ML’s great strength is that example cases are all you need to create a predictive model. The predictions work as long as the underlying process is not tampered with. But clients usually seek more: they yearn to understand the “data-generating machinery” in order to improve the outcome that the model predicts. Yet, this is dangerous without additional external information, including the direction of influence between variables. This talk illustrates how to achieve “peak interpretability” by using influence diagrams to model causal relationships, avoid mistaking correlation for causation, and quantify how outcomes will change when we manipulate key values.