Computer vision systems can employ multiple vision models to complete a single task or an array of tasks. Reasons may span from no single model being available that meets user requirements, hosting devices lacking the compute to execute a single model that contains the full required functionality, or training a new model requires extensive resources or expertise. Without intelligent input discrimination, these systems risk inefficient processing, leading to increased inference times and energy consumption. This paper investigates the impact of intelligent model activation regulation on energy efficiency and inference speed. We propose BICEC (Branched Image Classification Evaluative Controller), a lightweight solution based on a branched EfficientNetv2 architecture. BICEC adapts to existing vision systems without requiring system retraining by creating model-specific branches optimized for minimal size and near-optimal performance. Results show good performance for identifying when a model is relevant and significant reductions in system inference time and energy cost. While the scope of this work focuses on vision systems, we hope to exemplify how tighter control of AI systems can enhance sustainability and computational efficiency.