Flagging AI-Drafted Communications to Senior and Vulnerable Investors Under FINRA Rules 2165 and 4512
Neither Rule 2165 nor Rule 4512 directly regulates AI-drafted content. Firms apply their existing supervision, communications and customer-protection obligations to AI-assisted drafting. In practice that means joining account-level data (age, trusted contact status, exploitation concerns, suspected impairment) to the drafting workflow, defining routes for flagged and unknown-status customers, requiring human review where the firm decides it is needed, and retaining per-communication evidence of what the control did.
Routing and escalation steps
- 1
1. Query flag state
At each drafting request, retrieve age band, trusted contact status, exploitation-concern indicator and impairment indicator, with timestamps.
- 2
2. Classify
Assign unflagged, flagged (objective), flagged (judgment-based) or unknown. Define the unknown-status default in procedures, since it is a firm decision.
- 3
3. Apply the route
Allow with standard supervision, redact specified content, block, or escalate to a designated reviewer before sending.
- 4
4. Review
The reviewer approves, edits or rejects the draft. The decision and the reviewer's identity are recorded against the communication.
- 5
5. Escalate concerns
If a draft or the customer's response reveals possible exploitation, hand off to the firm's Rule 2165 hold-decision process, which stays a human decision.
- 6
6. Deliver and retain
Send only approved content and retain the communication, and for chatbots the chat session.
Evidence to retain for each communication
- Flag inputs and their timestamps, including who set any judgment-based flag and why
- Policy or criteria version in force at the time
- Action taken: allow, redact, block or escalate
- Reviewer identity, decision and any edits
- Final content sent, plus the chat session for chatbot interactions
- Training records and internal review records for any Rule 2165 hold decision
What the rules require, and what they do not
Rule 4512 concerns customer account information. Under Rule 4512(a)(1)(F), a firm must make a reasonable effort to obtain the name and contact information of a trusted contact person, age 18 or older, for each non-institutional customer account. FINRA describes the requirement as not limited to senior investors. FINRA also does not expect a member to verify the age of a designated trusted contact.
Rule 2165 concerns financial exploitation of Specified Adults. It permits, but does not require, a temporary hold on a securities transaction or a disbursement when the firm reasonably believes exploitation has occurred, is occurring, has been attempted or will be attempted. A Specified Adult is a natural person age 65 or older, or a person age 18 or older whom the firm reasonably believes has a mental or physical impairment that leaves them unable to protect their own interests. After placing a hold, the firm must notify all parties authorized to transact on the account and, if available, the trusted contact person, no later than two business days after the hold.
Neither rule sets requirements for how communications are drafted. Linking these data to AI drafting is a control the firm designs, not a rule requirement. The obligations that reach AI-drafted content come from elsewhere. FINRA states that its rules are technology neutral and apply to generative AI like any other technology. Rule 3110 requires a reasonably designed supervisory system tailored to the firm's business, and the content standards of Rule 2210 apply whether a human or a tool generates the communication. FINRA's 2026 oversight report adds that firms should consider how they will comply with FINRA rules before testing and deploying GenAI tools. As of November 2026, that is the framework a flagging workflow has to fit.
Data inputs and flag logic
A flagging workflow starts with a service the drafting workflow can query at the time of each request. Four inputs are enough to begin: customer date of birth or age band, the trusted contact on file (or a record of the reasonable effort to obtain one), an internal exploitation-concern or temporary-hold indicator, and a reviewer-set indicator of suspected impairment.
The flag logic should keep two kinds of criteria apart. Age 65 or older is an objective field and can be computed. Suspected impairment and exploitation concerns depend on a firm's reasonable belief. They are not computed attributes, so they should be set by a person, with the identity of the person and the reason recorded. FINRA's examination findings on Rule 2165 have included no documented internal review, which makes that record more than good hygiene.
The sources support only the Specified Adult definition as a defined category. Any broader definition of a vulnerable investor is a firm choice, and written procedures should label it that way and name who approves changes to it. Trusted contact status is worth carrying as an input, but a missing trusted contact is a data-completeness fact, not evidence of vulnerability. Treat it as a separate field so it does not silently widen or narrow the flagged population.
Governance and tradeoffs
Routing every flagged communication to a human reviewer is the most conservative design, but it creates volume and can slow servicing. Narrower routing, such as review only for specified content types, reduces load but needs a written rationale. Either way, the supervisory procedures should be reasonably designed for the firm's business, and the model risk process should evaluate the workflow before testing and deployment.
Flag-data quality is itself a control risk. Stale age data or missing trusted contact records can misroute a communication, so procedures should state how unknown states are handled and who monitors them. The inventory should also cover embedded GenAI features in third-party products, because FINRA applies its rules whether a firm builds a tool or uses another party's.
Document training and internal review. FINRA's findings for firms relying on Rule 2165 listed no documented training and no documented internal review, the latter meaning no retained record behind a decision to place a hold. Finally, monitor pending changes. FINRA has a proposed rule filing noticed in September 2026 that describes the Rule 4512 and 2165 framework, and Regulatory Notice 26-02 proposes a Rule 2166, a temporary delay of up to five business days on disbursements or transactions where fraud is reasonably believed, for all customers regardless of age or capacity. Neither is shown as adopted as of November 2026. Build flag criteria as versioned configuration so procedures can change without rebuilding the workflow.
Where runtime policy enforcement fits
The workflow above can be built in several ways. One option is to enforce policy in the execution path, at the level of individual interactions, rather than reviewing drafts after the fact. Trussed enforces policy in the execution path with actions to allow, redact, block or escalate for human approval, and it generates audit-ready evidence from that enforcement. Where the flag state is available to the policy layer at request time, those actions map directly to the routes described above, and the evidence records what was enforced for each interaction.
This matters for examiner scrutiny and model risk management because a written procedure and a record of what actually happened are different things. Trussed's depth in regulated financial sectors is in insurance, and it runs in production for a regulated enterprise, with 18+ months of continuous operation. A runtime layer does not decide whether a customer is impaired or whether to place a Rule 2165 hold. Those remain firm judgments, supported by the records above.
Flagging workflow for AI-drafted client communications
Data inputs
Age, trusted contact status, exploitation-concern records and reviewer-set impairment indicators.
Flag logic
Objective criteria kept separate from judgment-based criteria, with the owner and reason recorded.
Routing
Allow, redact, block or escalate for human approval, by flag state.
Evidence
Flag state, policy version, action, reviewer decision and final content tied to each communication.
Review your flagging and evidence design
Talk through how runtime enforcement and per-interaction evidence could support your supervision of AI-drafted client communications.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.