Data-driven modeling of complex dynamical systems can be very challenging. However, by explicitly considering the quality requirements of the intended use of the model this task can be significantly alleviated. Application oriented experiment design (AOED) is a systematic way to do this. It facilitates identification of system properties that are important for the application at hand, at the same time as it allows simplified model structures to be used since it, for reasons of experimental economy, avoids exciting system properties of little consequence for the application. In this talk we outline the theory for AOED and discuss how to use this technique in an on-line context such that the system is actively better and better probed in a sequential manner as more and more information is acquired. In particular we show how the technique can be integrated in model predictive control.