Building automation systems are believed to hold the key to significantly reducing the average energy consumption of our residential and commercial building stock, which in the U.S. is responsible for 41% of the total annual energy use in 2014. As these systems become more widespread and inexpensive, the complexity and challenges associated with their installation, maintenance and upkeep will increase. One of the primary challenges is the generation and update of the meta-data associated with the sensors and control points distributed throughout the facility. Previous research has attempted to reduce the human input required to perform these activities, by leveraging different signal processing and statistical analysis approaches to infer the sensor types and locations from measurements and/or tags obtained through a BAS. However, because of the relatively small sample size, the feasibility of applying these type approaches on large buildings, as well as their generalizability, remain as unsolved questions. In this paper, we propose a meta-data inference framework to learn from BAS measurement data in a semi-automated way. Furthermore, we evaluate the framework on two large buildings instrumented with thousands sensors and show the feasibility of applying data-driven approaches in the real world. We present the results of our study and provide recommendations for future work in this area.