Techniques for real-time offer customer preference learning are presented. Local agents on communication channels are equipped with predefined rules that capture actions and behaviors of customers interacting with an enterprise. The metrics associated with these actions and behaviors are plugged into the rules and in some cases combined with known pre-existing preferences for the customers for purposes of evaluating the rules and creating newly learned preferences for the customers. The newly learned preferences are dynamically fed into offer evaluation processing to determine whether to make offers to the customers.