In recent years, the use of AI systems has intensified, particularly in the field of education. It is a tool that could be useful in solving modelling problems, such as Fermi problems. These types of problems are suitable for introducing modelling in primary education classrooms. In the present study, we want to explore how pre-service primary school teachers use ChatGPT when solving two Fermi problems. The qualitative analysis of the interactions with ChatGPT and the written solutions produced by 94 pre-service teachers allows us to categorise the prompts by distinguishing categories that come from the prompting techniques from those that emerge from the phases of the modelling cycle. Through this system of categories, we characterise the use of AI by distinguishing three levels depending on the degree of responsibility given to the AI (from full delegation to instrumental use) in the modelling process: use of ChatGPT as problem problem-solving expert, assistant and support.