When an AI agent calls several systems for a user, the usual shortcuts are a shared account or passing one credential down the chain. Both hide who authorized what; a new, narrow credential per stop keeps that trail.
An AI agent on an employee's ERP account leaves only that name in the logs, so no one can tell who changed what or stop just the AI. Give it its own identity, deny by default, and escalate above set amounts.
When an AI workflow tool is exposed, the danger is the credentials stored in each flow. Close entry points, preserve logs, rotate by which flow held which key, and reissue with less access.