Mathematical under-preparedness is a persistent barrier for many university students in quantitative disciplines, particularly in settings where instructional time is limited and individual support is scarce. This paper evaluates a brief, reusable classroom exercise that introduces students to using a large language model (LLM) as an on-demand, interactive tutor within a flipped classroom at auniversity or high school level. The activity takes approximately fifteen minutes and combines small-group problem solving, individual LLM-guided coaching on the same problem, a class discussion and evaluating feedback quality. We collect student evidence during and after the activity using an in-class worksheet and survey items on engagement, perceived learning, and the relative value ofanLLM help compared with peers and existing course software. Students report high engagement and rate LLM assistance as at least as useful as in-class peer support, while describing the LLM as complementary to the course’s learning platform. Reported learning experiences are more positive among students using more advanced LLM versions. These findings suggest that a short, structured in-class exercisecan help students learn how to use LLMs productively for quantitative problem solving, and provides a practical template for instructors seeking to integrate generative AI into quantitative gateway courses.