Causal nets have been recently proposed as a suitable model for process mining, due to their declarative semantics and compact representation. However, the discovery of causal nets from a log is a complex problem. The current algorithmic support for the discovery of causal nets comprises either fast but inaccurate methods (compromising quality), or accurate algorithms that are computational demanding, thus limiting the size of the inputs they can process. In this paper a high-level strategy is presented, which uses appropriate clustering techniques to split the log into pieces, and benefits from the additive nature of causal nets. This allows amalgamating structurally the discovered Causal net of each piece to derive a valuable model. The claims in this paper are accompanied with experimental results showing the significance of the high-level strategy presented.