Permission design FAQ

Short answers to common questions about deciding what a customer-facing AI agent may read, recommend, change and route for approval.

What is permission design?

Permission design is the work of deciding what an AI agent may access, recommend, change and commit the business to doing. It turns broad goals, such as handling returns, into specific actions with evidence, limits, approval routes and recovery rules.

Where should I start?

Start with one customer task, then break it into actions. The permission design guide explains the concept, and How to Decide What Your Customer-Facing AI Agent Is Allowed to Do walks through a complete planning method.

What are the five permission levels?

PermissionByDesign uses five levels: Inform, Recommend, Act within limits, Request approval and Not permitted. They are distinct choices, not a ladder toward more autonomy. A useful agent may remain mostly informational, while a narrow routine action may be allowed within enforced limits.

Does every customer-facing action require employee approval?

No. Routine customer-facing actions can be authorized within enforced limits when the evidence, scope and recovery path are clear. Employee approval is most important for exceptions, high-impact actions, weak evidence, policy overrides and actions outside routine authority.

How is customer confirmation different from employee approval?

Customer confirmation means the customer confirms what they want, such as returning an item or using the original payment method. Employee approval is a business decision that a restricted action may proceed. Customer confirmation does not replace employee approval when the plan requires it.

Why does read-only access need planning?

Read-only access can still reveal private information. A returns agent may need one customer’s order and delivery record, but not unrelated account notes, sensitive internal fields or another customer’s orders. Plans should define approved sources, allowed fields, excluded information and what may appear in outbound messages.

Why do per-action limits sometimes need cumulative limits?

A per-action limit controls one execution. It does not automatically control repeated actions. Several individually allowed refunds may exceed a customer’s total allowance, or a bulk action may affect more records than intended. Cumulative limits define the total amount, count, period and scope, such as one order, one customer, one account or the agent’s overall activity.

What happens when evidence is missing?

Missing required evidence should pause execution. Approval does not automatically resolve missing evidence, and it never permits an action marked Not permitted. The plan should say what evidence is required, who can review an exception and what customer message is allowed while the case is paused.

What should happen when a tool times out?

An unknown outcome is different from a confirmed failure. If a refund request times out, the system should check transaction status before retrying. When an AI Agent’s Tool Fails explains pause, verification and duplicate-action protection.

Does the builder configure or certify an AI agent?

No. The Permission Plan Builder documents intended authority in the browser. It does not configure, certify, secure or connect to an AI agent. Connected production systems still need to enforce access controls, approval checks, execution limits and records.

How should I test a permission plan?

Test ordinary allowed actions and attempted exceptions. Include missing evidence, unavailable approvers, prohibited actions, duplicate or repeated actions, another customer’s record and a prompt-injection attempt such as an uploaded document that says to ignore the refund limit. Design for Correction and Recovery explains how recovery planning affects authority.