This paper presents a method which, on the basis of a parametric model of a stable process and of a record of the process response to a convenient input signal, provides a nonparametric overbound of the model error magnitude. The results obtained in no sense depend on the model complexity. Simulation examples are reported to show the method's effectiveness in improving the performance of model-based autotuners.