A vast number of game developers long for and seek to create games that impact and emotionally engages their players. However, it has proven to be a tough challenge to overcome, as a lot of narrative games today completely separate their emotional stories from their engaging gameplay. Extending this insight, this paper proposes the theoretical basis intendedto contribute to the understanding of how to create expressive AI agents, by answering the following question: How can developers tell stories through an AI agent's behaviour? A further key element in this paper is the creation of a conceptual framework, based on a literature study, exemplifying how AI agents can be designed to contribute and support the narrative in a game. The framework suggests that creating these types of AI agents is an iterative process of defining the core themes within a game, deriving an agent's purpose from these and building its behaviours from that purpose. Also, defining how to communicate these behaviours to the player. The framework also emphasizes the importance of user testing during the design process, as a way of evaluating the balance of transparency and emergence within the AI system of the game. The framework is exemplified with a case study conducted on an AI agent, designed and developed for the game Shelter 2: Paws by game studio Might and Delight. The results of the study shows that AI agent has a close connection to the narrative through the main themes of friendship, cooperation and growing up in the game. Finally, the AI agent design is evaluated through user tests, which shows that players interact with the agent as expected and share the game experience that is intended by the developers.