Current approaches to simulate occupants' wayfinding in AEC mostly employ direct routing algorithms that assume global knowledge of the navigation environment to compute a shortest path between two locations. This simplification overlooks evidence concerning the role of perception and cognition during wayfinding in complex buildings, leading to potentially erroneous predictions that may hinder architects' ability to design wayfinding by architecture. To bridge this gap, we present a novel simulation paradigm entitled Cognitive Occupancy Modeling in BIM to simulate wayfinding by means of a vision-based cognitive agent and a semantically-enriched navigation space extracted from BIM (Building Information Modeling). To evaluate the predictive power of the proposed paradigm against human behavior, we conducted a wayfinding experiment in Virtual Reality (VR) with 149 participants, followed by a series of simulation experiments with cognitive and direct routing agents. Results highlight a significant correspondence between human participants' and cognitive agents' wayfinding behavior that was not observed with direct routing agents, demonstrating the potential of cognitive modeling to inform building performance simulations in AEC.