The work bridges Learning to Defer (L2D) - an extension of supervised learning where the prediction task can be deferred to a human - and the causal framework of Potential Outcomes. This allows one to identify and estimate the causal effect of introducing a deferring system on the predictive performance.