Probabilistic Graphical Models play a crucial role in Microsoft's online services. In this talk, I will describe two powerful applications of machine learning in practice. TrueSkill is Xbox Live's Ranking and Matchmaking system and ensures that gamers online have balanced and exciting matches with equally skilled opponents. AdPredictor is the system that estimates click-through rates (CTR) for ad selection and pricing within Microsoft's search engine Bing. The two systems have in common that they are based on factor graph models and approximate Bayesian inference. They operate at a very large scale involving millions of gamers and billions of ad impressions, respectively. However, in this talk, I will put particular emphasis on those aspects of these applications that are not part of the generic machine learning setting: a) The difficulties that arise because these are closed-loop systems in which the predictions determine the future composition of the training sample. b) The consequences of the fact that these systems make decisions that have an impact on more or less rational agents (advertisers, users, gamers) with the ability to influence the training sample. Time permitting, I will show the two systems in action. This is based on joint work with Ralf Herbrich, Thomas Borchert, Tom Minka, and Joaquin Quińonero Candela.