Data volumes have continued to grow rapidly making it imperative to increasingly rely on machine learning for the discovery and classification of transients (supernovae, TDEs, flares etc.), many classes of variables (e.g. CVs, AGN, RS CVns etc.), as well as asteroids and comets near and far. Using the example of ZTF we describe some aspects of machine learning workflow for this purpose. Given the large number of possible discoveries our aim is to minimize follow-up observations while keeping the classifications interpretable.