The rapidly changing outlook of the electric grid with growing use of renewable energy imposes additional needs for operational flexibility. An opportunity to fulfill a portion of these grid services can come from the demand flexibility (DF) of commercial buildings. However, the practicality using these systems as grid resources hinges on the development of scalable and accurate load flexibility models. To satisfy both properties, we propose using internal HVAC system states to generate a feature called thermal unloading potential to predict demand flexibility. The proposed feature is designed to capture the balance between internal and external thermal loading. We evaluate this DF estimation framework with simulation experiments using three standard prototype office buildings during cooling operation with five distinct weather scenarios, yielding over 15,000 load shedding events. Subsequently, we fit regression models using our proposed feature and compare it to models that rely on outdoor air temperature (OAT) as predictor variable. Our results show that overall the proposed feature has a 55% reduction in error on average across all simulated events.