You could train an LSTM network to generate responses in the same way that google's inbox response suggestions work. Requires a mountain of source data, but hey, you probably have 30GB of email sitting in your Gmail, right?
Actually I think a NN might require way too much supervision. You could feed exits (emails that got no response) into a decision tree to know what to send at any point in the chain to maximize response.