A novel approach for recovering from errors during automatic assembly in typical mating operations is presented. It is based on automatic error detection w.r.t. a predefined process model, followed by choosing a recovery strategy from a repository. The latter comprises successful strategies that were recorded from human demonstration during a larger scale user study. The paper shows how to enhance the process model with additional data, how to record new strategies in case no suitable strategy is found, and how to select the most appropriate recovering strategy. The added value of the new approach is demonstrated by an experimental validation.