An AI agent policy sets the organisation's rules for agents that act: which uses are allowed, who approves an agent, how its authority is set and proven, which actions always need a person, what is recorded, and when an agent is suspended. The template below lists the sections, explains each, and fills in an example for a mid-sized lender.
What should an AI agent policy cover that an AI policy does not?
A general AI policy is usually about models and content: approved tools, data handling, bias, transparency. An agent policy adds the questions that only arise when software takes actions. How far may it act? Who decides? What happens when it is wrong? How do you stop it?
The answers turn on authority, not technology. A good agent policy says almost nothing about which model is used and a great deal about bounds, approvals and evidence. That is the approach in agent governance is a bound, not a model.
When should you use this template?
Write the policy before the first agent goes beyond a pilot, and review it at least annually. Your staff-facing rules sit beneath it in a separate acceptable use policy for staff using AI agents. Specific procedures, such as how quickly an agent's actions can be reversed, sit beneath it too; the reversal window policy template is one example.
For the drafting process itself, consultation, approval and rollout, see how to write an AI agent policy.
How do you fill in each section?
Write in rules, not aspirations. "Agents will be used responsibly" cannot be tested. "No agent may release a payment without a named person's approval" can.
- Authority model is the heart of the policy. State that authority is granted per capability, set by evidence, capped by a ceiling, and that it expires.
- Always-human actions lists what no agent may do at any depth in your organisation. Keep it short and absolute.
- Approval to deploy names the bodies and roles that approve each stage.
- Records states what is kept for every action and verdict, and for how long.
- Suspension says who can stop an agent and on what triggers, without needing anyone's permission.
Once the policy exists, connect your agents to it so they apply the current version; connecting an AI agent to your policy library explains how.
The AI agent policy template
| No. | Section | What to write |
|---|---|---|
| 1 | Purpose and scope | Which agents, which entities, which uses |
| 2 | Definitions | Agent, capability, depth, held, refused, owner |
| 3 | Roles and accountability | Board, executive owner, agent owners, second line, internal audit |
| 4 | Permitted and prohibited uses | Uses allowed, uses banned outright |
| 5 | Authority model | Per capability, evidence-based, ceiling, expiry |
| 6 | Always-human actions | Actions reserved for a person |
| 7 | Approval to deploy | Stages and approvers |
| 8 | Testing and certification | On whose cases, against what floor, how often |
| 9 | Data and sources | Allowed data classes; authoritative sources |
| 10 | Transparency | Disclosure to customers and staff |
| 11 | Monitoring and review | Daily, periodic, triggered reviews |
| 12 | Change control | What counts as a change; who approves |
| 13 | Records | What is kept, where, how long |
| 14 | Incidents and suspension | Triggers, who can suspend, reporting |
| 15 | Third parties | Vendor agents, contracts, exit |
| 16 | Policy owner and review | Owner, frequency, version |
A filled example
A hypothetical consumer lender, sections 4 to 6 and 14.
4. Permitted and prohibited uses. Agents may read, summarise, draft and carry out routine servicing within limits. Agents may not make a final decision to decline credit, set an individual's interest rate, or close an account.
5. Authority model. Each capability of each agent has a depth from 1 (read) to 4 (move money or hard-to-undo actions). Depth is set by evidence from testing on our own cases, capped by the ceiling the risk committee sets for that capability. Evidence expires after 90 days. A failed test drops the capability to read only that day.
6. Always-human actions. Releasing a payment above the agent's limit. Agreeing a forbearance arrangement. Sending a final response to a complaint. Writing off debt. Changing a customer's bank details.
14. Incidents and suspension. Any agent owner, the head of operational risk or the duty manager may suspend an agent immediately. Suspension does not need approval. Incidents are logged within one working day and reviewed by the second line.
The always-human list is short on purpose. Everything not on it is governed by depth and evidence, which lets the policy stay stable as agents improve. The shorter answer to the same question is in how to write an AI agent governance policy.
Questions people ask
Who should approve the AI agent policy?
The board or a board risk committee, on recommendation from the executive owner. Subordinate procedures can be approved by the executive owner.
Should the policy name specific vendors or models?
No. Vendors and models change faster than policy. Name them in procedures or the inventory, and keep the policy about authority, evidence and accountability.
How long should the policy be?
Most fit in eight to twelve pages. Detail on testing methods, thresholds and templates belongs in procedures beneath it.