Your compliance team needs a governance framework for AI agents, but you can't govern what you can't see. 80% of organizations have already encountered agentic AI risks, yet only 21% have mature governance programs. This template provides a working policy structure you can implement this week.
Purpose of the Template
This policy template sets discovery, approval, and monitoring requirements for AI agents created on platforms like Salesforce Agentforce, Microsoft Copilot Studio, and custom LLM integrations. It addresses the unique risks of AI agents, such as persistent permissions beyond employee tenure, autonomous access to sensitive systems, and decision-making without human oversight.
Use this template to:
- Document security requirements for AI agent creation
- Define approval workflows that support innovation
- Establish monitoring procedures for agents
- Create audit trails that satisfy SOC 2 Type II access control requirements
Prerequisites
Before implementing this policy, ensure you have:
Discovery capability: You can't enforce what you haven't found. Deploy either OAuth monitoring (tracks authorization grants to AI platforms), CASB integration (monitors API activity to agent platforms), or SSO log analysis (identifies authentication to agent creation tools). Each method has tradeoffs; OAuth monitoring catches agents at creation time but misses API key-based tools, while CASB sees runtime behavior but requires proxy deployment.
Stakeholder buy-in: This policy will reveal agents your business teams consider essential. Get your VP of Engineering and Chief Revenue Officer aligned on the "why" before you tell their teams "no."
Classification schema: You need a data classification policy that defines sensitive data in your environment. For PCI DSS, this means cardholder data environments (CDE). For SOC 2 Type II, it's customer data. For healthcare, it's PHI under HIPAA.
The Policy Template
# AI Agent Governance Policy
Version 1.0 | Effective Date: [DATE]
## 1. Scope
This policy applies to all AI agents that:
- Authenticate using corporate credentials
- Access systems containing [CLASSIFIED DATA TYPES]
- Make automated decisions affecting [BUSINESS PROCESSES]
- Integrate with [CRITICAL SYSTEMS LIST]
## 2. Discovery Requirements
### 2.1 Continuous Monitoring
Security Operations will maintain monitoring for:
- OAuth grants to AI agent platforms
- API key generation for LLM services
- SSO authentication to agent creation tools
### 2.2 Quarterly Inventory
Every 90 days, Security will:
- Export a complete list of authorized agents
- Identify agents without documented business owners
- Flag agents with access exceeding their documented purpose
## 3. Creation Approval Workflow
### 3.1 Pre-Approval Required
Before creating an AI agent, requestor must submit:
- Business justification (what problem does this solve?)
- Data access requirements (what systems will it touch?)
- Proposed permissions scope (read-only vs. write access)
- Retention period (how long will this agent run?)
### 3.2 Security Review Triggers
Automatic security review required if agent will:
- Access production databases
- Process [SENSITIVE DATA CLASSIFICATION]
- Execute privileged operations
- Integrate with external services
### 3.3 Approval Authority
- Low-risk agents (read-only, non-sensitive): Team lead approval
- Medium-risk agents (write access, internal data): Director + Security
- High-risk agents (privileged access, sensitive data): VP + CISO
## 4. Operational Controls
### 4.1 Least Privilege
Agents receive minimum permissions required for documented function. No "admin" or "service account" level access without CISO exception.
### 4.2 Credential Management
- Agents use dedicated service accounts (not personal credentials)
- API keys rotate every [90 DAYS]
- OAuth tokens expire after [TIMEFRAME] without activity
### 4.3 Activity Logging
All agent actions must generate audit logs including:
- Timestamp of action
- Data accessed or modified
- Decision rationale (where applicable)
- Human approver (for consequential actions)
## 5. Decommissioning
### 5.1 Automatic Triggers
Agents are automatically disabled when:
- Business owner leaves organization
- 90 days pass without activity
- Security review identifies policy violation
### 5.2 Manual Review
Quarterly review will identify agents for retirement based on:
- Duplicate functionality
- Obsolete business need
- Excessive permissions for current use
## 6. Compliance Mapping
- SOC 2 Type II CC6.1: Access controls and authorization
- [ISO 27001](https://www.iso.org/standard/27001): User access management
- PCI DSS v4.0.1 Requirement 7: Restrict access to system components
- [NIST CSF](https://www.nist.gov/cyberframework) v2.0 PR.AC-4: Access permissions managed
## 7. Enforcement
Violations result in:
- First offense: Agent disabled, owner training required
- Second offense: Escalation to management, access review
- Third offense: Revocation of agent creation privileges
Customizing the Template
Replace bracketed placeholders with your specifics:
[DATE]: Your implementation date[CLASSIFIED DATA TYPES]: "PII, payment card data, proprietary source code" or whatever matches your classification policy[BUSINESS PROCESSES]: "customer communications, financial transactions, code deployment"[CRITICAL SYSTEMS LIST]: Your actual system inventory[SENSITIVE DATA CLASSIFICATION]: Reference your existing data classification levels[90 DAYS]/[TIMEFRAME]: Match your existing credential rotation policies
Adjust approval thresholds based on your org structure. If you're a 500-person company, you probably don't need VP approval for medium-risk agents. Director + Security might suffice.
Add platform-specific sections if you know your teams are heavy users of particular tools. For Salesforce Agentforce deployments, specify which Salesforce objects require security review. For Microsoft Copilot Studio, define which Power Platform connectors are pre-approved.
Tune the enforcement section to your culture. Some organizations prefer coaching over punishment for first violations. Others need harder lines because they're in regulated industries.
Validation Steps
Week 1: Run your discovery tooling and generate the initial inventory. You're looking for two numbers: total agents found, and agents without documented owners. If that second number is over 30%, you've got a shadow AI problem that justifies this policy.
Week 2: Present the draft policy to three stakeholders: your most active agent creator (usually in Sales or Marketing), your most security-conscious engineering lead, and someone from Legal or Compliance. Their feedback will tell you if your approval workflow is realistic.
Week 3: Pilot the approval process with one team. Track how long approvals take and where requests get stuck. If low-risk approvals take more than 24 hours, your workflow is too heavy.
Month 2: Review your first month of agent decommissioning data. You should see some agents disabled due to inactivity or ownership changes. If you're not retiring any agents, your criteria are too loose or your monitoring isn't working.
Quarter 1: Run your first quarterly inventory and compare it to Week 1 baseline. You want to see the percentage of undocumented agents dropping, not rising. If it's rising, your discovery is working but your governance isn't being followed.
The goal isn't zero shadow AI agents. It's knowing what's running, who owns it, and whether the risk matches the business value. This template gives you the structure to make that judgment systematically instead of discovering problems during your next audit.
