<p>This study investigated the nature of modelling processes when prospective mathematics teachers engaged with a modelling activity supported by a purpose-built ChatGPT bot configured with prompts corresponding to the actions of Blum and Leiß’s (2005) modelling processes. Using an exploratory sequential mixed-methods design, the study employed two levels of analysis: (a) analysis of modelling actions across the modelling process, and (b) analysis of instrumental interaction through instrumental genesis, which is a framework for conceptualizing tool use as development through interaction, shaped by the intertwined processes of instrumentation and instrumentalization. Data were collected from screen–recorded participant–bot dialogues of 18 prospective teachers individually engaged in the modelling activity. Findings revealed three distinctive modelling routes: spiral (61%), characterized by frequent returns to earlier actions and iterative assumption revision; linear (28%), a largely unidirectional pass through the cycle with procedural bot use; and fragmented (11%), marked by skipped or superficial actions and instrumental bot use to obtain a solution. Modelling routes showed a strong association with interaction categories (instrumentation, instrumentalization, and turns-outside). At the actions level, understanding and simplification together accounted for 41% of all instrumentation events and 48% of all instrumentalization events, indicating strong scaffolding and participant adaptation during problem framing. These results contribute to teacher education by clarifying how conversational AI can support key aspects of modelling.</p>