Every enterprise has a settled process for granting a person access to a system. Almost none have one for granting an agent access. So agents get human credentials, or a shared service account, or a permission set nobody has reviewed since it was created. It works until an auditor asks who authorized a specific action, and the honest answer is that the question was never designed for.

This is a closed-door conversation among senior technology leaders about the accountability layer under agentic AI. Not the model, not the use case, the part where an autonomous system takes a real action in a real system of record and someone owns the consequence.