily does not perform lexicon surveillance or e-discovery, does not file anything with FINRA's Advertising Regulation Department, and does not classify your communications or count recipients against the 25-investor threshold. Those are supervisory controls your firm owns, in the system of record it already licences. If the archive itself is what you are shopping for, buy a dedicated archiving and supervision platform and judge it on capture completeness, retention configuration and production workflow. We sit upstream of that, not in place of it.

And if you are a registered rep on a firm-managed tenant, the choice may not be yours at all. Where IT controls which applications may connect and compliance keeps an approved-vendor list, the honest first step is asking them, not signing up.