In this paper we perform a comprehensive evaluation and cost-benefit analysis for training programs in Germany. We use an adapted version of the theoretical framework developed by Pavoni and Violante (2007) and extended by Wunsch (2007, 2012) and Pavoni, Setty and Violante (2010) for studying optimal welfare-to-work programs. We derive the optimal policy in terms of optimal benefit levels and optimal targeting, timing and duration of training programs based on a dynamic principal agent model. Quantitative results are obtained by simulating the model based on structural estimates of the input parameters and different training technologies from administrative data. We contrast the actual policy with the optimal one and quantify the budget and welfare implications of switching to the optimal scheme. We also consider constrained optima where only parts of the system are being optimized.