AI agents are moving from answering questions to accessing enterprise data, calling tools, and taking actions. That makes identity and authorization part of the agent architecture, not an afterthought. In this demo-heavy session, we will trace an agent request from user identity through data retrieval, reasoning, tool selection, and action execution to show where permissions can be lost or unintentionally broadened. We will examine delegated identity, least-privilege access, data entitlements, policy enforcement, tool authorization, approval gates, and auditability. Live demonstrations will contrast unsafe patterns, such as broad service-account access, with architectures that preserve the user's authorization context throughout the workflow. Attendees will leave with a practical reference architecture and security checklist for building enterprise AI agents that can access sensitive data and take useful actions without becoming more powerful than the people they represent.