So far, most of the work in Knowledge Representation has modelled concepts as classes, i.e., as sets of instances. However, as from the work in Teleosemantics, concepts can also be thought as abilities of performing certain (biological) functions. The shift is from the study of the means by which the world is represented to the study of the reasons and means by which such representations are generated. In this paper, which is grounded in the seminal contribution by the philosopher Ruth Millikan, we focus on substance concepts, namely on concepts as on recognition abilities, and on how this notion can be mapped to that of concepts as classes. The ultimate goal is to provide a unified theory of perception and knowledge representation that, eventually, will allow us to go beyond the limitations and lack of robustness of current Artificial Intelligence (AI) systems. We provide three main contributions: i) a model of concepts as abilities, with a focus on recognition abilities, ii) an early version of an Ontology of (Recognition) Abilities (called RAO) and iii) the beginning of a methodology for how to use RAO for discovering which classes, among those contained in the state of the art ontologies, correspond to recognition abilities.