An AI agent needs an operating boundary.

Make authority explicit

A prototype can leave permissions implicit because its creator supervises every attempt. A deployed workflow needs an explicit answer about which data it can read, which systems it can change, and which decisions require a person. The functional owner and platform owner have different responsibilities in setting that boundary.

A useful first step is an action inventory for one workflow. Each action records its required permission, observable result, and escalation condition. Uncertain outcomes need a recovery path before another attempt repeats an external change.

Evaluate the boundary as part of acceptance

A Production AI Lighthouse should exercise denied access, missing evidence, partial completion, and requests outside the agreed scope. Acceptance includes whether the workflow stops or escalates appropriately. Successful task execution and appropriate restraint both matter to the operating team.