Skip to content

Regulatory Framework

Comprehensive mapping of framework controls to US financial services regulatory requirements.

Overview

This document maps the FSI Agent Governance Framework controls to applicable US financial regulations. Organizations should use this mapping to prioritize control implementation based on their regulatory profile.

Disclaimer

This mapping is provided for informational purposes and does not constitute legal or regulatory advice. Regulatory interpretations vary by institution type and use case. Consult legal counsel for specific compliance requirements.

Primary US Financial Regulations

FINRA Rule 4511 — Books and Records

Overview: Requires firms to maintain records of all agent activities and communications.

Key Requirements:

  • Retention periods vary by record type (see matrix below)
  • All communications with customers
  • All agent outputs and decisions
  • Approval and supervisory records

Record Type Matters for Retention

Retention periods vary by record type. Agent conversation logs typically qualify as "communications" with 3-year retention under SEC 17a-4(b)(4), not the 6-year period for financial/customer records.

Retention Period Matrix

Record Type Retention Regulation Access Requirement
Communications (agent logs, chat, email) 3 years SEC 17a-4(b)(4) First 2 years readily accessible
Accounting/Financial Records 6 years SEC 17a-4(a) First 2 years readily accessible
Customer Account Records 6 years after account close SEC 17a-4(c)(e)(5) First 2 years readily accessible
FINRA-Specific Records (no SEC period) 6 years FINRA 4511(b) First 2 years easily accessible
Partnership/Corporate Records Life of enterprise + 3 years SEC 17a-4(d) Readily accessible
Audit Workpapers 7 years SOX 802 Accessible for examination

Agent Logs as Communications

Agent conversation logs (prompts, responses, interaction history) typically fall under the 3-year communications retention period per SEC 17a-4(b)(4), not the 6-year financial records period. However, if agent interactions generate or modify financial records, those outputs follow the applicable 6-year period.

Applicable Controls:

Control Requirement Mapping
1.7 Comprehensive Audit Logging Retention per record type (3 years for communications, 6 years for financial records)
1.9 Data Retention and Deletion Retention policies per FINRA timeline
1.21 Adversarial Input Logging Record security incidents and attacks
2.12 Supervision and Oversight Compliance Officer oversight
2.13 Documentation and Record Keeping All records documented
3.1 Agent Inventory Central registry of all agents
3.3 Compliance and Regulatory Reporting Regular compliance reports
3.4 Incident Reporting Document all incidents
3.10 Hallucination Feedback Loop Record and track accuracy issues

Zone Requirements:

  • Zone 2: Maintain 1-year audit logs, document approval process, monthly compliance reviews
  • Zone 3: Maintain audit logs per retention matrix (3 years for communications, 6+ years for financial records; first 2 years readily accessible), comprehensive real-time monitoring, immediate incident escalation

FINRA Rule 3110 — Supervision

Overview: Requires written policies and procedures for supervision of agents and AI technologies.

Key Requirements:

  1. Written supervisory procedures
  2. Qualified supervisor assignment
  3. Ongoing supervision and review
  4. Documentation of supervisory activities

FINRA Rule 3120 — Supervisory Control System

Overview: Requires annual testing and verification of supervisory procedures established under Rule 3110.

Key Requirements:

  1. Annual testing of supervisory control systems
  2. Documented testing procedures and results
  3. Escalation of identified exceptions
  4. Remediation of control deficiencies

AI Agent Governance Application:

Test Area Annual Testing Requirement
WSP Adherence Verify AI agent supervision procedures are followed
HITL Functionality Test that human review triggers function correctly
Escalation Procedures Verify escalation routing works as designed
Review Queue SLA Audit that reviews complete within target timeframes
Sampling Protocol Confirm sampling rates match policy

See: FINRA Rule 3120

Applicable Controls:

Control Requirement Mapping
2.3 Change Management Change control and approval
2.5 Testing and Validation QA before production
2.6 Model Risk Management OCC Bulletin 2026-13 (formerly OCC Bulletin 2011-12) / Fed SR 26-2 (formerly SR 11-7) alignment
2.11 Bias Testing Fairness assessment
2.12 Supervision and Oversight Define supervisory procedures
2.17 Multi-Agent Orchestration Limits Supervise agent interactions
2.18 Conflict of Interest Testing Test for recommendation biases
3.3 Compliance Reporting Supervision documentation

Zone Requirements:

  • Zone 1: No formal supervision required
  • Zone 2: Basic supervisory procedures, quarterly compliance reviews, annual testing
  • Zone 3: Comprehensive supervision, real-time monitoring, mandatory incident escalation, monthly compliance certification

FINRA AI Supervision Requirements

FINRA Notice 25-07 — Monitored Proposal (Not Adopted)

FINRA Regulatory Notice 25-07 (April 2025) is a Request for Comment (RFC) on workplace modernization, not an adopted rule. Its AI-relevant content is concentrated in Section E.3 (recordkeeping challenges for AI-generated communications under Exchange Act Rule 17a-4(b)(4)) and Section G (AI-driven fraud). The comment window closed July 2025; no final rule has been adopted. This framework monitors 25-07 for potential future recordkeeping requirements but does not treat it as binding. For current, adopted AI supervision guidance, refer to FINRA Regulatory Notice 24-09 (Gen AI guidance), FINRA Rule 3110 (Supervision), FINRA Rule 2111 (Suitability), and FINRA's Annual Regulatory Oversight Report for current AI examination priorities.

FINRA Regulatory Notice 24-09 (June 2024)

FINRA Notice 24-09 provides official guidance on generative AI and large language model (LLM) obligations. Key points:

  • Technology-neutral principle: Existing FINRA rules apply equally to AI-generated content
  • Rule 3110 supervision: Firms must establish supervisory procedures for AI tools
  • Rule 2210 communications: AI-generated customer communications must meet content standards
  • Model risk management: Firms should apply appropriate governance to AI systems

See: FINRA Regulatory Notice 24-09

FINRA 2026 Annual Regulatory Oversight Report (December 2025)

The 2026 Annual Regulatory Oversight Report contains FINRA's most detailed AI agent supervision guidance to date, with a dedicated section on generative AI and agentic systems.

Key AI Agent Guidance from 2026 Report

Topic Requirement Framework Control
AI as Supervisory Function Document WSPs for AI supervision substitution; define boundaries for AI vs. human oversight 2.12
Audit Trail Completeness Retain prompts, model state, and reasoning chain—not just outputs 1.7
Decision Reconstruction Demonstrate how agents reached conclusions for examination 1.7, 2.13
Agent Autonomy Limits Dedicated supervisory procedures for autonomous AI agents 2.12, 2.17
Rule 3120 Testing Annual testing of AI supervisory controls per Rule 3120 2.12

See: FINRA 2026 Annual Regulatory Oversight Report

Overview: FINRA's AI supervision requirements derive from existing rules that apply to associated persons' use of AI tools for customer communications and recommendations.

Key Requirements:

  1. Written Supervisory Procedures (Rule 3110) — Document AI tool approval, supervisory review, escalation paths
  2. Suitability (Rule 2111) — Validate AI recommendations meet suitability standards
  3. Recordkeeping (Rule 4511) — Retain AI-generated communications and agent logs

Applicable Controls:

Control Topic Mapping
1.7 Comprehensive Audit Logging Records retention for AI communications
2.5 Testing and Validation Agent accuracy testing
2.6 Model Risk Management Formal framework per SR 26-2 (formerly SR 11-7)
2.11 Bias Testing Fairness assessment per SR 26-2 (formerly SR 11-7)
2.12 Supervision Written supervisory procedures
3.2 Usage Analytics Performance monitoring
3.10 Hallucination Feedback Loop Monitor output accuracy

Framework Approach: The framework applies FINRA's existing supervision principles to AI agents, treating them as tools requiring documented procedures, ongoing monitoring, and supervisory oversight.

SEC Rule 17a-3/4 — Recordkeeping

Overview: Requires SEC-registered firms to maintain records of all transactions and communications. Retention periods vary by record type — see the Retention Period Matrix in the FINRA 4511 section above for details.

Record Categories:

  • Agent Communications: All user interactions, outputs, decisions (3 years per SEC 17a-4(b)(4), first 2 years readily accessible)
  • Transaction Records: If agent processes transactions, provides advice, executes trades (6 years per SEC 17a-4(a), first 2 years readily accessible)
  • Governance Records: Approvals, change logs, incident reports, validation results (6 years minimum per FINRA 4511(b))

Applicable Controls:

Control Requirement Mapping
1.7 Comprehensive Audit Logging Retention per record type (3-6 years), first 2 years readily accessible
1.9 Data Retention Retention policies enforced
1.19 eDiscovery for Agent Interactions Search and export for regulators
2.13 Documentation and Record Keeping All records documented
3.1 Agent Inventory Registry of agents as records
3.3 Compliance Reporting Evidence retention
4.6 Grounding Scope Governance Knowledge source records
4.7 M365 Copilot Data Governance M365 Copilot usage logging

SEC Regulation Best Interest (Reg BI)

Overview: Regulation Best Interest (17 CFR 240.15l-1, SEC Release No. 34-86031, compliance date June 30, 2020) establishes a "best interest" standard of conduct for broker-dealers when recommending securities transactions or investment strategies to retail customers. Reg BI examinations remain an annual SEC and FINRA priority. AI agents that participate in generating, screening, scoring, or framing recommendations may bring those recommendations within Reg BI scope, even when a human registered representative delivers the final advice.

The Four Reg BI Obligations (17 CFR 240.15l-1(a)(2)):

# Obligation AI-Agent Scope Primary Controls
1 Care Obligation Reasonable basis, customer-specific suitability, series-of-transactions analysis for AI-assisted recommendations 2.5, 2.6, 2.11, 3.10
2 Disclosure Obligation Form CRS / Reg BI disclosure should describe AI use, decision categories, and human-review checkpoints 2.19, 2.13, 2.21
3 Conflict of Interest Obligation Identify, disclose, eliminate, or mitigate conflicts embedded in model training data, product-coverage scoring, vendor revenue arrangements 2.18, 2.7, 1.5
4 Compliance Obligation Written supervisory procedures reasonably designed to achieve Reg BI compliance, including AI-specific testing, escalation, and remediation 2.12, 1.21, 3.3, 3.4

Zone Requirements:

  • Zone 1 (Personal): WSPs should explicitly prohibit personal-productivity agents from generating customer-deliverable recommendation content
  • Zone 2 (Team): Disclosure documentation, supervisory review, and validation evidence required for any agent that drafts, screens, or aggregates recommendation inputs
  • Zone 3 (Enterprise / Customer-Facing): Full Reg BI control set with documented Care, Disclosure, Conflict-of-Interest, and Compliance obligation evidence; bias-testing and conflict-testing reports retained per FINRA 4511 / SEC 17a-4(b)(4)

Framework Approach: The framework helps support Reg BI through 14 mapped controls. Reg BI is a registered-broker-dealer obligation; legal and compliance review is required to confirm the firm's WSPs, Form CRS, and Reg BI disclosure documents address the firm's specific facts and AI use cases. See the Reg BI section in Regulatory Mappings for detailed control mappings.

SEC Regulation SCI — Systems Compliance and Integrity

Overview: Regulation SCI (17 CFR §§ 242.1000–242.1007, SEC Release No. 34-73639) establishes uniform technology infrastructure requirements for "SCI entities" — large broker-dealers (per Rule 1000 thresholds), national securities exchanges, registered clearing agencies, plan processors, certain SROs, and SCI ATSes. The 2024 amendments expanded the SCI entity definition and tightened incident reporting, with phased compliance dates extending into 2026.

Applicability test: Reg SCI applies only to SCI entities. Most retail and mid-tier broker-dealers are out of direct SCI scope but may inherit SCI obligations contractually through connectivity to SCI entities.

AI-Agent Implication: When an AI agent operates within an SCI entity's "SCI systems" (trading, clearance, settlement, order routing, market data, regulation, surveillance), the agent inherits the SCI entity's Reg SCI obligations: capacity, integrity, resiliency, availability, security, BCP/DR, SCI event reporting, and five-year recordkeeping.

Key Reg SCI Obligations and Mapped Controls:

Obligation Reg SCI Reference Primary Controls
Policies and procedures (capacity, integrity, resiliency, availability, security) Rule 1001(a) 2.1, 2.3, 2.4, 2.6
BCP / DR (operational capability and 2-hour resumption objective) Rule 1001(a)(2)(v) 2.4, 2.7
SCI event notification (immediate / 24-hour / quarterly) Rule 1002 3.4, 1.21
Annual independent SCI review Rule 1003(b) 3.1, 3.3, 1.7
Industry-wide BCP/DR testing Rule 1004 2.4
Recordkeeping (5 years; first 2 readily accessible) Rule 1005 1.7, 1.9, 2.13

Zone Requirements:

  • Zone 1 (Personal): WSPs at SCI entities should explicitly prohibit personal-productivity agents from operating within or transmitting to/from SCI systems
  • Zone 2 (Team): SCI scoping decision required; agents touching SCI systems inherit BCP coverage and SCI inventory inclusion
  • Zone 3 (Enterprise): Full SCI control set; BCP/DR aligned to the entity's recovery time objective; SCI event runbooks; SCI review participation; five-year retention

Framework Approach: The framework helps support Reg SCI through 12 mapped controls. Reg SCI is entity-specific; legal counsel and the SCI entity's regulatory operations team must confirm scope and classification for each AI-agent surface. See the Reg SCI section in Regulatory Mappings for detailed mapping.

SOX Sections 302/404 — Internal Controls

Overview: Requires public companies to maintain effective internal controls over financial reporting.

Key Requirements:

  • Management certification of internal controls (302)
  • Assessment of internal control effectiveness (404)
  • Audit trail for financial data access
  • Change control for systems affecting financials

Applicable Controls:

Control Requirement Mapping
1.7 Comprehensive Audit Logging Audit trail for all access
1.11 Conditional Access and MFA Access control
1.18 Application-Level RBAC Role-based access
2.3 Change Management Change control procedures
2.8 Segregation of Duties SoD controls
3.3 Compliance Reporting Control effectiveness evidence

Zone Requirements:

  • Zone 1: Not applicable
  • Zone 2: Limited scope if agent touches financial data
  • Zone 3: Full SOX compliance for agents affecting financial reporting

AI System Coverage

SOX does not explicitly address AI or automated systems. AI agents affecting financial reporting are governed implicitly through existing ICFR frameworks. The PCAOB is conducting research to determine whether new standards are needed for AI in audits and financial reporting (July 2024 Spotlight on GenAI).

GLBA Section 501(b) — Safeguards Rule

Overview: Requires financial institutions to protect the security and confidentiality of customer information.

Key Requirements:

  • Administrative, technical, and physical safeguards
  • Risk assessment and management
  • Service provider oversight
  • Incident response procedures

Applicable Controls:

Control Requirement Mapping
1.5 DLP and Sensitivity Labels Data protection
1.11 Conditional Access and MFA Access safeguards
1.15 Encryption Technical safeguards
1.17 Endpoint DLP Endpoint protection
2.7 Vendor Risk Management Service provider oversight
3.4 Incident Reporting Incident response

OCC Bulletin 2026-13 (formerly OCC Bulletin 2011-12) / Fed SR 26-2 (formerly SR 11-7) — Model Risk Management

Overview: Guidance on model risk management for banks using models in decision-making.

Generative AI and agentic AI excluded from scope

The interagency guidance underlying OCC Bulletin 2026-13 / Fed SR 26-2 explicitly states: "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance." In this framework, direct mappings under this section apply to traditional model risk (for example, algorithmic credit scoring or fraud models). GenAI- and agentic-AI-specific controls are included here only as analogous sound-risk-management principles. See Regulatory Mappings for the fuller scope caveat, including the guidance's "most relevant to" institutions over $30 billion in assets.

Key Requirements:

  1. Model validation and testing
  2. Ongoing monitoring and performance tracking
  3. Model governance and documentation
  4. Independent validation

Applicable Controls:

Control Requirement Mapping
2.5 Testing and Validation Model testing
2.6 Model Risk Management Comprehensive MRM framework
2.11 Bias Testing Fairness validation for traditional models; analogous principle for GenAI reviews
2.16 RAG Source Integrity Analogous principle for GenAI grounding-source validation
3.2 Usage Analytics Performance monitoring
3.10 Hallucination Feedback Analogous principle for GenAI output-quality monitoring

Applicability:

  • National banks (OCC)
  • State member banks (Federal Reserve)
  • State non-member banks (FDIC applies interagency guidance)

CFTC Rule 1.31 — Recordkeeping for Swap Dealers

Overview: CFTC Rule 1.31 requires swap dealers and major swap participants to maintain records of transactions, communications, and related activities in a format that is identifiable and searchable.

AI Agent Relevance: Agent interactions in commodities trading contexts — including automated communications, trade-related outputs, and decision logs — may fall within 1.31 record-keeping scope. Organizations subject to CFTC oversight should consider mapping agent governance controls (particularly Controls 1.7, 1.9, and 2.13) to their 1.31 record-keeping requirements. Implementation specifics depend on institutional use of agents in CFTC-regulated activities.

Regulation-Zone Mapping

FINRA Examination Focus by Zone

Zone FINRA Focus Key Controls Examination Depth
Zone 1 None N/A None
Zone 2 Supervisory controls Rule 3110 Moderate
Zone 3 Complete oversight Rules 3110, 4511 Comprehensive

SEC Examination Focus by Zone

Zone SEC Focus Key Controls Examination Depth
Zone 1 None N/A None
Zone 2 Data handling Reg S-P Limited
Zone 3 Complete compliance Rules 17a-3/4, AI disclosure Comprehensive

SOX Internal Controls by Zone

Zone SOX Scope Documentation Testing
Zone 1 None N/A None
Zone 2 Limited Process docs Annual
Zone 3 Full All controls Annual + Continuous

Regulatory Priority by Institution Type

Broker-Dealers (FINRA/SEC)

Priority Controls:

  1. Control 2.12 (Supervision) — FINRA 3110
  2. Control 1.7 (Audit Logging) — FINRA 4511, SEC 17a-4
  3. Control 2.11 (Bias Testing) — FINRA 3110 (supervision of AI tools)
  4. Control 3.3 (Compliance Reporting) — Examination readiness

Banks (OCC/Fed)

Priority Controls:

  1. Control 2.6 (Model Risk Management) — OCC Bulletin 2026-13 (formerly OCC 2011-12), Fed SR 26-2 (formerly SR 11-7)
  2. Control 2.11 (Bias Testing) — Fair lending
  3. Control 1.7 (Audit Logging) — Records requirements
  4. Control 1.5 (DLP) — GLBA 501(b)

Investment Advisers (SEC)

Priority Controls:

  1. Control 2.12 (Supervision) — Reg BI
  2. Control 2.19 (AI Disclosure) — Client communication
  3. Control 1.7 (Audit Logging) — SEC 17a-4
  4. Control 3.1 (Agent Inventory) — Examination readiness

Credit Unions (NCUA)

Priority Controls:

  1. Control 1.5 (DLP) — Part 748 security program
  2. Control 1.7 (Audit Logging) — Records requirements
  3. Control 2.8 (Segregation of Duties) — Internal controls
  4. Control 3.4 (Incident Reporting) — Security program

State-Level Regulations

NYDFS Part 500 (23 NYCRR 500)

The New York Department of Financial Services Cybersecurity Regulation (23 NYCRR Part 500), as amended in 2023, applies to all NYDFS-licensed financial institutions. For AI agent deployments, the most directly applicable requirements are:

  • §500.11 Third-Party Service Provider Security Policy
  • §500.15 Encryption of Nonpublic Information
  • §500.16 Incident Response Plan
  • §500.06 Audit Trail
  • §500.07 Access Privileges and Management
  • §500.12 Multi-Factor Authentication

A complete per-section mapping with applicable framework controls is available in the Regulatory Mappings reference (search for "NYDFS Part 500"). Additionally, the 2024 NYDFS AI Cybersecurity Guidance letter (October 16, 2024) and the dual-signature certification requirement (April 15, 2024) are also reflected in the detailed mapping.

Applicability: Banks, insurance companies, mortgage providers, money transmitters, and other entities licensed by NYDFS.

Other State-Level Regulations (For Awareness)

The following state regulations may apply but are beyond the primary scope of this framework:

Regulation Jurisdiction Focus Framework Relevance
CCPA/CPRA California Consumer privacy Controls 1.5, 1.9, 4.4
Colorado AI Act Colorado High-risk AI Controls 2.6, 2.11, 2.19

Organizations should conduct separate analysis for state-specific requirements.

Examination Readiness Checklist

Pre-Examination Preparation

  • Agent inventory current and complete (Control 3.1)
  • Audit logs accessible for required retention period (Control 1.7)
  • Supervisory procedures documented (Control 2.12)
  • Change records available (Control 2.3)
  • Incident reports filed (Control 3.4)
  • Training records current (Control 2.14)

Common Examiner Requests

Request Control Documentation
List of all AI agents 3.1 Agent inventory
Agent approval records 2.12 Governance committee minutes
Audit logs for specific agent 1.7 Purview Audit export
Supervisory procedures 2.12 Written procedures document
Incident history 3.4 Incident reports
Testing documentation 2.5 Test results and validation

Framework Coverage Summary

Regulation Controls Mapped Framework Coverage
FINRA 4511 9 controls Books and records
FINRA 3110 8 controls Supervision
FINRA 3110/2111 11 controls AI supervision
SEC 17a-3/4 8 controls Recordkeeping
SOX 302/404 6 controls Internal controls
GLBA 501(b) 6 controls Safeguards
OCC Bulletin 2026-13 (formerly OCC 2011-12) / Fed SR 26-2 (formerly SR 11-7) 6 controls Model risk
CFTC 1.31 3 controls Recordkeeping

Total: 78 controls across 4 pillars providing mapped coverage to primary US financial regulations.

Note

Coverage indicates which framework controls address aspects of each regulation. Actual compliance requires implementation, validation, and ongoing maintenance.

Microsoft CAPE alignment cross-reference

FSI-AgentGov has included a Microsoft CAPE alignment layer since v1.5.0 (current at v1.6.2) that maps Microsoft's six Frontier Transformation Patterns onto FSI zones, controls, and regulatory exposure. This alignment does not modify the underlying 78-control framework, which remains the definitive governance structure for US financial services AI deployments. CAPE is a Microsoft business-strategy framework designed to accelerate enterprise AI transformation, not a regulatory framework. Any deployment using CAPE Patterns 4–6 sits inside the FSI regulatory perimeter and triggers the same regulatory obligations as any other AI deployment subject to FINRA, SEC, OCC, Fed, GLBA, SOX, and state regulatory oversight.

The Microsoft CAPE crosswalk provides the canonical mapping between CAPE vocabulary and FSI governance requirements. For each CAPE pattern, the crosswalk identifies:

  • Primary regulations triggered by the pattern
  • Default FSI zone classification
  • Mandatory FSI-AgentGov controls required to support compliance with regulatory obligations
  • Autonomy cap (the documented human-in-the-loop boundaries required for examiner-defensible posture)
  • Examiner red flags (typical regulatory examination questions and evidence requests)
  • CAPE language to reframe (industry-agnostic CAPE descriptors that require FSI-specific translation for regulatory contexts)

Pattern × Zone default mapping:

Pattern 6 Autonomy Guardrail

Fully autonomous customer-impacting Pattern 6 deployments are not currently supported in Zone 3 without documented regulator pre-approval. This framework position reflects the current regulatory environment where fully autonomous, self-optimizing AI systems that directly affect customers have not been explicitly addressed by US financial services regulators. Organizations considering such deployments should obtain documented guidance from their primary regulator(s) before production use.

Cross-references:

FSI Agent Governance Framework v1.6.2 - May 2026