As more and more data about mining complex operations is collected and stored, it becomes increasingly important for computer systems to help human operators make better, more informed decisions. This can be done indirectly, through improved visualization or prediction, or directly, by suggesting decisions that respond to new information. This paper contributes to the direct approach by showing how state-of-the-art data-driven decision-making can be used for optimizing material flows in a large mining complex. To this end, a combination of neural networks and policy gradient reinforcement learning is used for computing material destination decisions that automatically respond to new information. Results using a computational model of a large copper mining complex show that the proposed method significantly outperforms an optimized cut-off grade policy similar to the one currently used at the mine.