Model predictive control (MPC) has shown potential in improving building performance but is bottlenecked by the difficulty in constructing control-oriented models. The challenge lies in evaluating the sufficiency of the model and the data usage beforehand. This paper bridges the knowledge gaps in the interactions between data requirements, model quality, and control performance by integrating real-world measurements and simulation-based experiments. The data usage related to occupancy and Internal heat gain (IHG) was studied considering its importance and the absence of consensus in the literature. Varying occupant-related data sources were tested as RC model inputs, including none, schedule, electricity consumption, CO2 ppm, and ideal measurement. Combinations of model inputs and complexities were examined for prediction and control in an office, a classroom, and multi-zone offices on one floor. The results indicated that the usefulness of data is jointly affected by three factors: measurement suitability, model complexity, and modeling purpose. Given the adequate model structure, satisfying prediction and control performance was achieved in offices with no detailed measurement. Meanwhile, electricity and CO2 were needed together to capture the IHG influence and realize the good performance for classrooms. The experiments also uncovered the heterogeneous requirements on models from traditional prediction tests and the control tasks. Lower prediction error did not always mean better control. More importantly, we provided the first quantitative demonstration of the complementary relationship between model adequacy and data informativeness with respect to different purposes. This study advocates the pioneering idea of sparse data usage and parsimonious modeling, which promotes the actual application of MPC in buildings by guiding control-oriented model development.