From an academic point of view, Machine Learning models generally try to minimize the errors they make in their predictions. As a theoretical approach it is very efficient, but in addition ML models used in the real world must be efficient: we want them to have the greatest possible benefit with the lowest associated cost (considering the many meanings of the term). In this regard we use different techniques: from the adjustment of thresholds to the use of algorithms with uncertainty ranges, through the use of custom error metrics that take into account the cost associated with a model. In this session we will talk about real cases learned: Machine Learning from the real world.