IT systems have a high impact on organizations and businesses today and are a key to globalization. Evolution of complexity calls for changes in the way these systems are built and managed. As a consequence, a paradigm shift towards Model-Driven Engineering and Policy-Based Management can be observed, both aiming at a better management of complexity. A lot of human effort is still necessary to develop and operate complex systems. In this thesis, an innovative approach to reduce that effort is presented. Domain-specific models are used to describe the IT system and its behavior at a high level of abstraction initially. They are refined into technical representations at lower levels and finally transformed into event condition action policies. These executable policies define which actions must or must not be performed to achieve the objectives of the system. The refinement of models and the generation of policy implementations are both fully automated. As high-level changes are automatically materialized in the respective policies, system behavior can be adapted at runtime simply by changing the high-level models. This allows to operate even complex systems without considering their technical details. Finally, the correctness of the refinement is proven with a relational algebra. The benefits of the apporach are demonstrated and evaluated with two real-world case studies, dealing with the management of mobile networks and the calculation of bonus payments.