An AI agent policy is the set of rules that decides what your agents may do: which actions they can take on their own, which need a person's approval, and which are never allowed. Most organisations have an "AI usage policy" document for staff. Agents need something different — rules precise enough to be checked automatically, every time an agent tries to act.
A document isn't an AI agent policy
A written policy like "agents must not make commitments above their authority" sets intent. It doesn't stop anything at 3am on a Sunday. An enforceable AI agent policy is expressed in terms a checkpoint can evaluate: the action, its arguments, the agent, the user and the context. Keep the document — it explains why. The rules do the work.
The three outcomes
Every rule ends in one of three outcomes:
- Allow — the action runs immediately.
- Require approval — the action waits for a named person. See human in the loop for AI agents.
- Deny — the action doesn't run, and the agent receives a clear reason.
The art is putting each action in the right bucket. Most should be allowed; a few should be held; some must never happen.
What a rule looks like
policy "refunds" version "2.4" {
action = "issue_refund"
allow when amount <= 100
require_approval when amount <= 1000 approver = "finance-lead"
deny when customer.flagged
deny otherwise reason = "exceeds agent authority"
}
A good rule is scoped to one action, readable by the process owner, and specific about thresholds. More examples by use case are in the use case library.
What rules can check
- The action —
issue_refund,send_email,delete_resource - Its arguments — amount, recipient, environment, row count
- The agent — which agent and version, which team owns it
- The user — who the agent is acting for, and whether they're verified
- Context — customer tier, region, time, daily totals so far
Who owns AI agent policy
The person who owns the process owns its rules: finance for money, support for customer contact, security for access. Engineering reviews how the rules are implemented and tested. Treat rules like code — versioned, reviewed and reversible — which is the idea behind policy as code for AI agents.
How to roll out an AI agent policy
- Inventory actions. List every tool each agent can call.
- Classify risk. Mark actions that move money, touch personal data, contact customers or change production. An AI agent risk assessment helps.
- Draft rules with process owners. Start strict; loosen with evidence.
- Dry-run before enforcing. Replay past activity to see what each rule would change — see how to test AI agent guardrails.
- Enforce and record. Turn rules on and keep an audit trail of every decision.
- Review regularly. Revisit when agents gain tools or when the logs show friction.
Mistakes to avoid
- Rules in prompts. They're suggestions to the model, not enforcement — see AI agent guardrails.
- Approval for everything. It creates fatigue, and people stop reading.
- No default. Decide what happens when no rule matches. Deny is the safe choice.
- Rules nobody can read. If the process owner can't review it, they can't own it.
Go deeper
- Policy as code for AI agents — versioned rules you can roll back
- How to test AI agent guardrails — dry-run before production
- AI agent governance — a practical operating model
- Agentic AI risks — what can go wrong when AI acts
- AI agent risk assessment — a pre-launch template
Where Gaten fits
Gaten enforces your AI agent policy on every tool call, with readable versioned rules, dry-runs, approvals and a full decision log. See how policies work.
Frequently asked questions
What is an AI agent policy?
A set of enforceable rules that decides, for each action an agent attempts, whether it's allowed, needs a person's approval, or is denied.
How is it different from an AI acceptable use policy?
An acceptable use policy guides people using AI tools. An AI agent policy governs what autonomous agents may do, and is checked automatically on every action.
Who should write AI agent policies?
The owners of the processes agents touch — finance, support, security — with engineering making sure the rules are implemented, tested and versioned.

