Learning predictors that incorporate human decision has been the focus of extensive research in recent years. These predictors are used in order to increase the final accuracy and reduce the risk in high-stake tasks. One of the strategies to keep the human in the loop involves using learn-to-defer methods, in which the prediction is made either by AI or is deferred to the human expert. This strategy has attracted considerable attention due to the reduction of expert bandwidth as well as increasing the accuracy. However, we show that learn-to-defer methods are not optimal considering their understanding of the task and human decision. In this paper, we first derive the optimal predictor that provides the defer option, while incorporating human decision into its final prediction. We further show strict improvement of this method upon learn-to-defer methods, both theoretically and empirically. The code for this paper is open-sourced on \url{https://anonymous.4open.science/r/BeyondDefer-9073/}.