Governance Fundamentals

Core concepts and principles for AI agent governance in financial services.

Adaptive Governance Philosophy

This framework is built on the principle of adaptive governance: controls that scale proportionally with risk rather than applying uniform restrictions across all agent use cases. A low-risk personal productivity agent is not the same as an agent connected to a core business system — governance should reflect that difference.

Adaptive governance rests on three operational pillars:

Classify risk clearly — The three governance zones (Personal, Team, Enterprise) provide graduated risk classification that determines the level of oversight, approval, and audit requirements for each agent.
Enforce through the platform — Governance is most effective when it is inherent to the platform, not layered on through policy documents alone. Managed environments, environment groups, DLP policies, and environment routing enforce boundaries automatically — reducing reliance on manual processes and individual compliance.
Create promotion paths — Rather than restricting all agent development, the framework provides a supported path from personal experimentation (Zone 1) through team collaboration (Zone 2) to enterprise production (Zone 3). Sharing limits paired with zone progression criteria create a clear on-ramp: agents can scale when they meet the governance requirements of the next zone.

Why This Matters

Governance models that only restrict — without providing supported paths for innovation — tend to drive shadow IT. By combining platform-enforced controls with clear promotion criteria, organizations can let teams experiment freely in Zone 1 while requiring deliberate promotion, review, and accountability when agents are ready to reach broader audiences.

Framework Overview

The FSI Agent Governance Framework provides complete guidance for deploying, governing, and managing Microsoft 365 agents (Copilot Studio, Agent Builder, and related AI services) in regulated financial services environments.

Version: 1.2.53 (March 2026) Target Audience: US Financial Services Organizations Regulatory Focus: FINRA, SEC, SOX, GLBA, OCC, Federal Reserve, FDIC, NCUA, CFTC

Warning

This framework is provided for informational purposes only and does not constitute legal, regulatory, or compliance advice. See Disclaimer for full details.

Scope and Assumptions

What This Framework Covers

This framework provides governance guidance for:

Copilot Studio agents
Agent Builder agents
Power Platform environments hosting agents
SharePoint as a knowledge source for agents

Agent Types In Scope

Agent Type	In Scope?	Governance Location
Copilot Studio custom agents	Primary focus	Power Platform Admin Center
Agent Builder agents (M365)	Primary focus	M365 Admin Center + PPAC
Microsoft-built agents (Researcher, Analyst, Facilitator)	Partial	M365 Admin Center (limited controls)
Third-party/partner agents	Varies by integration	M365 Admin Center
Agent 365 control plane	GA (May 1, 2026)	M365 Admin Center

Microsoft Agent 365 — Generally Available May 1, 2026

Microsoft Agent 365 is the centralized control plane for agent governance, announced at Ignite 2025 and generally available as of May 1, 2026. It provides centralized agent registry, Entra Agent ID, lifecycle management, and observability dashboards. At GA, Agent 365 transitions from preview per-agent-instance licensing to per-user licensing — agents acting on behalf of a licensed user are covered under that user's Agent 365 or Microsoft 365 E7 license. See Agent 365 overview for current details.

Entra Agent ID (Public Preview)

Entra Agent ID provides first-class agent identities in Microsoft Entra, enabling Conditional Access, Identity Protection, and lifecycle management for AI agents. Currently in Public Preview since November 2025. See Agent Identity Architecture for implementation guidance and FSI readiness assessment.

Microsoft-Built Agent Applicability

Microsoft-built agents (Researcher, Analyst, Facilitator) have limited governance options. Control applicability:

Control Category	Applicability	Notes
1.2 Agent Registry	Partial	Visible in M365 Admin Center only
1.5 DLP/Sensitivity Labels	Full	Tenant-wide policies apply
1.7 Audit Logging	Full	Logged via OfficeActivity
2.1 Managed Environments	N/A	Not in Power Platform
2.3 Change Management	N/A	Managed by Microsoft
3.1 Agent Inventory	Full	Visible in M365 Admin Center
3.2 Usage Analytics	Full	Copilot usage reports
4.1-4.9 SharePoint	Full	Grounding data governance applies

Zone Classification: Based on data access - Zone 1 (personal M365), Zone 2 (team sites), Zone 3 (regulated content).

Limited Customization

Microsoft-built agents cannot be configured. Governance focuses on data access control, usage monitoring, and output supervision.

What This Framework Does NOT Cover

Non-US regulations (EU AI Act, GDPR, DORA, MiFID II are out of scope)
Non-Microsoft AI platforms (OpenAI direct, AWS Bedrock, Google Vertex AI, etc.)
Custom ML model development, training, or validation
Quantitative model risk management (requires dedicated MRM programs)
State privacy laws (CCPA/CPRA require separate analysis)
Third-party AI integrations outside Microsoft 365 ecosystem

Key Assumptions

Assumption	Rationale
Microsoft 365 E3/E5	Required for Copilot Studio, Purview, and advanced governance features
Microsoft Entra ID	Identity and access management foundation
Microsoft Purview	Compliance and data governance capabilities
Power Platform licensing	Required for environment management and DLP policies
Foundational IT controls	Network security, endpoint protection, backup/recovery assumed in place
Identity and access hygiene	Existing permissions are correctly scoped before agent deployment

Agents Expose Existing Permission Problems

Agents generally operate under the calling user's identity and permissions — they do not create new access rights. This means agents do not introduce permission problems; they surface existing over-permissioning faster by making previously obscure but technically accessible content more discoverable. Organizations should audit and remediate identity and access management (IAM) gaps — particularly oversharing in SharePoint, broad security group memberships, and excessive delegated permissions — before deploying agents broadly. See DSPM for AI and M365 Copilot Data Governance for assessment tools and remediation guidance.

Four Governance Pillars

The framework organizes 78 controls across four pillars:

Pillar	Controls	Focus	Examples
1. Security	29	Protect data and systems	DLP, Audit, Encryption, MFA, eDiscovery, Network Isolation, Information Barriers
2. Management	26	Govern lifecycle and risk	Change Control, Testing, Model Risk, Multi-Agent Orchestration, HITL Framework
3. Reporting	14	Monitor and track	Inventory, Usage, PPAC, Sentinel, Hallucination Feedback
4. SharePoint	9	SharePoint-specific controls	Access, Retention, External Sharing, Grounding Scope, Copilot Data Governance

Note: Pillar 4 specializes Pillars 1-3 for SharePoint as an agent knowledge source. Controls address SharePoint-specific implementation of data protection, access governance, and content management.

Three Governance Zones

Agents are classified into zones based on risk level:

Zone	Risk Level	Data Access	Approval	Audit Retention
Zone 1: Personal	Low	M365 Graph only	Self-service	30 days
Zone 2: Team	Medium	Internal data	Manager approval	1 year
Zone 3: Enterprise	High	Regulated data	Governance Committee	10 years

See Zones and Tiers for detailed zone definitions and selection criteria.

Governance Triangle

Effective agent governance operates through three interconnected layers:

Policy Layer (Technical Controls)

Automated guardrails that enforce governance without manual intervention:

Component	Purpose	Implementation
Environment Groups	Consistent policy across environments	PPAC environment groups
Group Rules	Connector, sharing, channel controls	Environment group rules
DLP Policies	Data boundary enforcement	Data policies
Environment Routing	Automatic maker placement	Default environment routing

Process Layer (Operational Workflows)

How governance decisions are made and executed:

Agent Lifecycle Management — Creation, testing, deployment, monitoring, retirement
Approval Workflows — Zone-appropriate authorization paths
Change Control — Controlled promotion between environments
Incident Response — Detection, investigation, remediation procedures
Compliance Reviews — Scheduled verification of control effectiveness

People Layer (Organizational Structure)

Accountability and human oversight:

Role	Governance Function	Zone Focus
AI Governance Lead	Framework ownership, policy decisions	All zones
Power Platform Admin	Technical implementation, environment management	Zones 2-3
Compliance Officer	Regulatory alignment, audit coordination	Zones 2-3
Business Owner	Agent sponsorship, use case validation	Per agent
Security / CISO	Threat monitoring, incident response	Zone 3

How the Layers Interact

Policy enables Process — Technical controls automate workflow enforcement
Process guides People — Defined procedures clarify responsibilities
People inform Policy — Human judgment shapes control configuration

Trust by Design with Built-In Verification

The Governance Triangle reflects a trust-by-design approach: strong proactive controls (Policy Layer) set boundaries that allow agents to operate with appropriate autonomy, while reactive controls (Process Layer) verify those boundaries hold through monitoring, diagnostics, and audit trails.

This mirrors how organizations manage other operational risks. Consider expense approvals: humans unintentionally approve incorrect items regularly, and that risk is managed through audits, compensating controls, and limits on blast radius — not by eliminating approvals entirely. Agent governance follows the same principle: know what happened, understand why it happened, and contain the impact when outcomes differ from expectations.

Proactive controls — DLP policies, sharing limits, environment groups, publishing restrictions
Reactive controls — Audit logging, Sentinel integration, compliance reporting, incident response
Compensating controls — Hallucination feedback loops, access reviews, orphaned agent detection

FSI Note

In regulated environments, all three layers must be documented and auditable. Examiners expect evidence of policy configuration, process execution, and role assignment.

Governance Maturity Levels

Each control supports three implementation levels:

Level	Name	Description	Typical Use
Level 1	Baseline	Minimum required implementation	Initial deployment, Zone 1
Level 2-3	Recommended	Best practice implementation	Zone 2, most production agents
Level 4	Regulated/High-Risk	Comprehensive controls	Zone 3, customer-facing agents

Maturity Assessment

For compliance tracking, use the 5-point maturity scale: Level 0 (Not implemented), Level 1 (Baseline), Level 2-3 (Recommended), Level 4 (Regulated). See FAQ for percentages.

Control Implementation Approach

Assess — Current state vs. required level
Implement — Follow playbook guidance
Verify — Use verification procedures
Document — Record evidence for audit
Review — Schedule recurring reviews

Integration with Existing Governance

This framework is designed to complement, not replace existing enterprise governance programs:

Integrate controls with your existing IT risk management framework
Align with enterprise information security policies
Coordinate with records retention and eDiscovery requirements
Map to your organization's internal audit program

Note

Organizations should validate all controls against their specific regulatory obligations and existing policy frameworks.

Framework Documentation Structure

Three-Layer Architecture

Layer	Audience	Update Frequency	Content
Framework (this layer)	Executives, compliance, governance	1-2x per year	Principles, zones, regulatory context
Control Catalog	Compliance officers, architects	Quarterly	Control objectives, requirements
Playbooks	Platform teams, operations	Continuous	Portal steps, scripts, screenshots

For Governance Questions:

Framework Overview — This layer
Zones and Tiers — Classification requirements
Operating Model — Roles and responsibilities

For Control Requirements:

Control Catalog — All 78 controls
Regulatory Framework — Regulation-to-control mappings

For Implementation:

Playbooks — Step-by-step procedures
Adoption Roadmap — Phased implementation

Quick Reference

Getting Started

Read Executive Summary (10 minutes)
Review Zones and Tiers to classify agents (15 minutes)
Check Regulatory Framework for applicable regulations (10 minutes)

For Implementation

Use the Adoption Roadmap for phased approach
Reference Control Catalog for requirements
Follow Playbooks for step-by-step procedures
Schedule reviews per Governance Cadence

For Governance

Use the Operating Model for roles and responsibilities
Establish governance committee per Zones and Tiers
Schedule recurring compliance reviews
Track incidents per playbook procedures