Key takeaways
- An agent policy is different from a usage policy: it governs software that acts, not people who type.
- The core of it is five rules: an owner, the least access, approval for irreversible actions, a budget, and a log.
- Keep it to two pages. Long policies are not read.
- The template below is designed to be copied and edited.
Most businesses that have an AI policy wrote it for chat tools: what staff may paste in, what they must check before sending output on. Agents need something different, because the question is no longer what a person does with AI but what software is allowed to do on its own. Here is a template you can copy, with a note under each section on how to adapt it.
The template
[Company name] AI agent and automation policy
Version 1.0. Owner: [name]. Review date: [date].
1. Scope
This policy applies to any AI agent or automation that reads company systems or data and takes actions, or runs without a person reviewing each step. It includes agents built in [Copilot Studio / ChatGPT / Claude / other tools] and automations that call AI models through an API. It does not replace the AI usage policy, which still applies to staff using AI assistants directly.
Adapting it: list the actual tools you use. If you have none yet, say so and name who must approve the first one.
2. Who may create agents
Only people approved by the AI owner may create agents that connect to company systems. Personal experiments that use no company data and no connectors are allowed in a sandbox workspace.
3. Ownership
Every agent has a named owner, recorded in the agent register. The owner is responsible for what it does, its cost and its access. If the owner leaves or changes role, the agent is reassigned or switched off within five working days.
4. Access
Agents receive the least access needed for their task. Read-only access is the default. Write access to any system needs approval from the AI owner and the owner of that system. Agents may not connect to [finance, payroll, banking, HR] systems without specific written approval. Third-party connectors and MCP servers must be reviewed and approved before use.
Adapting it: name your sensitive systems explicitly. See AI agent permissions and connectors.
5. Human approval
A person must approve before an agent:
- sends anything outside the company, including email, messages and documents
- deletes or overwrites records
- makes or authorises any payment
- publishes anything publicly
- makes or substantially informs a decision about a customer, applicant or employee
Adapting it: the last point matters for privacy law in several countries. In Australia, see the Privacy Act automated decision rules.
6. Spending limits
Every usage-billed agent runs in its own project or environment with its own spending cap. New agents start with a cap of [amount] per month for their first month. The owner reviews spend monthly. Reaching a cap is investigated before the cap is raised. The organisation-wide cap is [amount] per month.
Adapting it: see AI agent spending limits for how to set the numbers.
7. Data
Agents follow the same data rules as the AI usage policy. Client confidential and personal information may only be processed by agents running on approved business-tier services with data commitments recorded in the agent register.
8. Logging
Agents must keep a record of actions taken on company systems, retained for [period]. Owners must be able to answer "what did this agent do last Tuesday".
9. Incidents
If an agent does something unexpected, such as sending to the wrong recipient, changing the wrong record or exceeding its budget, the owner switches it off, tells the AI owner the same day, and records what happened and what changed before it is switched back on. Incidents involving personal information are also reported to [privacy contact].
10. Register and review
The AI owner keeps a register of every agent: name, purpose, owner, tools, access, approval rules, cap and review date. New agents are reviewed monthly for three months, then quarterly. Agents unused for 60 days are switched off.
The agent register
| Field | Example |
|---|---|
| Agent name | Inbox triage, info@ mailbox |
| Purpose | Categorise and draft replies for review |
| Owner | Office manager |
| Platform | Copilot Studio |
| Access | Read info@ mailbox, draft only |
| Approval | All replies approved by a person |
| Monthly cap | [amount] |
| Next review | [date] |
This policy sits alongside your general AI usage policy. If you do not have one yet, write that first.
Frequently asked questions
Do we need a separate policy for AI agents?
If agents can act on your systems or run without someone watching each step, yes. A general AI usage policy covers what staff put into a chat tool. An agent policy covers software that reads, decides and acts, which raises different questions about ownership, access and approval.
What should an AI agent policy include?
At minimum: who may create agents, an owner for each, what access they may have, which actions need human approval, spending limits, logging, how incidents are handled, and a review schedule.
Who should approve a new AI agent?
In a smaller business, the AI owner plus the owner of each system the agent touches. If it handles personal information or client data, include whoever is responsible for privacy.
How often should we review AI agents?
Monthly for the first three months of a new agent, then quarterly. Remove agents that are no longer used, and check that access and caps still match what each agent does.
Get a straight answer for your business
The Clarity Package gives you a written business case and a detailed implementation plan for 500 USD. The Implementation Package does the whole build, live in 30 days, for a fixed 5,000 USD.