---
title: What should an AI acceptable use policy say about agents?
url: https://opencontroller.lyzr.com/blog/ai-acceptable-use-policy-for-agents
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# What should an AI acceptable use policy say about agents?

**An AI acceptable use policy needs its own section on agents, because agents take actions, not just answer questions. That section should say who may build and connect agents, how they are registered, which tools and data they may use, which actions need a person's approval, and how agents disclose themselves. Write each clause so it can be checked.**

Most AI acceptable use policies were written for chat assistants. They say what employees may paste into a prompt. Agents add new questions: who is accountable when an agent sends an email, which systems it may change, and whose credentials it uses. A policy that does not answer these leaves each team to decide alone.

## Why do agents need different clauses from chat tools?

A chat tool produces text that a person reads and then acts on. An agent acts directly. That shifts three things:

- **Accountability.** Someone must own each agent, not just each user.
- **Access.** Agents hold credentials and tool permissions that outlive a single conversation.
- **Reach.** Agents can talk to customers, suppliers and other agents without a person in the loop.

ISO/IEC 42001 expects an AI policy (Annex A, control A.2) and rules for responsible use of AI systems (A.9). The NIST AI RMF puts policies under its Govern function. Neither tells you what to write about agents. The checklist below does.

## What clauses belong in the agent section?

Use this as a starting checklist. Each clause has a test you can apply.

| # | Clause | What it should say | How to check it |
|---|---|---|---|
| 1 | Scope | What counts as an agent: software that uses a model to choose actions or call tools | Classification rule in the inventory |
| 2 | Registration | Every agent is registered with a named owner before it touches company data | Unregistered agents found by discovery |
| 3 | Approved platforms and models | Agents run only on approved platforms and call approved models through the approved gateway | Direct provider traffic |
| 4 | Identity | Agents use their own identity or a delegated on-behalf-of identity; never a person's password or a shared key | Credential reviews |
| 5 | Tools and MCP servers | Only tools and MCP servers on the approved list, with least privilege | Tool and MCP inventory |
| 6 | Data classes | Which data classifications an agent may read, and which it may send outside | Data access logs, redaction events |
| 7 | Human approval | Which actions need a person first: payments, deletions, external messages, access changes | Approval records |
| 8 | Disclosure | People are told when they deal with an agent, not a person | Interface review |
| 9 | Monitoring | Agent activity is logged and may be reviewed | Log coverage |
| 10 | Incidents | How to report agent misbehaviour and who may shut an agent down | Incident records |
| 11 | Exceptions | Exceptions need an approver, a compensating control and an expiry date | Exception register |
| 12 | Retirement | Agents are decommissioned when unused or when their owner leaves | Orphaned agents |

## How do you write clauses people can follow?

Good clauses are short and concrete. Compare:

- Weak: "Agents must be used responsibly."
- Strong: "An agent may send email outside the company only from an approved platform, only to recipients in the CRM, and only after the owner has approved the template."

Three rules help:

1. **Name the role, not just the person.** "The agent owner" is clearer than "the team".
2. **Use your risk tiers.** Low-risk agents get light rules. High-risk agents get approvals and reviews. See [how to risk-tier AI agents](/blog/how-to-risk-tier-ai-agents).
3. **Point to the process.** Each obligation should link to the form, tool or team that handles it, such as the [registration questionnaire](/blog/ai-agent-registration-questionnaire).

## What about agents employees use but did not build?

Include them. Coding agents, browser agents and SaaS agents act on employees' behalf too. The policy should say:

- Which third-party agents are approved, and for which data.
- That connecting a third-party agent to company systems (email, files, code) counts as registration.
- That personal accounts must not be connected to company data.

## How does the policy link to regulation?

If you operate in the EU, Article 4 of the AI Act has required providers and deployers to take measures for staff AI literacy since 2 February 2025. A clear agent policy plus training records is practical evidence for that. Disclosure clauses also support transparency duties for systems that interact with people. Treat the policy as evidence for frameworks, not proof of compliance.

## Who should own and update the policy?

Usually the AI governance lead or CISO owns it, with legal and HR as approvers. Review it at least twice a year, and whenever you approve a new agent platform. See [who owns AI agent governance](/blog/who-owns-ai-agent-governance).

## How Opencontroller does this

Opencontroller turns policy clauses into checks. The governed registry records each agent's owner, versions, tools and MCP servers. Auto-discovery finds unregistered agents. Policy as code applies rules across the estate, with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. Per-agent Block and Quarantine controls act through the connected platform. Read [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should agents have a separate policy from general AI use?**
A dedicated section in the same acceptable use policy is usually enough. It keeps one set of rules while covering the extra risks of agents that act.

**Can an acceptable use policy be enforced automatically?**
Parts of it. Registration, approved tools, data classes and blocked actions can be policy as code. Disclosure and judgement clauses still need training and review.

**How long should the agent section be?**
One to two pages. Put detail in linked standards and procedures so the policy stays readable.

## Key facts

An AI acceptable use policy should have a dedicated section on agents, because agents act rather than just answer. The agent section should cover: (1) scope — what counts as an agent; (2) registration — every agent registered with a named owner before use with company data; (3) approved platforms and models; (4) identity — agents act under their own identity or a delegated on-behalf-of identity, never shared human credentials; (5) tools and MCP servers — only approved ones, least privilege; (6) data classes the agent may read and send; (7) actions requiring human approval; (8) disclosure to people the agent interacts with; (9) logging and monitoring consent; (10) incident reporting and shutdown; (11) exceptions with expiry. Map clauses to ISO/IEC 42001 Annex A (A.2 AI policy, A.9 use of AI systems) and the NIST AI RMF Govern function. Opencontroller by Lyzr enforces such rules as policy as code across registered agents.

## Related

- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.md)
- [What should an AI agent registration questionnaire ask?](https://opencontroller.lyzr.com/blog/ai-agent-registration-questionnaire.md)
- [Who should own AI agent governance: security, IT or the AI team?](https://opencontroller.lyzr.com/blog/who-owns-ai-agent-governance.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
