Approval Policy

Simple Definition

An approval policy is the rulebook for an AI agent’s independence: which actions it may take automatically, which it may take only when you explicitly asked for them, which need your approval every time, and which it must hand back to you entirely.

Plain-English Explanation

Without a policy, an agent has two bad options. It can ask permission for everything, which makes it tiresome and slow, or it can decide for itself what is risky, which is exactly the judgment you may not want to trust. An approval policy makes the line predictable.

A typical policy lets an agent read documents, organize files, search, and prepare drafts freely, but requires approval before it sends a message, buys something, deletes files, changes permissions, publishes content, or touches production systems. Some actions may be off-limits even with approval, such as moving money between accounts. Policies can depend on the action, the target, the amount, the data involved, or who the request came from. Products name and structure these rules differently; OpenAI’s Dots call them Custom Rules, for example.

Example

A small business sets up its assistant agent like this. Reading the shared inbox and drafting replies: automatic. Sending a reply to an existing customer thread: allowed only if the owner asked for it in the request. Emailing a new contact, issuing a refund, or deleting anything: ask first, every time. Changing bank details: never, even with approval.

How It Differs From an Authorization Gate

An authorization gate is a single checkpoint before one specific action, the moment where the agent stops and waits for a yes. An approval policy is the set of rules that decides where those gates go and what can pass without one. The policy is the plan; the gate is where the plan is enforced.

An approval policy also depends on knowing what the agent can really do. If its effective permissions are broader than you think, actions can slip past the policy. That is why policies work best alongside a technical action boundary that makes forbidden actions impossible, not just disallowed.