This paper describes theory-informed AI-powered chatbots as Socratic tutors in a first-year Manufacturing Technology course (Productietechniek 1) at The Hague University of Applied Sciences. The chatbots combine large language models (ChatGPT and Gemini) with Retrieval-Augmented Generation (RAG), drawing on curriculum-specific resources including lecture materials and worked examples. Unlike much AI-in-education research, the implementation is grounded in established learning theories: Cognitive Load Theory, Explicit Instruction, Self-Regulated Learning, Formative Assessment, Self-Determination Theory, and Sociocultural Learning Theory. Students used the chatbot collaboratively in small groups to analyse products and design production processes. Across five weeks of structured teacher observation in four parallel class groups, the two observers noted patterns consistent with the theoretical predictions: sustained on-task behaviour, more explicit technical reasoning in group dialogue, engagement with AI fallibility as a learning moment, and emerging articulation of what a professional product requires. These observations are offered as candidate design hypotheses rather than evaluated outcomes. In the subset of chatbot exchanges directly observed, incorrect AI output was frequently traceable to incomplete or imprecise student input, and these moments appeared to function productively within group discussion. This work contributes to the CDIO initiative by demonstrating how AI can enhance formative feedback, student agency, and professional competence development as a theory-informed conversational learning partner.
