This work presents the deployment of a machine learning (ML)-based HVAC-driven strategy through a co-simulation framework that integrates EnergyPlus with real-time decision-making to shift the cooling demand in a logistics hub, accommodating electric-truck charging. Building on prior findings, the proposed approach overcools the indoor space, maintaining a reduced setpoint for a 90-min period, followed by a setpoint relaxation phase. Unlike fixed-start-time interventions, the proposed approach dynamically schedules the onset of the setpoint changes so that the setpoint relaxation phase (charging window) lasts until the end of warehouse working hours, thereby ensuring uninterrupted charging while maintaining occupant comfort. ML models are initially trained on simulations obtained during a development period, where interventions are applied at a fixed start time. To assess the practical feasibility of this approach, the ML models are then embedded in a co-simulation framework. At each timestep, an agent uses real-time data with the models to predict the relaxation phase duration and decide when to trigger or delay pre-cooling. This dynamic scheduling is crucial to ensure that the setpoint relaxation does not end prematurely, which would cause normal cooling to resume and interrupt the charging window. The co-simulation enables safe deployment of the ML-driven strategy by testing models across varying conditions and timestamps, assessing how prediction errors translate into energy penalties without risking comfort, equipment, or operations. The results demonstrate that after introducing a safety margin, an average charging window of 48.77 min is obtained, while eliminating energy penalties and load surges associated with premature HVAC reactivation.