Drug discovery requires balancing efficacy with pharmacokinetic and safety properties. Although machine learning models for ADMET (absorption, distribution, metabolism, excretion and toxicology) prediction are increasingly sophisticated, their impact is often limited by challenges in deployment, user adoption, interpretability and project-specific applicability. In this perspective, we describe Bayer's predicTeam approach, which assumes end-to-end responsibility for the prediction lifecycle, from endpoint selection and data curation to deployment, user training and performance monitoring. Key elements include close collaboration with experimental scientists, automated retraining workflows, structured user education and project-specific applicability reports. Drawing on more than a decade of experience, we argue that data quality, curation, governance and organizational accountability are often more important for impact than algorithmic innovation alone, particularly as AI-driven drug discovery continues to evolve.