Modern power grids are increasingly challenged by the growing reliance on renewable energy sources. Due to their inherent intermittency, these sources can cause voltage fluctuations and phase imbalances, particularly during grid disturbances. Among these perturbations, voltage sags are the most critical ones, occurring within seconds or even milliseconds. A required mitigation strategy involves injecting reactive current into the phase experiencing low voltage. Traditional approaches for determining the appropriate amount of reactive current rely on grid codes, which define the minimum value based on measured voltages; therefore, there is no optimization, nor adaptability to better attend to the grid’s needs. In this study, we propose an alternative solution based on Soft Actor-Critic (SAC), a model-free and off-policy reinforcement learning (RL) algorithm which addresses the weaknesses of previous approaches. Simulation results during the inference phase demonstrate that the SAC-based method closely matches the performance of optimization-based approaches, while offering better generalization to unseen data within a fast response time in milliseconds.