I once (2000 or so) wrote a very smart webmail client used by my then client's customer service. The most used feature were context-sensitive canned responses. It would check the message for keywords (Bayesian filtering came later) and suggest the most likely answer. It was an ugly JSP-based thing that took me a month to write. It should be much quicker to do now with modern technology.
You need to make the canned responses use actual client data, have some variability and track what canned responses were used before for the client. We also allowed the human operator to customize the message and check if what the machine thinks is appropriate really is. We had more than 80% accuracy and we continuously measured that against the keywords found in the previous messages.
