From Sidhant Tamrkar | Product & Market Analysis
FINRA Generative AI Guidance: 8 of 16 Asks Bind. Here Is the Vendor Evidence.
On this page
FINRA has written no AI rule, and half of what its documents ask about generative AI is not binding. We read 6 FINRA documents and sorted their asks by the verb FINRA used. Of 16 asks, 8 bind member firms today. FINRA generative AI guidance supplies the other 8, and your buyer will still want evidence for most of them.
Key takeaways
- FINRA regulates AI through existing rules. Regulatory Notice 24-09 says its rules are technology neutral and apply whether a member firm builds a GenAI tool or buys one, including features embedded in existing products.
- Only 8 of 16 FINRA asks use binding language. Our count covers 6 FINRA documents. The remainder comprises 2 should statements, 5 optional practices and 1 open question.
- Prompt logs are not a stated FINRA recordkeeping duty. The 2026 oversight report lists them under practices firms may want to consider, and its Books and Records topic never mentions AI.
- Responsibility stays with the firm, so vendors sell evidence. Notice 21-29 says outsourcing does not relieve a member of its ultimate compliance responsibility, which turns every row of the map into a procurement question.
What does FINRA require for generative AI? It requires no AI-specific control. Rule 3110 supervision, Rule 2210 communications, SEC Rule 17a-4 recordkeeping and Regulation Best Interest apply to GenAI as they apply to any other tool. Firms stay responsible when a vendor builds the tool, so vendors must supply the evidence.
FINRA generative AI obligations start from technology neutrality
FINRA's position fits in 1 line. Regulatory Notice 24-09, published 27 June 2024, says its rules are "intended to be technology neutral" and keep applying when firms use GenAI. The same notice says it "does not create new legal or regulatory requirements".
The notice also closes the exit buyers might hope for. FINRA rules apply to proprietary tools and to third-party technology, "including through embedded features in existing third-party products". An AI feature switched on inside software the firm already licenses is therefore in scope, by our reading, even if nobody ran a new procurement.
Regulatory Notice 21-29, from August 2021, sets the vendor frame. Quoting earlier FINRA guidance, it says outsourcing "does not relieve members of their ultimate responsibility for compliance with all applicable federal securities laws" and FINRA rules.
That has a blunt consequence for your sales deck. FINRA rules bind member firms, not their software suppliers. We would strike "FINRA compliant" from any vendor page, because the product cannot hold that status. What your product can do is make the firm's compliance cheaper to prove.
Formal AI rulemaking has not started either. Walter Saurack of Duane Morris wrote on 22 July 2026 that FINRA "has not started formal rulemaking specific to AI," while expecting further guidance. The nearest thing is a communications proposal, covered below, that names GenAI as 1 of its drivers.
The FINRA AI guidance map: 16 asks and the vendor evidence behind each
We built the table from 6 FINRA documents: Notices 21-29, 24-09, 25-07 and 26-14, the advertising FAQ, and the 2026 oversight report. Each ask is sorted by FINRA's own verb. Must, or a rule that applies, counts as binding. Should stays should. "May want to consider" and listed effective practices count as optional. A question counts as open.
The working is simple: 8 binding + 2 should + 5 optional + 1 open = 16, so 50% of the asks bind. The vendor evidence column is our inference, not FINRA text. No FINRA document we read prescribes a vendor artefact.
| # | Source | What it asks of the firm | Status | Vendor evidence it implies |
|---|---|---|---|---|
| 1 | Rule 3110; Notice 24-09 | A "reasonably designed supervisory system tailored to its business". | Binding | Control description the firm can adapt for its written procedures: who reviews what, and what escalates. |
| 2 | Rule 2210(a), (b); FAQ B.4 | Treat chatbot output as correspondence, retail or institutional, based on nature and number of recipients. | Binding | Recipient count and audience metadata on every output. |
| 3 | Rule 3110(b)(4), 3110.06 through .09; FAQ B.4 | Supervise chatbot communications. | Binding | Review queue, sampling settings and reviewer identity export. |
| 4 | Rules 2210, 2220; FAQ D.8 | Keep AI output fair and balanced, with no false, misleading, promissory or exaggerated claims. | Binding | Test results for those failure types, by model version. |
| 5 | Rule 2210(b)(1); Notice 26-14 | Principal approves each retail communication before use. | Binding, change proposed | Hold-and-release gate recording principal and time, configurable by category. |
| 6 | SEC Rule 17a-4(b)(4); Rules 2210(b)(4), 3110.09, 4511(a); FAQ D.8 | Preserve business communications, including AI-generated ones. | Binding | Export of the exact delivered text to the firm's archive. |
| 7 | SEC Rule 17a-4(f)(2) | Electronic records need a time-stamped audit trail or a non-rewriteable format. | Binding | Written statement of the path used, plus a demonstration of edit history. |
| 8 | SEC Rule 15l-1; FAQ D.8 footnote | Regulation Best Interest applies if AI recommends a securities transaction or strategy to a retail customer. | Binding | Recommendation flag, plus a way to block or route that output to a person. |
| 9 | Notice 24-09 | Evaluate GenAI tools before deployment. | Should | Pre-deployment pack: model, intended use, known limits, test results. |
| 10 | Notice 24-09 | If GenAI is part of supervision, address model risk, data privacy and accuracy. | Should | Miss rates for surveillance use, model change log, data flow diagram. |
| 11 | Notice 21-29 | Due diligence on continuity, data protection, subcontractors and audit reports. | Optional | SOC 2 Type II report, continuity plan, subprocessor list with controls. |
| 12 | Notice 21-29; 2026 report | Contract for regulator access to records, data disposition and cyber event notice, and bar sensitive data from open-source GenAI tools. | Optional | Contract rider, plus technical proof of where data flows. |
| 13 | 2026 report, third-party topic | Assess vendor GenAI use; inventory components, versions, data types and fourth parties; review default settings. | Optional | Machine-readable inventory of models, sub-vendors and default settings. |
| 14 | 2026 report, GenAI topic | Monitor prompts, responses and outputs; store logs; track model version; human review. | Optional | Exportable logs carrying model version, with retention controls. |
| 15 | 2026 report, GenAI topic | For agents, consider system access, human oversight, action tracking and guardrails. | Optional | Scoped permissions list, approval gates, action trace, tested limits. |
| 16 | Notice 25-07 E.3; Notice 26-14 item 8.1 | FINRA asks how firms handle records for chatbots, transcripts, summaries and automatically generated communications. | Open | Retention settings per artefact type, so the firm can apply its own decision. |
Rule 2210 applies to AI-generated communications regardless of author
The clearest FINRA language on GenAI sits in its advertising FAQ, posted 10 May 2024. Question D.8 asks whether a firm is responsible for content created using AI. The answer starts with "Yes" and says responsibility holds "regardless of whether they are generated by a human or AI technology".
Classification turns on who received the output
FAQ B.4 says chatbot communications may be correspondence, retail or institutional, "depending on the nature and number of persons receiving" them. Each category carries different review duties. So the supervision question depends on a fact many AI products do not record per output, which is how many people saw it.
We think this is the most underbuilt field in financial services AI. A chat log tells the firm what was said. It does not tell the firm which FINRA category the message fell into, and the firm needs that before it can choose a review path.
Content standards need tests, not an accuracy score
D.8 says communications must be fair and balanced and bars "false, misleading, promissory or exaggerated statements or claims". A headline accuracy figure answers none of those. An answer can be factually correct, omit the risk and still fail fair balance. Ask your evaluation team for a separate test set on each failure type, reported by model version.
A footnote to D.8 adds a line that vendors skim past. If AI recommends a securities transaction or strategy to a retail customer, Regulation Best Interest applies. Any product that can produce a recommendation needs a way to detect that output, then block it or route it to a person.
Broker-dealer AI supervision records: what is settled and what is not
Recordkeeping is where commentary has run furthest from the text. Most of the settled ground is old. The unsettled ground involves exactly the outputs firms now generate most.
The retention rule predates GenAI
SEC Rule 17a-4(b)(4) covers communications sent and received, including inter-office ones, "relating to its business as such". Those records must be kept for 3 years, the first 2 easily accessible. Notice 25-07 concedes that the phrase "business as such" is "not defined". Rule 17a-4(f)(2) then offers 2 storage paths for electronic records: a complete time-stamped audit trail, or a non-rewriteable, non-erasable format.
The prompt log sentence is softer than reported
The 2026 report's GenAI topic introduces its list with "may want to consider". Monitoring, it says, "may include storing prompt and output logs for accountability and troubleshooting". Those are 2 layers of optional wording, and the stated purpose is accountability and troubleshooting, not retention.
Archive Intel, which sells archiving, argued in September 2026 that the passage is governance guidance rather than a recordkeeping requirement. We agree on the text, and an archiving vendor has little commercial reason to argue for less archiving. But the text does not settle the law. Saurack's post says neither the SEC nor FINRA "has stated yet whether a prompt or its output qualifies".
Our recommendation is to log everything, then let the firm set retention. Designing the fields is its own job, and our 12-field minimum audit record for AI agents is a starting template.
Meeting summaries remain a live question
The 2026 report says the top GenAI use case among members is "Summarization and Information Extraction". Notice 25-07, from April 2025, asked firms what recordkeeping challenges AI creates, naming "AI-generated transcripts or summaries of meetings". We found no FINRA document answering that question. So the most common category of use overlaps with the least settled record type. Note-takers also carry separate exposure, covered in the analysis of consent and training rights on recordings.
The Rule 2210 proposal would change the AI review queue, not the liability
Regulatory Notice 26-14, published 9 July 2026, is the closest FINRA has come to rewriting a rule with GenAI in mind. It says pre-use approval "presents significant challenges" for some communications, "particularly communications on social media and some generated by AI". Comments closed on 11 September 2026. The current rule stays in force until any change is approved and takes effect.
| Item | Today | As proposed |
|---|---|---|
| Principal approval before use | Required for each retail communication. | Written procedures decide which categories need it. |
| Retail research reports | Pre-use approval. | Pre-use approval kept. |
| Basis for the decision | None needed, since all are approved. | Non-exhaustive list of 8 factors in Supplementary Material .01 to the rule. |
| Where pre-use review is not universal | Does not arise. | Training, documentation of training, surveillance and follow-up. |
| Responsibility for AI output | The member. | The member. |
The factors translate directly into product metadata. They include whether a communication makes a recommendation, promotes a third-party product, appears tailored to a specific audience, or includes performance data, rankings or comparisons. A vendor that labels every draft with those attributes hands the firm its routing logic.
FINRA is explicit that tools can sit inside that system "provided they are vetted, tested and monitored". It also suggests firms consider the qualifications of the people reviewing the AI tool. We would not sell the proposal as approval relief yet. Build the labels now because they help under either version, and keep the hold-and-release gate.
The notice itself is loose on 1 point. It says procedures "should include" training and surveillance in 1 passage, and "must include" them in its economic impact section. Treat them as required when you scope the product.
FINRA vendor due diligence for AI: the questions you will receive
Notice 21-29 and the 2026 report's third-party risk topic are the closest thing to a FINRA vendor questionnaire. Neither is phrased as a mandate. Both are likely to shape what a firm's vendor risk team asks you, which is why we put them in the table.
What due diligence covers
Notice 21-29 suggests firms ask about vendor continuity plans, data protection and controls over subcontractors. It suggests they consider obtaining SSAE 18 Type II or SOC 2 reports, and asks whether they document their findings. Our breakdown of the artefacts SOC 2 auditors now request for AI covers most of the overlap, so 1 evidence set can serve both.
Where the 2026 report names GenAI
The 2026 report lists "assessing the third-party vendor's use of GenAI" as due diligence. Its contract example is specific: language prohibiting firm or customer sensitive information "from being ingested into a third-party vendor's open-source GenAI tool". That reaches your own model stack, so be ready to disclose which models sit behind the feature and who hosts them.
The inventory practice is the sleeper. FINRA suggests firms list vendor services and components "including the version," the data types vendors access, and any fourth parties handling firm data. For an AI product, the model provider is typically that fourth party. A silent model swap would then make the firm's inventory wrong without anyone at the firm doing anything.
| Question | Weak answer | Answer that works |
|---|---|---|
| Are you FINRA compliant? | Yes, we are FINRA compliant. | FINRA rules bind you. Here is our map of which control supports which rule. |
| Is it accurate? | A single accuracy percentage. | Results by failure type and model version, with test set size and date. |
| Can we retain everything? | We keep logs. | What is logged, the export format, the retention setting and who can delete. |
| Which models do you use? | Leading frontier models. | Named models and hosts, with notice terms for any change. |
FINRA AI agent guidance is short, optional and demanding to meet
The 2026 report defines AI agents as systems "capable of autonomously performing and completing tasks on behalf of a user". It names 6 agent risks: autonomy, scope and authority, auditability, data sensitivity, domain knowledge, and rewards and reinforcement. It then says the usual GenAI risks of bias, hallucinations and privacy still apply.
Its 4 considerations cover how to monitor system access and data handling, where to place human oversight, how to track actions and decisions, and how to set guardrails. FINRA says agents "may call for supervisory processes that are specific to the type and scope of the AI agent". That points to a separate control set per agent, not 1 policy for all of them.
Lawyers are already reading beyond the wording. Snell & Wilmer wrote in December 2025 that agent logs "should be treated as regulatory records subject to Rule 17a-4 retention requirements". That is the firm's recommendation, not FINRA's. It is still a sensible design target if you sell agents to broker-dealers.
Oversight placement deserves more care than a blanket approval step. Reviewers who approve nearly every prompt provide little supervision, a failure examined in the piece on designing the human loop before buying the agent. Show the firm where a person intervenes, and what they check when they do.
Where this map is weakest
The table looks more certain than it is. 2 problems deserve stating before you take it into a procurement meeting.
The 8 of 16 count is our judgement
FINRA publishes no such list. Splitting row 2 from row 3, or merging the agent and monitoring rows, would move the ratio. Rows 6, 7 and 8 rest largely on SEC rules that FINRA points to, not FINRA's own text. We also read 6 documents and not every speech, podcast or exam finding. Use the count to see the shape, not as a compliance score. None of this is legal advice.
Optional wording does not mean optional in an exam
This is the strongest objection to our framing. Rule 3110 requires a reasonably designed system, and a firm may struggle to defend that design without the practices FINRA listed. A firm that skipped logging because it was only "may want to consider" could still have a supervision problem. So treat the optional band as a list of likely exam questions.
The enforcement record cannot settle this yet. Our searches found no FINRA disciplinary action centred on GenAI supervision, but we did not run a full check of FINRA's disciplinary database. Absence from search results is weak evidence, and an unreported case would change how the optional band reads.
Frequently asked questions
Does FINRA have rules for generative AI?
FINRA has no AI-specific rule. Regulatory Notice 24-09 says its rules are technology neutral and continue to apply when member firms use generative AI, whether they build a tool or buy it from a third party. The main rules involved are Rule 3110 on supervision and Rule 2210 on communications, along with recordkeeping requirements. A July 2026 proposal, Notice 26-14, would change how retail communications are reviewed.
Is a firm responsible for AI-generated communications under FINRA rules?
Yes. FINRA advertising FAQ D.8 says firms are responsible for their communications whether a human or AI technology generated them. They must meet supervision requirements, applicable FINRA and SEC recordkeeping requirements, and the content standards in Rules 2210 and 2220. If AI recommends a securities transaction or investment strategy to a retail customer, a footnote to the FAQ says Regulation Best Interest also applies.
Does FINRA require firms to store AI prompts and outputs?
FINRA has not stated a prompt logging requirement in the documents we read. Its 2026 oversight report says monitoring may include storing prompt and output logs, inside a list of practices firms may want to consider. Delivered communications are a different matter, because existing retention rules cover them. Whether a prompt itself is a required record remains unresolved, so most firms should capture logs and decide retention deliberately.
What does FINRA expect from AI vendors?
FINRA rules bind member firms, not their software vendors, so its expectations reach vendors through the firm. Notice 21-29 suggests due diligence on continuity, data protection, subcontractors and audit reports. The 2026 report adds assessing a vendor's GenAI use and contract language keeping sensitive data out of open-source GenAI tools. Vendors should prepare documented evidence for each item, because the firm keeps ultimate responsibility.
Do FINRA rules apply to AI agents?
Yes, because existing rules are technology neutral, though FINRA has issued no agent-specific rule. The 2026 oversight report names risks such as autonomy, scope and authority, and auditability. It suggests firms consider how to monitor system access and data handling, where to place human oversight, how to track agent actions and decisions, and how to set guardrails. Those considerations are phrased as optional, yet Rule 3110 still applies.
How to use this map this week
Copy the 16 rows into a spreadsheet and add 1 column: the artefact you could send today. Mark each row as exists, exists but cannot be exported, or missing. Start with rows 2, 6 and 14, because they depend on data that is impossible to recreate after the event.
Then find every sentence in your sales material that says "FINRA compliant" and replace it with a link to that spreadsheet's output. A compliance reviewer will trust a mapped gap more than an unqualified claim. It also gives your buyer something to forward.
Use this
Attach the completed 16-row map to your next broker-dealer security questionnaire, and date it. FINRA's position is still moving, so the date is part of the evidence.
References
- FINRA, Regulatory Notice 24-09, 27 June 2024. Used for technology neutrality, third-party scope and pre-deployment evaluation.
- FINRA, 2026 Annual Regulatory Oversight Report, GenAI topic and Third-Party Risk topic, plus the Books and Records topic. Release date of 9 December 2025 per Debevoise, 12 December 2025.
- FINRA, Frequently Asked Questions About Advertising Regulation, Questions B.4 and D.8, posted 10 May 2024.
- FINRA, Regulatory Notice 21-29, 13 August 2021. Used for vendor due diligence and contract items.
- FINRA, Regulatory Notice 25-07, 14 April 2025. A request for comment, used for question E.3 and the undefined business as such phrase.
- FINRA, Regulatory Notice 26-14, 9 July 2026. A proposal, with comments due 11 September 2026. We read the notice, not the rule text in its Attachment A.
- SEC, 17 CFR 240.17a-4, via Cornell Legal Information Institute. FINRA Rule 4511 as described on FINRA's Books and Records page.
- Commentary, not primary: Walter Saurack, Duane Morris, AI and Broker-Dealer Compliance, 22 July 2026. Archive Intel, What FINRA's 2026 Oversight Report Means for AI and Recordkeeping, 8 September 2026. Snell & Wilmer, FINRA's 2026 Oversight Report Signals a Supervisory Reckoning for Autonomous AI, 17 December 2025.
Last verified 6 November 2026. The weakest thing about this source base: the status split and vendor evidence column are our own reading, and FINRA has not endorsed either. The status of Notice 26-14 should be rechecked before relying on row 5.
Related reading