Energy is used in buildings to provide services to occupants, but occupants are poorly represented in mainstream building energy models. It is argued that this stems from the 'vernacular architecture' of building energy models, an architecture characterised by agglomeration of selected, epistemically diverse forms of knowledge sanctioned by the dominant 'folk ontology' of the field that favours the material, visual, physical and causal in model building over the configurational, invisible, variable and correlational. Within this context, a new form of occupant representation using Bayesian networks is outlined, evaluated, and its integration into existing energy models discussed.