Control 3.6: Supervision and Oversight (FINRA Rule 3110 / SEC Reg BI)
Control ID: 3.6 Pillar: Compliance & Audit Regulatory Reference: FINRA 3110 (Supervision), FINRA 3120 (Supervisory Control System), SEC Regulation Best Interest (Reg BI), FINRA Regulatory Notice 24-09 Last Verified: 2026-02-17 Governance Levels: Baseline / Recommended / Regulated
Objective
Establish written supervisory procedures, qualified supervisor assignments, ongoing supervisory activities, and annual testing programs that address Microsoft 365 Copilot usage by associated persons, helping meet FINRA Rule 3110 supervisory obligations and SEC Regulation Best Interest requirements for AI-assisted activities — including the expanded supervisory obligations that arise from deploying M365 Copilot agents.
Why This Matters for FSI
FINRA Rule 3110 requires that each member firm establish, maintain, and enforce written supervisory procedures (WSPs) and a system to supervise the activities of each associated person that is reasonably designed to achieve compliance with applicable securities laws, regulations, and FINRA rules. This obligation is absolute and extends to every tool and technology used by associated persons -- including Microsoft 365 Copilot.
FINRA Rule 3120 requires that firms designate and specifically identify a qualified registered principal who is responsible for establishing, maintaining, and enforcing a system of supervisory control policies and procedures. The rule further requires annual testing of supervisory procedures.
SEC Regulation Best Interest requires that broker-dealers act in the best interest of retail customers when making recommendations, and that they establish policies and procedures reasonably designed to address conflicts of interest. When Copilot assists in generating recommendations or recommendation-adjacent content, supervisory controls must help verify that Reg BI obligations are met.
FINRA Regulatory Notice 24-09 explicitly reminds firms that existing supervisory obligations apply to AI-generated content, and that firms must "reasonably supervise" the use of AI tools by associated persons. This means supervision cannot be delegated to the AI tool itself -- human supervisory oversight is required.
Control Description
This control covers the comprehensive supervisory framework required for Copilot usage, including written supervisory procedures, qualified supervisor assignment, supervisory review activities, annual testing requirements, and documentation obligations.
Supervisory Framework Components
| Component | FINRA Rule | Copilot Application |
|---|---|---|
| Written supervisory procedures | 3110(b) | WSPs must specifically address Copilot usage policies, prohibited activities, review requirements, and escalation procedures |
| Supervisory system | 3110(a) | The overall system must account for Copilot as a tool used in supervised activities |
| Qualified supervisor | 3110(a) | Designated supervisors must understand Copilot capabilities and limitations |
| Review of communications | 3110(b)(4) | Supervisory review must cover Copilot-assisted communications |
| Review of customer accounts | 3110(b)(2) | Supervisory review of Copilot-assisted account activities |
| Annual compliance meeting | 3110(a)(7) | Annual meeting must address Copilot usage and compliance |
| Supervisory control system | 3120 | Supervisory controls must test Copilot-specific procedures |
| Annual testing | 3120(b) | Annual testing must include Copilot-related supervisory procedures, including AI tool supervision effectiveness |
| Agent supervision | 3110(a), 3120 | WSPs must address each deployed Copilot agent's scope, authorized actions, and supervisory review cadence |
Written Supervisory Procedures -- Copilot Addendum
The firm's WSPs should include a Copilot-specific addendum addressing:
- Authorized uses of Copilot -- Which business activities may use Copilot assistance
- Prohibited uses of Copilot -- Activities where Copilot may not be used (e.g., generating investment recommendations without human review, producing compliance reports without verification)
- Pre-use review requirements -- Communications requiring supervisory approval before distribution
- Post-use review procedures -- Sampling and review cadence for Copilot-assisted activities
- Escalation procedures -- When and how to escalate Copilot-related compliance concerns
- Training requirements -- Required training before Copilot access is granted
- Recordkeeping -- How Copilot usage records are maintained and accessible
- Incident response -- Procedures for addressing Copilot-related compliance failures
- Agent scope and oversight -- For each deployed Copilot agent (Teams channel agents, declarative agents): authorized scope, permissible actions, supervisory review cadence, and agent-generated output review requirements
Supervisory Review Triggers
Certain Copilot-assisted activities should trigger mandatory supervisory review:
| Trigger | Review Requirement | Timing |
|---|---|---|
| Copilot-drafted communication to retail client about investment products | Principal review | Pre-send (retail communications) or within 24 hours (correspondence) |
| Copilot-generated financial analysis shared with clients | Supervisor verification of data accuracy | Before client delivery |
| Copilot-assisted response to regulatory inquiry | Senior compliance review | Before submission |
| Copilot-drafted marketing or advertising materials | Principal pre-approval per FINRA 2210 | Before use |
| Copilot-generated recommendation or suitability analysis | Supervisor review for Reg BI compliance | Before client communication |
| Copilot-assisted handling of customer complaint | Compliance review | Before response sent |
| Agent-generated content distributed to clients (email drafted by agent, document generated by agent) | Supervisory review equivalent to human-authored content | Pre-send or within 24 hours per communication type |
Agentic AI Supervision Requirements
FINRA 2026 Agentic AI Supervision and M365 Copilot Agents
FINRA's evolving supervisory expectations signal that firms deploying AI agents must extend their FINRA Rule 3110(a) supervisory systems to cover autonomous and semi-autonomous agent actions. For Microsoft 365 Copilot, this means bringing the following agent types within the scope of written supervisory procedures:
Teams channel agents that respond to user queries within Teams channels require supervisory oversight equivalent to any other tool used by associated persons. Where a Teams channel agent assists in responding to client-facing inquiries or generating content that is forwarded to clients, those interactions fall within the firm's supervisory obligations under FINRA 3110.
Declarative agents built on the Microsoft 365 Copilot platform that access organizational data and respond to user queries must be covered by the firm's written supervisory procedures. Each declarative agent should be documented in the WSP addendum with its authorized scope, the data sources it can access, and the supervisory review process for outputs it generates. Agents that are given access to client data or produce client-facing content carry the highest supervisory exposure.
Agent actions that result in client-facing outputs — emails drafted by agents, documents generated by agents, summaries forwarded to clients — require the same pre-use or post-use supervisory review as human-authored content. The automated origin of the content does not reduce the firm's supervisory obligation; it increases the importance of having documented oversight procedures because the volume and velocity of agent-generated content may exceed what traditional spot-check sampling captures.
Institutions should prepare for FINRA examination scrutiny of agent supervision as part of the 2026 Annual Risk Monitoring and Examination Priorities cycle. Firms that have already extended their supervisory frameworks to cover M365 Copilot agents and documented this coverage in their WSPs will be better positioned to demonstrate supervisory effectiveness.
Full-Chain Telemetry for Agent Decision Reconstruction
When multiple agents or Copilot interactions contribute to a decision or output — for example, a declarative agent queries SharePoint data, summarizes it, and the associated person forwards the summary to a client — the firm must be able to reconstruct the full chain of AI interactions that led to the final output. This capability supports both FINRA Rule 3110 supervisory documentation and FINRA Rule 3120(b) annual testing of supervisory effectiveness.
The telemetry chain for decision reconstruction follows this path:
- CopilotInteraction audit events — captured in Microsoft Purview audit log (see Control 3.1), records the initial Copilot prompts and responses
- Agent-specific audit fields —
AgentId,AgentName, andXPIA(cross-prompt injection attempt) flags within CopilotInteraction records identify agent-specific interactions - Communication Compliance records — captures the outbound communication or document if it passed through a supervised location (see Control 3.4)
- Final business record — the retained communication or document in the firm's records system (see Control 3.11)
Tier guidance for decision reconstruction capability:
| Tier | Requirement |
|---|---|
| Baseline | Log all agent interactions using CopilotInteraction audit events; preserve agent-specific audit fields |
| Recommended | Correlate agent audit trails with business outcomes — match agent session IDs to downstream communications reviewed by supervisors |
| Regulated | Implement full decision reconstruction capability with documented chain-of-custody for all agent-assisted client-facing outputs; maintain reconstruction records available for examination response |
Copilot Surface Coverage
| Copilot Surface | Supervisory Concern | Review Mechanism |
|---|---|---|
| Outlook Copilot | Drafting client communications with investment content | Communication Compliance + supervisory review workflow |
| Teams Copilot | Meeting recaps shared with clients; Copilot-assisted client conversations | Teams message monitoring + post-meeting review |
| Word Copilot | Generating client proposals, investment memos, compliance reports | Document review before external distribution |
| Excel Copilot | Financial analysis, portfolio modeling, performance calculations | Data accuracy verification before client use |
| PowerPoint Copilot | Client presentation decks, investment reviews | Pre-meeting supervisory review |
| Microsoft 365 Copilot Chat | Research, information gathering that informs client activities | Periodic sampling of Copilot Chat usage patterns |
| Copilot Pages | Collaborative content that may be shared externally | External sharing controls + review requirement |
| Teams channel agents | Automated responses to user queries within Teams channels | Supervisory review per FINRA 3110(a); WSP coverage required |
| Declarative agents | Agents that access organizational data and produce user-facing responses | WSP scope documentation; supervisory review cadence for client-facing outputs |
Governance Levels
Baseline
- Update written supervisory procedures to address Copilot usage by associated persons
- Designate qualified supervisors responsible for Copilot usage oversight
- Require supervisory pre-approval for all Copilot-drafted retail communications
- Configure at least one Communication Compliance policy with "Microsoft 365 Copilot and Microsoft 365 Copilot Chat" selected as a monitored location (Purview portal > Communication compliance > Policies > Create policy > Locations). Copilot content is a location within standard CC policies — no separate policy type is required
- Set communication direction to Internal for user-to-AI interaction monitoring
- Establish a post-use review sampling rate for Copilot-assisted correspondence (minimum 5%) via CC policy > Review percentage > 1–100% (portal-only configuration; not available via PowerShell)
- Include Copilot usage in the annual compliance meeting agenda
- Document Copilot-authorized and prohibited uses in firm policies
- Maintain a log of all deployed Copilot agents (Teams channel agents, declarative agents) with their authorized scope and data access
Recommended
- Increase post-use review sampling rate to 25% of Copilot-assisted communications (FSI-recommended baseline); configure via CC policy > Review percentage in the Purview portal
- Implement Communication Compliance policies with automated flagging for supervisory review (see Control 3.4), including:
- Trainable classifiers: Financial regulatory text, Inappropriate content, Azure AI Content Safety classifiers for Copilot prompt/response review
- Sensitive information types: Up to 30 SITs per policy; include FSI-relevant SITs (ABA Routing Number, U.S. Social Security Number, Credit Card Number, U.S. Bank Account Number)
- Direction: Internal (for user-to-AI interactions)
- Assign Communication Compliance reviewer roles with appropriate permissions:
- Communication Compliance Admin: Full policy management
- Communication Compliance Analyst: Review flagged items, tag, escalate
- Communication Compliance Investigator: All Analyst permissions plus view message body content
- Communication Compliance Viewer: View reports only
- Create Copilot-specific supervisory review checklists for each communication type
- Establish monthly supervisory review reports documenting volumes reviewed, issues found, and dispositions
- Train supervisors on Copilot capabilities, limitations, and common failure modes including agent behaviors
- Implement enhanced supervision procedures for associated persons with recurring Copilot-related compliance issues
- Integrate Copilot supervisory review data into the firm's compliance dashboard
- Update WSPs to include agent-specific supervisory procedures for all deployed Teams channel agents and declarative agents
- Implement recommended-tier decision reconstruction: correlate agent audit trails with downstream supervised communications
Regulated
- Implement pre-send supervisory review for all Copilot-drafted client-facing communications (not just retail communications)
- Escalate to 100% review sampling rate for priority users (executives, trading desk, M&A staff) via CC policy > Review percentage
- Define and document the escalation path: Analyst → Investigator → Legal/Compliance team (via case escalation in CC portal); escalate to eDiscovery case directly from a Communication Compliance alert
- Deploy automated Copilot usage analytics to identify high-risk usage patterns (e.g., high-volume Copilot use for client communications, Copilot use during trading hours by restricted persons)
- Automate policy provisioning via PowerShell where supported:
- Modern:
New-CommunicationCompliancePolicy+New-CommunicationComplianceRule(requires E5 Compliance or Communication Compliance add-on; verify cmdlet availability withGet-Command *CommunicationCompliance*) - Legacy:
New-SupervisoryReviewPolicyV2(still functional for supervisory review scenarios) - Limitation: Sampling rate is portal-only and not configurable via PowerShell
- Modern:
- Conduct quarterly FINRA 3120 testing of Copilot-specific supervisory procedures with documented results, including explicit testing of agent supervision effectiveness
- Maintain real-time supervisory dashboards showing Copilot usage by associated persons, including agent interaction volumes
- Implement supervisory exception reports for Copilot activities outside normal patterns
- Require annual certification by designated supervisors confirming Copilot supervisory procedures are effective
- Prepare examination-ready evidence packages demonstrating supervisory system effectiveness for Copilot activities
- Document supervisory decisions and rationale in a format that supports regulatory examination response
- Implement full decision reconstruction capability with documented chain-of-custody for all agent-assisted client-facing outputs
Setup & Configuration
Step 1: Update Written Supervisory Procedures
- Draft a Copilot Usage Addendum to the firm's existing WSPs
- Include the following sections:
- Scope and applicability (which associated persons, which Copilot features, which deployed agents)
- Authorized uses (permissible business activities)
- Prohibited uses (activities that may not use Copilot)
- Communication review requirements (pre-use, post-use, sampling rates)
- Supervisory review procedures and escalation
- Training requirements before Copilot access
- Recordkeeping requirements
- Consequences for policy violations
- Agent inventory with scope, authorized actions, and review cadence for each deployed Teams channel agent and declarative agent
- Route the WSP addendum through compliance, legal, and senior management approval
- Distribute to all associated persons and obtain acknowledgment
Step 2: Designate Qualified Supervisors
- Identify registered principals who will supervise Copilot-related activities
- Verify supervisors hold appropriate registrations (Series 24, Series 26, or equivalent)
- Provide Copilot-specific training to designated supervisors covering:
- How Copilot works (grounding, prompt-response cycle, limitations)
- Common Copilot failure modes in financial communications
- How to identify Copilot-assisted content in the review queue
- How to identify agent-generated content (AgentId, AgentName fields in audit records)
- Firm-specific review criteria and escalation procedures
- Document supervisor assignments in the firm's supervisory designation system
Step 3: Configure Supervisory Review Workflows
- Implement the Communication Compliance policies defined in Control 3.4
- Configure review queues routed to designated supervisors
- Establish review SLAs:
- Retail communications: Pre-send approval (same business day)
- Correspondence: Post-use review within 24 hours
- Institutional communications: Post-use review within 48 hours
- Agent-generated client-facing content: Post-use review within 24 hours
- Configure escalation for items exceeding SLA
Step 4: Establish Sampling Procedures
For post-use review of correspondence and institutional communications:
- Define sampling methodology (random, risk-based, or stratified)
- Set sampling rates:
- Baseline: 5% of Copilot-assisted communications
- Recommended: 10% of Copilot-assisted communications
- Regulated: 100% of outbound Copilot-assisted client communications (pre-send review)
- Document sampling methodology and results monthly
- Adjust sampling rates based on violation rates and risk assessment
Step 5: Configure Annual Testing Program
Per FINRA Rule 3120, establish annual testing of Copilot supervisory procedures:
- Define test scope: all Copilot-related supervisory procedures
- Assign testing to qualified personnel independent of the supervision function
- Test plan should include:
- Verification that WSPs are current and address all Copilot surfaces
- Sample review of supervisory approvals for completeness and timeliness
- Testing of communication compliance policy effectiveness (false positive/negative rates)
- Verification that training requirements are enforced
- Testing of recordkeeping completeness
- Verification of escalation procedures
- Explicit testing of agent supervision effectiveness — verify that WSPs cover all deployed agents and that agent-generated outputs are being reviewed
- Document findings and remediation in the annual supervisory testing report
- Present results to senior management per FINRA 3120(c)
Financial Sector Considerations
SEC 2026 Internal AI Examination Focus
The SEC's 2026 examination priorities are expected to include focused review of how registered entities govern internal AI tools, including Microsoft 365 Copilot. Institutions should prepare for examination questions about AI tool inventories, supervisory procedures specific to AI-assisted activities, and evidence of ongoing monitoring of AI-generated outputs. Firms should ensure they can demonstrate not only that written supervisory procedures reference AI tools, but that active supervisory review is occurring and that annual testing per FINRA 3120(b) explicitly includes AI tool supervision effectiveness as a tested element.
SEC Regulation Best Interest Alignment
When Copilot assists in generating content that constitutes or accompanies a "recommendation" to a retail customer, Reg BI obligations apply:
- Disclosure obligation: The associated person must disclose the material facts about the recommendation, including that AI assisted in its preparation
- Care obligation: The associated person must exercise reasonable diligence, care, and skill in evaluating the Copilot-generated recommendation
- Conflict of interest obligation: Firms must identify and mitigate conflicts, including potential biases in Copilot-generated recommendations
- Compliance obligation: Firms must establish policies and procedures reasonably designed to achieve compliance with Reg BI, including for AI-assisted recommendations
Heightened Supervision
Firms should consider implementing heightened supervision for:
- Associated persons who are new Copilot users (first 90 days)
- Associated persons with a history of compliance issues
- Associated persons in high-risk roles (client-facing, recommendation-making)
- Activities in high-risk product areas (complex products, alternative investments)
- Newly deployed Copilot agents until their behavior is understood and documented
Branch Office Considerations
For firms with branch offices, supervisory procedures must address:
- How branch-level supervisors oversee Copilot usage at the branch level
- Whether centralized or decentralized supervisory review is used for Copilot communications
- How remote-work Copilot usage is supervised
- Branch audit procedures for Copilot compliance
Evidence of Supervision
Examiners expect to see documented evidence of supervisory activity. For Copilot supervision, maintain:
- Review logs with reviewer identity, date, communication reviewed, and disposition
- Sampling methodology documentation and sample selection records
- Escalation records with investigation outcomes
- Training completion records for supervisors and supervised persons
- WSP version history showing Copilot-related updates
- Annual testing results with remediation documentation
- Agent inventory showing each deployed agent, its authorized scope, and supervisory review cadence
Verification Criteria
| # | Verification Step | Expected Outcome | Governance Level |
|---|---|---|---|
| 1 | Review WSPs for Copilot-specific addendum | WSPs include comprehensive Copilot usage procedures including agent scope | Baseline |
| 2 | Verify supervisor designations | Qualified principals are designated for Copilot supervisory oversight | Baseline |
| 3 | Confirm pre-approval workflow for retail communications | Test retail communication is held for principal approval | Baseline |
| 4 | Review sampling documentation | Sampling methodology is documented and sample reviews are recorded | Recommended |
| 5 | Verify supervisor training completion | All designated Copilot supervisors have completed required training including agent supervision | Recommended |
| 6 | Test escalation workflow | Flagged Copilot communication is escalated within SLA when reviewer does not act | Recommended |
| 7 | Review annual testing report | FINRA 3120 testing covers Copilot-specific procedures including agent supervision effectiveness with documented findings | Regulated |
| 8 | Verify real-time supervisory dashboard | Dashboard shows Copilot usage volumes, review status, and exception alerts | Regulated |
| 9 | Test examination-ready evidence package | Evidence package demonstrates complete supervisory system for Copilot activities | Regulated |
| 10 | Verify enhanced supervision procedures | Associated persons with compliance history have documented heightened Copilot supervision | Regulated |
Additional Resources
- FINRA Rule 3110 (Supervision)
- FINRA Rule 3120 (Supervisory Control System)
- FINRA Regulatory Notice 24-09 (AI and Member Firm Obligations)
- SEC Regulation Best Interest
- FINRA Report on Examination and Risk Monitoring
- Control 3.1 -- Audit Log and Data Access Governance
- Control 3.4 -- Communication Compliance Monitoring
- Control 3.5 -- FINRA Rule 2210 Compliance
- Control 3.11 -- Record Keeping and Books-and-Records Compliance
-
Related Controls: 3.4 Communication Compliance, 3.5 FINRA 2210 Compliance, 3.1 Copilot Audit Logging
FSI Copilot Governance Framework v1.2.1 - March 2026