In December 2025 FINRA released its 2026 Annual Regulatory Oversight Report. The document devoted a discrete section to generative and agentic AI, explicitly framing AI supervision under Rule 3110's reasonable-design obligation and reasserting Rule 4511 recordkeeping over AI artifacts (prompts, outputs, edits, signoffs). Within forty-eight hours, Sidley Austin, Snell & Wilmer, Debevoise & Plimpton, ACA Group, and Smarsh had published client alerts unpacking the operational implications. By February 2026, every CCO at every FINRA-member broker-dealer had been asked the same question by management: "what does our supervisory system actually do when a registered representative uses Jump to take notes, an internal LLM to draft a rollover memo, an agentic workflow to push a trade ticket through the OMS, and Smarsh to archive the result?" This lesson maps the three FINRA rules — 2210 (communications), 3110 (supervision), 4511 (recordkeeping) — and Regulatory Notice 24-09 onto the actual AI-touched workflows of a 2026 BD or hybrid practice, and lays out what a "reasonably designed" supervisory system looks like in the production environment.

FINRA's 2026 Oversight Report and Regulatory Notice 24-09 as the Operational Frame

The FINRA 2026 Annual Regulatory Oversight Report's GenAI section consolidated several prior threads: the July 2024 Regulatory Notice 24-09 (the staff's substantive GenAI guidance), the recurring exam-priority focus on communications supervision, and the 2025 enforcement signals on shadow AI use. The 2026 framing has three operational components.

First, AI use is firmly within the existing rule framework — FINRA did not create new rules for AI; existing Rules 2210, 3110, and 4511 (plus 4530 customer-complaint reporting, 4530 fraud reporting, and 2010 just-and-equitable principles) apply to AI-touched activity. Second, the 2026 Report explicitly frames agentic AI (AI taking action — sending an email, placing a trade, filing a form) under Rule 3110's reasonable-design obligation, requiring supervision architecture that anticipates AI action and includes books-and-records, pre-action compliance gating, post-action review, kill-switch design, and the supervisory log. Third, the 2026 Report reasserts that the Rule 4511 retention obligation extends to AI artifacts — system prompts, user prompts, AI tool / model version, AI outputs, human edits, supervisory signoffs — and that "shadow AI" (advisor use of unauthorized AI tools outside firm visibility) is itself a supervisory failure regardless of whether any specific recommendation has gone wrong.

Reg Notice 24-09's substantive guidance, as the Snell & Wilmer and Debevoise teardowns highlight, emphasized that AI tools must be evaluated for accuracy, fairness, privacy, security, and reliability, and that vendor due diligence, training, written policies, and supervisory testing all flow from the existing rule architecture. The Sidley alert specifically called out the reasonable-design standard's flexibility — what is reasonable depends on the firm's size, complexity, and AI footprint — and the ACA Group alert connected the supervisory framework to the SEC's parallel Rule 206(4)-7 compliance program requirements for RIAs. The Smarsh teardowns emphasized the practical recordkeeping implications: tamper-proof, timestamped retention across the full prompt-to-output-to-signoff chain.

Rule 2210 — Communications with the Public Applied to AI-Drafted Client Content

FINRA Rule 2210 governs communications with the public — retail communications, correspondence, institutional communications, and public appearances — with content standards (fair and balanced, not misleading), pre-use approval / principal review obligations in specified circumstances, recordkeeping, and filing requirements for certain categories. AI-drafted client-facing content is squarely within the rule.

Retail Communications

A "retail communication" under Rule 2210(a)(5) is any written (including electronic) communication that is distributed or made available to more than 25 retail investors within any 30 calendar-day period. Retail communications generally require principal pre-use approval under Rule 2210(b)(1) (with certain exceptions). AI-drafted website copy, mass emails, social posts, podcast descriptions, AI-generated capabilities decks distributed to clients and prospects — all sit in this category and require principal review against the rule's content standards and the firm's WSPs.

Correspondence

"Correspondence" under Rule 2210(a)(2) is any written (including electronic) communication distributed or made available to 25 or fewer retail investors within any 30 calendar-day period. Correspondence generally requires supervision and review under Rule 3110(b) but does not always require pre-use principal approval — though firms commonly extend principal pre-use review to AI-drafted correspondence given the volume and the AI-amplified risk of content drift. AI-drafted one-on-one follow-up emails, individualized client updates, and personalized planning memos sit in this category.

Content Standards Applied to AI Output

Rule 2210(d)(1) requires all communications to be fair and balanced, provide a sound basis for evaluating facts in regard to any particular security or service, and not omit material facts or qualifications. AI tools that draft fund recommendations, market commentary, or planning memos can produce content that emphasizes upside without proportionate risk disclosure, omits material qualifications, or makes claims requiring substantiation. The principal review under Rule 2210(b) is the firm's checkpoint; the AI red-team first-pass screen plus the human principal signoff is the L4 Ch3 design pattern.

Filing Requirements for Certain Categories

Rule 2210(c) requires filing with FINRA's Advertising Regulation Department of certain communication categories — for example, new-member retail communications, certain communications concerning registered investment companies including AI-generated content involving rankings, performance, and certain product categories. AI-generated retail content covering these categories must flow through the filing workflow if applicable.

Rule 3110 — Supervision and the Reasonably-Designed Standard Applied to AI

FINRA Rule 3110 requires each member to establish and maintain a supervisory system to supervise the activities of each associated person reasonably designed to achieve compliance with applicable securities laws and regulations and FINRA rules. The 2026 Report's framing extended this to AI: the supervisory system must be reasonably designed to achieve compliance even when AI tools are in the workflow, and agentic AI raises the supervisory bar.

The Supervisory System Components

Rule 3110(a) requires designation of qualified principals, designation of an OSJ (Office of Supervisory Jurisdiction) for each location, assignment of each registered person to a principal for supervision, and reasonable inspection of each location. Rule 3110(b) requires written procedures (WSPs) for supervision, principal review of correspondence and internal communications, principal pre-use approval of retail communications, and review of transactions for compliance.

What Reasonable AI Supervision Looks Like

The 2026 Report's reasonable-design framing for AI requires the supervisory system to address: (a) which AI tools are approved (the firm's AI inventory), (b) which data categories may be processed by each tool (the data-classification matrix tied to Reg S-P and GLBA), (c) the approval workflow for new use cases, (d) training requirements for registered persons using AI, (e) the supervisory review queue (principal review of AI-drafted communications, alternatives-considered adequacy for Reg BI, conflict drift detection), (f) the exception-handling workflow for AI outputs that flag concerns, (g) incident response for AI failures (hallucination affecting a recommendation, data leak, prompt injection, output compromise), and (h) the kill-switch architecture for agentic AI taking action. The Sidley alert specifically highlighted the flexibility — what is "reasonable" scales to the firm's size, complexity, and AI footprint.

Agentic AI Under 3110 Reasonable Design

The 2026 Report's most-cited language deals with agentic AI — AI that takes action rather than producing a draft for human approval. When AI sends an email, places a trade, files a form, processes an RMD, or initiates an ACATs request, the supervisory architecture changes. The Snell & Wilmer alert summarized the change: pre-action compliance gating (the agent's action passes a programmatic check before execution), post-action review (every agentic action is logged for principal review), kill-switch design (the firm can halt the agent's operations on demand), and the supervisory log under Rule 4511 retaining the goal, the agent's reasoning chain, the tool calls, and the outcome. The L4 Ch3 lesson on Agentic-AI WSPs operationalizes this architecture in detail.

Shadow AI as a Supervisory Failure

The 2026 Report's framing of shadow AI (advisor use of unauthorized AI tools outside firm visibility) treats the visibility gap itself as a supervisory failure. The Smarsh teardown emphasized that even if no specific recommendation has gone wrong, the firm's inability to see what AI is being used by whom for what is the failure. The remediation: an AI tool inventory, an acceptable-use policy enforced through identity-and-access management and endpoint monitoring, and a registered-person attestation cycle.

Rule 4511 — Recordkeeping Applied to AI Prompts, Outputs, and Meeting Transcripts

FINRA Rule 4511 requires members to make and preserve books and records as required by FINRA rules, the Exchange Act, and the rules thereunder. The default retention period is three years; specific categories have longer periods (six years for many transactional records, lifetime of the account for certain account records). The 2026 Report's reassertion of Rule 4511 retention over AI artifacts is operationally consequential.

What Counts as a Record

The 2026 Report and the Smarsh teardowns frame the AI production chain as records: (a) the meeting transcript (Jump, Zocks, FinMate, Sybill, Zeplyn) is a record of the meeting; (b) the system prompt and the user prompt are records of the input; (c) the AI tool and model version are records of the production environment; (d) the AI output is the produced artifact; (e) the human edits are records of registered-person revision; (f) the principal / compliance signoff is the supervisory record. The full chain is retained because the firm must be able to reconstruct how any AI-touched artifact was produced if an examiner asks.

Tamper-Proof, Timestamped Retention

Smarsh, Global Relay, and the major archiving vendors emphasize three operational requirements for AI artifacts: (a) tamper-proof storage (the artifact, once committed, cannot be silently altered — write-once, read-many architecture or equivalent), (b) timestamped capture (the artifact's creation moment is fixed), (c) cross-referenced retrieval (the meeting transcript, the prompt, the output, the edits, and the signoff are linked as one searchable record set). The Smarsh 2026 release on AI communications compliance, the Global Relay AI-aware archive search, and the practitioner consensus across Sidley, Debevoise, ACA Group all converge on this architecture.

Retention Periods

The default Rule 4511 period is three years. Specific records (account records, customer complaints, communications with the public) carry longer obligations. For dually-registered firms, the longer of Rule 4511 and SEC Rule 204-2 (five years for advisers, first two easily accessible) governs the practical retention. The practitioner convention: retain to the longer period, ensure cross-system consistency, and document the retention schedule in WSPs.

Off-Channel and Meeting Recordings

The 2024-2025 off-channel-communications enforcement (text-message, WhatsApp, personal email) made one principle explicit: the channel doesn't determine the recordkeeping obligation — the content does. AI meeting transcription (Jump, Zocks) extends this principle. The meeting audio (where recorded with consent and retained by the AI tool), the AI-generated transcript, and the AI-generated summary all join the record set. Off-channel AI use (an advisor's personal ChatGPT session covering client business) is a recordkeeping failure on the same logic as off-channel text messaging.

Putting It Together — The BD AI Supervisory Architecture

The 2026 BD with an AI-touched workflow needs five operational components:

1. AI Tool Inventory and Acceptable Use Policy — every tool the firm approves, what data each may process, who is approved to use each. Maps to Rule 3110 reasonable design and the Reg S-P / GLBA data classification.
2. Principal Pre-Use Review Workflow — AI-drafted retail communications under Rule 2210(b)(1), AI-drafted correspondence per WSP design, AI-drafted Reg BI recommendation memos. Risk-based sampling and AI red-team first-pass screening for volume management (L4 Ch3 design).
3. Production Chain Retention — meeting transcript + system prompt + user prompt + tool/model version + output + edits + signoff, retained tamper-proof and timestamped per Rule 4511 (and Rule 204-2 for hybrid firms).
4. Agentic AI Architecture — pre-action compliance gating, post-action review, kill-switch design, supervisory log. The L4 Ch3 lesson on Agentic-AI WSPs.
5. Incident Response and Shadow AI Detection — IRP under Reg S-P for AI-related breaches (30-day notification clock under the May 2024 amendments), shadow AI detection through IAM and endpoint monitoring, registered-person attestation cycle, and exception-handling workflow for hallucinations affecting client recommendations.

Cross-References and the Next Lesson

This lesson sits between the Reg BI lesson (L1 Ch4 L2) and the Reg S-P / state cyber lesson (L1 Ch4 L4 — next). The Marketing Rule lesson (L1 Ch4 L1) established the Rule 2210 interlock for hybrid practices; this lesson develops the BD-side rules in detail. The L4 Ch3 supervisory-procedures chapter (especially the agentic-AI WSPs lesson) is the deep-dive on the supervisory architecture, the L4 Ch3 principal-review lesson is the deep-dive on the AI review queue at scale, and the L3 Ch10 archiving lessons operationalize the Rule 4511 retention pipeline (Zocks-to-Wealthbox-to-Smarsh).

Key Takeaways

• The FINRA 2026 Annual Regulatory Oversight Report's GenAI section frames AI supervision under existing Rules 2210/3110/4511 (no new AI rules), explicitly extends Rule 3110 reasonable design to agentic AI, and reasserts Rule 4511 retention over AI artifacts (prompts, outputs, edits, signoffs). Reg Notice 24-09 is the substantive predecessor guidance.
• Rule 2210 applies to AI-drafted communications: retail communications (more than 25 retail investors in 30 days) generally require principal pre-use approval under 2210(b)(1); correspondence (25 or fewer) requires supervision under 3110(b) with firms commonly extending pre-use review to AI-drafted correspondence; content standards under 2210(d) require fair-and-balanced, sound-basis content with material qualifications not omitted.
• Rule 3110 reasonable design for AI requires: AI tool inventory, data-classification matrix tied to Reg S-P / GLBA, approval workflow for new use cases, training, supervisory review queue, exception handling, incident response, and kill-switch architecture for agentic AI. The standard scales to firm size, complexity, and AI footprint.
• Agentic AI under 3110 requires pre-action compliance gating, post-action review, kill-switch design, and a supervisory log retaining goal, reasoning chain, tool calls, and outcome — operationalized in L4 Ch3.
• Shadow AI is itself a supervisory failure — the visibility gap is the issue, not just incidents that occur in the gap. Remediation: AI inventory, acceptable-use policy enforced through IAM and endpoint monitoring, registered-person attestation.
• Rule 4511 record retention extends to the full AI production chain: transcript + system prompt + user prompt + tool/model version + output + edits + signoff, tamper-proof, timestamped, cross-referenced. Default three-year period; longer for specified categories. Dually-registered firms retain to the longer of Rule 4511 and SEC Rule 204-2 (five years, first two easily accessible).
• Off-channel AI use (personal ChatGPT for client business) is a recordkeeping failure on the same logic as off-channel text messaging — the content, not the channel, determines the obligation.