The thesis examines the role of artificial intelligence in supporting exam preparation in Italian STEM degrees, where high dropout and delay rates are concentrated around a few “gatekeeper” courses such as Calculus 1–2, Physics, and Linear Algebra. It asks whether vertically specialised, exam specific AI tutoring platforms can provide effective learning support and achieve sustainable unit economics in a context where general purpose chatbots are already widely used and largely free. Theoretically, the work reviews the evolution of Intelligent Tutoring Systems from rule based architectures to data driven and LLM based tutors, summarising evidence on learning gains, engagement, equity, and risks such as accuracy issues, cognitive offloading, academic integrity, and the constraints introduced by emerging regulation (e.g.