Disclosed herein is an automated personalized message composition system, method and architecture. A composition style of a user is learned for each of a number of categories, such that each category has a corresponding composition style. The user's composition style determined for a given category can be used to personalize a non-personalized message for the user. A personalized message including elements of the user's composition style. The composition style elements from the user's composition style replacing a number of non-personalized elements in the non-personalized message.