Many applications, such as smart building automation, crowd flow analysis, action recognition, and assisted living, rely on occupancy information. Common to each is the need for the ability to measure occupancy, subject to constraints of cost, privacy, scalability and generalization. This work investigates a scalable Wireless Sensor Network (WSN) with CO2-based estimation as a viable solution. To support many applications, such a system must be transferable and must function without knowing the physical system model; the CO2-occupancy dynamics should be learned directly from system observations. Using the data captured from occupancy experiments, five different machine learning models were trained on the task of occupancy estimation. These models were subject to different training conditions to assess the consequences of machine learning design decisions on performance and to study how these consequences relate to the WSN constraints. Model specific design consequences are identified, and their impact on designing a generalized occupancy estimation system are discussed.