Start Here

Start here

Permission design turns policy into action limits.

A customer-facing AI agent needs more than instructions. It needs clear boundaries for what it can do directly, what it can recommend, what needs approval, and what must be logged for review.

The basic question

For each task, ask: what business action could the agent take, and what would make that action risky, costly, irreversible, or sensitive?

A useful permission plan names the action, the normal boundary, the approval trigger, the approver, the evidence required, and the test scenarios that should be run before launch.

Five permission levels

  • Inform: The agent can retrieve or explain information.
  • Recommend: The agent can propose an action but cannot execute it.
  • Act within limits: The agent can complete the action when defined conditions are met.
  • Request approval: The agent must route the action to a person or approved workflow before it happens.
  • Not permitted: The agent must not propose or execute the action.

Ready to map a real customer task? Use the Permission Plan Builder to record actions, limits, approvals, recovery decisions and review triggers.

Prefer a guided PDF first? Download the Permission Design Playbook for a 12-page walkthrough of the same planning decisions.

Questions about the planning terms? Read the permission design FAQ.

Controls are not the same as instructions

An instruction tells the agent what it should do. A control restricts what the system will actually allow. Permission design is most useful when business teams can describe the desired boundary clearly enough for technical teams to enforce it.

What to document

  • The customer-facing task and the action the agent may take.
  • The data or evidence the agent must inspect before acting.
  • The limits for routine cases.
  • The approval triggers for exceptions.
  • The record that should remain after the action.
  • The scenarios that should be tested before launch.

Build your first permission plan

Ready to put these principles into practice? Read How to Decide What Your Customer-Facing AI Agent Is Allowed to Do for a practical guide to defining allowed actions, limits and approval requirements.

Further reading

For a security perspective on agent permissions, read Udaya Bhaskar Vemuri’s AI Agent Access Control: Give Agents What They Need and Nothing More on VKTR. It explains how access can expand over time and why businesses need clear limits, approval requirements and records of agent activity.