FINRA Rule 3110 + GenAI: what financial advisors need to know in 2026.

Drew Harris · CEO and Chief Product and Technology Officer · 2026-04-21 · 9 min read

complianceregulatoryzero-hallucinationpii

What Rule 3110 has always required

FINRA Rule 3110 (Supervision) establishes the baseline: member firms must establish and maintain a supervisory system reasonably designed to achieve compliance with applicable securities laws, rules, and FINRA regulations. In practice, that means:

Supervisory procedures documented in writing
Qualified supervisors assigned to each associated person
Review of communications with the public (emails, chat, social media, now AI-generated content)
Records of supervisory review retained per the rule's retention requirements
Internal inspections at prescribed cadences

The rule is technology-agnostic by design. It has always applied to any communication a registered rep sends. What changed in 2026 is FINRA's explicit guidance on how it applies to GenAI.

What changed in the 2026 report

The 2026 FINRA Annual Regulatory Oversight Report, GenAI section makes several operational clarifications:

Every GenAI-produced communication with a client or prospect is subject to Rule 3110. There is no "it was the AI, not the rep" exemption. The associated person using the tool is responsible; the firm supervising that person is responsible.
Firms must implement recordkeeping for prompt inputs, model outputs, and model versions in use at the time of each communication. The supervisory audit must be able to reconstruct what the AI produced and why.
Firms must supervise the AI system itself, not only the rep using it. That includes vetting the vendor, understanding the model's limitations, and reviewing how the system handles edge cases.
Agentic AI warrants heightened supervision. When an AI system takes autonomous actions (books appointments, sends messages, executes workflows) on behalf of a rep, the supervisory framework must anticipate and constrain those actions.

For advisors considering an AI protégé or clone, the question shifts from "is this product OK under existing rules?" to "does this product produce the records, version history, and supervisory hooks my compliance function needs?"

Rule 3110 was a supervision rule. The 2026 report is a records-and-systems rule. GenAI output without a prompt log, a version pin, and a reviewable trail creates a compliance gap the firm, not the vendor, will be held to.

The four supervisory requirements for GenAI

Distilling the 2026 guidance into evaluation criteria for an AI clone platform:

1. Prompt + output logging

Every interaction between the client and the AI should produce a record of:

What the client said (prompt/query)
What the AI produced (output/response)
Which knowledge base chunks (if RAG-based) were retrieved and scored
Which model version produced the output
Timestamp, session ID, and associated-person attribution

This is the recordkeeping substrate. Without it, the firm cannot reconstruct what happened during a client interaction, cannot demonstrate supervisory review, and cannot respond to a regulator's question about a specific conversation.

2. Model-version pinning

Models update. An answer produced by one version of a language model in March is not identical to one produced by a later version in June. For supervisory purposes, a firm needs to know which version was in force at the time of each interaction.

Our take: model-version pinning is a platform-level feature. The firm does not want to manage which model a given protégé runs on. The platform should pin, record, and version-tag every session.

3. Supervisory review workflows

Rule 3110 requires review of communications. A platform that produces 10,000 AI-client messages a month without a structured supervisory review path is, by construction, unsupervisable.

What "good" looks like:

Sampling logic (random, risk-flagged, high-volume client)
A reviewer's UI that surfaces flagged interactions
Reviewer disposition records (reviewed, issue noted, escalated, closed)
Written procedures for what triggers escalation

4. Auditable source-of-truth retrieval

For RAG-based systems (which includes nearly every AI clone platform), the firm needs to see:

What knowledge base content existed at the time of the interaction
What was retrieved for the specific query
What was not retrieved and why

This is the zero-hallucination architecture story told in compliance terms. A firm that can't trace an AI's answer back to source content cannot defend that answer under Rule 3110.

What "agentic AI" means to FINRA

"Agentic" is one of the more-abused words in AI marketing. FINRA's guidance narrows it to systems where the AI takes actions on behalf of the rep: scheduling meetings, sending follow-up communications, initiating workflows, or interacting with third-party systems.

Heightened supervision requirements for agentic systems:

Pre-action review: high-stakes actions should require human confirmation
Scope limits: the AI should have a documented, narrow set of actions it can take
Action logs: distinct from conversation logs; every autonomous action is individually recorded
Rollback capability: when an action was in error, the firm must be able to reverse it

For an AI clone platform, this is where the rules engine becomes a compliance feature as well as a productivity feature. Rules that fire on session events (send an email, book a call, escalate to human) are the agentic surface; treating them as supervised actions rather than background automations is the 3110-aligned posture.

How to evaluate an AI clone platform for Rule 3110 fit

A practical checklist. Run it in conversation with the vendor, not on their marketing site.

Can I export, for a given time window, every prompt and output attributable to a specific client? If not, you can't satisfy supervisory review.
Can I see which model version produced each output? If the vendor "just uses the latest" without pinning, you're inheriting model drift as a compliance risk.
Can I see the knowledge-base content and retrieval score for each answer? Without this, you can't defend an answer's provenance.
Is there a supervisory review UI, or am I expected to build one on top of raw logs? The maturity difference is months of compliance work.
Does the platform's rules engine / automation require human approval for client-facing actions? If not, the agentic surface is under-supervised by default.
Does the vendor have a written response to Rule 3110 / the 2026 FINRA guidance? "We're compliance-adjacent" is not an answer. Specific mechanisms mapped to specific requirements is.
What are the data retention defaults? Rule 4511 / Rule 3110 have records retention periods. The vendor's defaults should meet or exceed them.

A vendor who can walk through all seven with specific product surfaces is compliance-ready. A vendor whose answers are qualitative is not.

Our current posture (honestly stated)

We'd rather be specific about what we have and don't have than overclaim. Current Apex Replicant posture for FINRA Rule 3110 considerations:

What we have

Full session transcripts: every voice and text interaction is logged with timestamp and session ID
Structured session insights: seven-category summaries via Claude AI (session insights feature)
Zero-hallucination architecture: retrieval-gated generation; below-threshold responses route to explicit "I don't know" (patent-protected retrieval architecture)
PII redaction on ingest: regulated-grade detection pipeline (PII handling details)
Rules engine with action logging: every rule execution is recorded
Structured output per protégé: JSON schema for intake that produces records, not just prose

What we don't yet have (stated plainly)

Named FINRA-aligned audit / compliance certification: we are compliance-adjacent, not certified
Retrieval audit trail (customer-exposed): not built yet. Session insights capture what was retrieved for each answer at the aggregate level, but a per-answer "which KB chunks at what similarity score" trail exportable to a supervisor is not shipped. We would add this on request for a financial-vertical deployment
Reviewer-specific UI optimized for 3110 supervisory workflows: logs and transcripts are exportable, but the supervisor review interface is on our roadmap, not shipped
Model-version pinning as a user-configurable feature: we pin internally but have not exposed the configuration to customers
Written vendor-response document mapping our product to each 3110 sub-requirement: in development; contact our team for current draft

For a firm evaluating AI clones for a financial-advisor deployment, the right posture is: ask for a direct walkthrough with our team, bring your compliance counsel, and evaluate against your specific supervisory program. We will not claim readiness we don't have.

FAQ

Is Apex Replicant "FINRA-compliant"? No AI platform can be "FINRA-compliant"; compliance is a firm-level property determined by how the firm uses the tools. What a platform can be is FINRA-aligned: shipping the primitives (logging, version pinning, review workflows, audit trails) that a firm needs to build compliant usage on top. Our current posture is architecturally strong, operationally partial. See "what we don't yet have" above.

Does the 2026 report apply to me if I'm a solo RIA or a state-registered adviser? FINRA rules apply to FINRA members (broker-dealers and their registered persons). RIAs are regulated by the SEC or state securities regulators, who have issued separate GenAI guidance. The practical supervisory expectations are similar: records, version tracking, review. Your specific rule set depends on your registration. Consult your compliance counsel.

Can I run an AI protégé for financial advice today under Rule 3110? You can run an AI system whose outputs constitute communications with clients, subject to full 3110 supervision. Whether your specific firm can do so safely depends on your supervisory procedures, your records systems, and the specific platform you use. This is a compliance-counsel conversation, not a marketing decision.

What about recordkeeping rules (Rule 4511, 17a-3, 17a-4)? Those govern retention duration and format. They apply to AI-generated records the same way they apply to email. Check with your compliance team on retention defaults for any AI platform. Our session records are exportable and retained by default; contact us for specific retention configurations.

How does this relate to SEC guidance on AI for advisers? SEC Staff Bulletin (2024) and subsequent guidance emphasize adviser disclosure obligations when AI is used in client-facing communications, conflicts-of-interest considerations, and the duty of care. The 2026 FINRA report focuses on supervisory recordkeeping. Both are relevant for dual-hat firms. Again, compliance counsel.

Is there a vertical-specific Apex Replicant offering for finance? Our finance vertical page details our current feature emphasis and compliance-posture flags for the vertical. The honest short version: we are positioning for compliance-first financial advisor use cases and will not ship named certifications we haven't earned. Contact us for the current draft of our FINRA-alignment documentation.