A mixed parametric-nonparametric approach to H-infinity identification is proposed, aimed to estimate a low-order approximating model and an identification error giving a measure of the model perturbation in a form well suited for H-infinity control methodologies. It is shown how to estimate the modeling errors of the identified parametric model and how to evaluate the performance values that can be guaranteed when the H-infinity controller is designed and applied to the real system. This performance value is then used for selecting the most "suitable" order of the parametric model class.