Within Motivational Interviewing (MI), client utterances are coded as for or against a certain behaviour change, along with commitment strength; this is essential to ensure therapists soften rather than persisting goal-related actions in the face of resistance. Prior works in MI agents have been scripted or semi-scripted, limiting users’ natural language expressions. With the aim of automating the MI interactions, we propose and explore the task of automated identification of client motivational language. Employing Large Language Models (LLMs), we compare in-context learning (ICL) and instruction fine-tuning (IFT) with varying training sizes for this identification task. Our experiments show that both approaches can learn under low-resourced settings. Our results demonstrate that IFT, though cheaper, is more stable to prompt choice, and yields better performance with more data. Given the detected motivation, we further present an approach to the analysis of therapists’ strategies for balancing building rapport with clients with advancing the treatment plan. A framework of MI agents is developed using insights from the data and the psychotherapy literature.