Early applications of virtual reality (VR) technology in psychological assessment, treatment, and research have yielded promising results. In particular, an increasing number of studies analyze the unique features of the experience made by patients during their exposure to virtual environments. However, the majority of these studies explore how patients navigate in the virtual spaces and interact with virtual objects. Only a few of them investigate the features of inhabited virtual environments, where real people and autonomous virtual humans are able to interact and to cooperate. In particular, there is a lack of discussion of the role that such autonomous virtual humans could have in VR-aided psychotherapy. The main goal of this paper is to identify a framework for future research in this area. Three levels of analysis are identified. The purpose of the first two levels is the identification of the key "physical" features (e.g., appearance, structure) and "internal" characteristics (e.g., behavior, degree of autonomy, perceptual capabilities) needed by an effective simulation. The third level is concerned with the evaluation of the interaction characteristics required for a successful relationship between the patient and the virtual human.