A registered representative gives an AI assistant a few notes and asks it to draft a client email. The result is polished. It explains a product, mentions a fee, and closes with a confident recommendation. It also adds a statistic that was never in the notes and cites a source that does not exist.
That last part may be easy to miss. Fluency makes an answer feel finished, especially when the reviewer already knows the subject and reads quickly. NIST warns that generative AI can produce false content with confident wording and can invent logic or citations that make an answer look better supported than it is.[3]
For a small broker-dealer, the useful question is not whether AI may help with a draft. It is what has to happen between the draft and the client.
FINRA says its rules still apply when member firms use generative AI, and the content standards for communications with the public apply whether a person or a technology tool generated the communication.[1] Its guidance does not turn every AI draft into the same type of communication or prescribe one approval path. The firm still has to classify and supervise the communication under the rules and procedures that apply to its business.[4]
This article is an awareness and workflow aid, not legal advice or a determination that a message meets FINRA, SEC, state, or firm requirements. Compliance should set the review and retention rules. Employees need a simple way to follow them.
Keep the first draft in a review lane
An AI-generated email should begin in a place where it cannot reach a client by accident. That might be a draft folder, an approved workspace, or a ticket that routes the text to a reviewer. The exact tool matters less than the boundary: generation and delivery should be separate actions.
Do not give an experimental assistant permission to send email merely because it can draft one. Do not let a representative paste text from a personal AI account into the client system without following the firm’s approved-use and data-handling rules. A fast drafting shortcut should not bypass supervision, books-and-records procedures, privacy controls, or the ordinary communication channel.
Label the draft while it is under review. A short line such as “AI-assisted draft; not approved for delivery” removes ambiguity when the text is copied into a shared queue. The label is internal handling information, not a substitute for review and not necessarily something that belongs in the final message.
Read once for invented facts
The first review should ignore style. Check every factual statement against material the firm already trusts.
Mark names, dates, prices, fees, rates, product features, performance figures, eligibility terms, account details, tax statements, regulatory references, and quoted language. For each marked item, the reviewer should be able to point to the source used for that exact claim. If the source is missing, stale, or unclear, remove the claim or send it to the appropriate subject-matter reviewer.
Never trust a citation because it looks plausible. Open the source. Confirm that it exists, says what the email claims, applies to the right product and time period, and has not been replaced. NIST specifically notes that a model may generate false citations that encourage people to trust an incorrect answer.[3]
A useful source note is brief: document name, version or date, and the section supporting the statement. Do not paste a client record or an entire proprietary document into an unapproved tool just to make verification easier.
FINRA’s 2026 oversight report describes hallucinations as inaccurate or misleading information presented as fact. It also notes that incorrect interpretations of rules, policies, client data, or market data can affect decisions.[2] That is why “it sounds right” cannot be an acceptance test.
Read again for missing context
A draft can contain individually accurate sentences and still mislead by leaving something out.
Ask what a client could reasonably take away from the message. Does a description of a benefit omit a material condition or risk? Is a fee described without the circumstances in which it applies? Does a comparison use the same period, basis, and assumptions? Has the draft turned general education into a personal recommendation? Does the subject line promise more than the body supports?
This pass needs a person who understands the firm’s products and communication standards. Grammar software may catch a typo. It will not reliably notice that the email describes liquidity but omits a restriction that changes the client’s decision.
Keep the review tied to the actual message. A long AI policy will not help someone decide whether one sentence needs a qualification. Give reviewers examples drawn from the firm’s approved communications, along with a clear route to compliance when the answer is uncertain.
Check the prompt and the output
Reviewing only the final prose leaves half the activity unexplained. The firm should know what information went into the tool, which account and model were used, and whether the output entered another system.
The record does not need to become a warehouse of sensitive prompts. Compliance, privacy, legal, and records staff should decide what must be retained and what should not be collected. Employees should know how to report an accidental disclosure, an unexpected connector, or a draft that included another client’s information.
The 2026 FINRA report lists prompt and output logs, model-version tracking, validation, and human review among practices firms may use to monitor generative AI.[2] A small firm can translate that into a modest record: the business purpose, approved tool, drafter, reviewer, disposition, and a reference to the source material. Follow the firm’s formal requirements rather than creating an unofficial shadow archive.
Make approval visible
“Someone looked at it” is not a control. The workflow should show who reviewed the draft, what they reviewed it for, and whether the message was approved, returned, or rejected.
The required reviewer may change with the communication’s content, audience, and distribution. Staff should not have to interpret that routing from memory. Put the decision into the process:
- The drafter identifies the intended recipient and purpose.
- The workflow routes the message under the firm’s communication procedures.
- The reviewer checks facts, sources, material context, and required wording.
- Only the approved version moves to the firm’s delivery channel.
- Material edits after approval return to review.
If an AI feature can rewrite text inside the delivery system, decide whether a last-second rewrite invalidates the approval. Otherwise, the version that reaches the client may not be the version anyone reviewed.
Train with a deliberately flawed message
People learn this faster by marking up one email than by hearing a lecture about hallucinations.
Create a fictional client message with a handful of planted problems: one invented statistic, one real source that does not support the claim, an outdated fee, a missing qualification, and an unsupported recommendation. Use fictional names and accounts. Ask the team to review the email using the firm’s real routing process, then compare what each person caught.
Finish by testing the handoff. Can the employee find the approved source? Do they know who handles a questionable product claim? Can they report that protected information went into the wrong tool? Does the approved copy stay fixed after review?
Record the process gaps, not a score that embarrasses the person who missed a sentence. The goal is to improve the review lane: clearer source locations, better routing, tighter tool permissions, and a reliable stop before delivery.
AI can save time on a first draft. The firm remains responsible for the communication that leaves its systems. A visible review lane gives employees room to use the tool without confusing polished language with verified work.
For practical guidance on cyber risk and safer AI use, subscribe to Cyber & AI Insights.
If your firm wants a practical awareness session built around its approved AI and communication workflows, schedule a conversation with Your Data Guardian.
Sources
[1] https://www.finra.org/rules-guidance/notices/24-09 — FINRA Regulatory Notice 24-09
> “For example, FINRA has provided guidance that the content standards of Rule 2210 (Communications with the Public) apply whether member firms’ communications are generated by a human or technology tool, and that guidance discusses the specific application of the rules to certain AI-generated communications.”
[2] https://www.finra.org/rules-guidance/guidance/reports/2026-finra-annual-regulatory-oversight-report/gen-ai — FINRA 2026 Oversight Report: GenAI
> “Hallucinations refer to instances where the model generates information that is inaccurate or misleading, yet is presented as factual information.”
> “This may include storing prompt and output logs for accountability and troubleshooting; tracking which model version was used and when; and validation and human-in-the-loop review of model outputs, including performing regular checks for errors or bias.”
[3] https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf — NIST AI 600-1: Generative AI Profile
> “GAI outputs may also include confabulated logic or citations that purport to justify or explain the system’s answer, which may further mislead humans into inappropriately trusting the system’s output.”
[4] https://www.finra.org/rules-guidance/guidance/faqs/advertising-regulation — FINRA Rule 2210 Frequently Asked Questions
> “Depending on the nature and number of persons receiving the chatbot communications, they may be subject to FINRA communications rules as correspondence, retail communications, or institutional communications. Therefore, the firm must supervise the chatbot communications in accordance with applicable FINRA rules.”