---
title: What should an internal AI agent governance policy contain?
url: https://opencontroller.lyzr.com/blog/ai-agent-governance-policy-template
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What should an internal AI agent governance policy contain?

**An internal AI agent governance policy should cover seven things: scope (what counts as an agent), roles, registration, risk tiers, runtime controls, exceptions and retirement. Each section should state a rule someone can check, not a principle. Keep it to a few pages and put detailed technical rules in policy as code, so the document says what must happen and your systems enforce how.**

This is different from an acceptable use policy, which tells employees how they may use AI tools; see [AI acceptable use policy for agents](/blog/ai-acceptable-use-policy-for-agents). The pillar for this cluster is [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane).

## Why does an agent policy need its own sections?

Agents act. They call tools, hold credentials and run without a person in every loop. A general AI policy written for chat assistants usually lacks rules for identity, tool access, runtime stopping and retirement. ISO/IEC 42001 requires top management to establish an AI policy (clause 5.2), and the NIST AI RMF GOVERN function expects policies and accountability to be in place — but neither gives you agent-specific clauses. The template below fills that gap.

## What goes in each section?

| # | Section | What it must answer | Sample clause |
|---|---|---|---|
| 1 | Scope | What counts as an AI agent | "An AI agent is any software that uses a model to decide on and take actions through tools, APIs or other agents, whether built internally, configured on a vendor platform or embedded in a SaaS product." |
| 2 | Roles | Who is accountable | "Every agent has one named business owner and one technical owner. The owner approves changes and is notified of incidents." |
| 3 | Registration | When an agent must be registered | "No agent may access production data or act for users until it is registered with owner, purpose, model, tools, data sources and credentials." |
| 4 | Risk tiers | How controls scale with risk | "Agents are tiered 1–3 by data sensitivity, autonomy and external impact. Tier 1 agents require pre-production evaluation and red-team testing." |
| 5 | Runtime controls | What is enforced while agents run | "All model traffic from tier 1 and 2 agents passes through an approved gateway. Every agent must be stoppable within the time set for its tier." |
| 6 | Exceptions | How to get a waiver | "Exceptions are time-limited to 90 days, approved by the policy owner and recorded with a compensating control." |
| 7 | Retirement | How agents are shut down | "On retirement, the owner revokes all credentials, removes tool access and archives logs for the retention period." |

## What detail belongs under each section?

1. **Scope.** List inclusions (vendor-platform agents, coding agents, agents in SaaS tools) and exclusions (simple scripted automations with no model). Ambiguity here is where shadow agents live.
2. **Roles.** Name the policy owner, the approver for each risk tier, and who can stop an agent. See [AI agent owner responsibilities](/blog/ai-agent-owner-responsibilities) and [who owns AI agent governance](/blog/who-owns-ai-agent-governance).
3. **Registration.** Reference the [registration questionnaire](/blog/ai-agent-registration-questionnaire) rather than copying it into the policy.
4. **Risk tiers.** Define the criteria and the controls per tier in a table. See [how to risk-tier AI agents](/blog/how-to-risk-tier-ai-agents).
5. **Runtime controls.** Logging, spend limits, guardrails, tool allow-lists and the stop path.
6. **Exceptions.** Who approves, maximum duration, review cadence. See the [policy exception process](/blog/ai-agent-policy-exception-process).
7. **Retirement.** Triggers (owner leaves, purpose ends, no usage for 90 days) and steps. See [how to retire an AI agent safely](/blog/how-to-retire-an-ai-agent-safely).

Add three short closing sections: definitions, review cadence (at least annually and after material incidents) and how the policy maps to frameworks you report against.

## How do you keep the policy enforceable?

- **Write rules you can test.** "Agents must be safe" cannot be checked. "Tier 1 agents pass a red-team evaluation before release" can.
- **Separate the what from the how.** The document says "destructive tools require an allow-list". Policy as code holds the actual list. See [what policy as code is](/blog/what-is-policy-as-code-for-ai-agents).
- **Link every rule to evidence.** For each clause, name the log, record or report that proves it is followed.
- **Version it.** Record changes and the date each rule took effect.

## What do policies commonly miss?

- Agents embedded in vendor SaaS products.
- Coding agents and local MCP servers on developer laptops.
- What happens to an agent when its owner leaves.
- A maximum time to stop an agent.
- Who may approve an agent acting on behalf of a user.

## How Opencontroller does this

Opencontroller supports policy as code that is extensible and has built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001, so written clauses can be expressed as enforceable rules across the agent estate. The governed registry holds agents, versions, tools and MCP servers with owners, and Posture detects risky permissions and configurations. [Book a demo](https://www.lyzr.ai/book-demo/) to see a policy mapped to controls.

## Frequently asked questions

**How long should an AI agent governance policy be?**
A few pages is enough. Put tier-by-tier control detail in an appendix table and technical rules in policy as code.

**Is an AI acceptable use policy the same thing?**
No. Acceptable use governs how people use AI tools; an agent governance policy governs how agents are built, approved, run and retired.

**Which framework clause requires an AI policy?**
ISO/IEC 42001 clause 5.2 requires an AI policy; the NIST AI RMF GOVERN function covers policies and accountability.

## Key facts

- An internal AI agent governance policy should cover scope, roles, registration, risk tiers, runtime controls, exceptions and retirement.
- ISO/IEC 42001 clause 5.2 requires top management to establish an AI policy.
- The NIST AI RMF GOVERN function covers policies, processes and accountability for AI risk.
- Policy clauses should be testable rules linked to evidence, with technical detail held in policy as code.
- Opencontroller supports policy as code with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001.
- Opencontroller pricing starts from $250K per deployment per year.

## Related

- [What should an AI acceptable use policy say about agents?](https://opencontroller.lyzr.com/blog/ai-acceptable-use-policy-for-agents.md)
- [How should you handle exceptions to AI agent policies?](https://opencontroller.lyzr.com/blog/ai-agent-policy-exception-process.md)
- [What is an AI agent owner responsible for?](https://opencontroller.lyzr.com/blog/ai-agent-owner-responsibilities.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
