Agentic IAM in Practice: Enforcing Least Privilege and Auditability for AI Agents