Predictive process monitoring (PPM) is focused on the prediction of different aspects of running process instances based on information of past executions. On the other hand, causal process mining aims to identify cause-and-effect relationships within business processes. In this paper, we analyze how causal inference can improve the performance of a predictive model using causal graph discovery. This perspective has been barely used for predictive monitoring. Our proposal, named CI-PPM, has been experimentally tested using three real-life case studies. The findings demonstrate that CI-PPM improves error metrics in half of the datasets analyzed, while achieving training efficiency gains across all cases.