A computer implemented method, computer system and computer program product are provided for transferring in a BOT conversation. According to the method, a user input is received, by a device operatively coupled to one or more processing units, from a user. A first response from a conversation BOT responding to the user input is obtained by the device. The first response is evaluated by the device according to configured rules to determine whether a human agent is needed, wherein the evaluation according to at least one of the configured rules is carried out by a trained engine of a reinforcement learning model. Finally, responding to determine the human agent is needed, a transferring recommendation is sent to the user by the device.