Large organizations develop guidelines, processes, and tools to help them implement and manage IT. In many cases, these encode policies, the central means of control for Autonomic Computing systems. Policies consist of goals, scope, constraints, and tradeoffs in system management, and define the boundaries of delegation of work between people and systems. Yet little is known about policy-based interaction among people and between people and systems. Are policies created once and used routinely over long periods? Or do policies change frequently? How do policies work in the real world? Our field observations suggest that policies are continually negotiated and re-interpreted as they are applied to specific instances. This implies several design points, including the need to make policy application transparent and flexible, enabling lightweight, rapid, incremental, and reversible interactions with policy-based systems.