RAG systems can retrieve relevant information and still generate incorrect or unsupported answers. As AI applications move into production, engineers need reliable ways to detect these failures and determine when a model should respond, escalate, or abstain. Drawing on research into educational recommendation systems and risk-controlled Agentic RAG, this session explores how retrieval uncertainty, generation uncertainty, and statistical calibration can improve the reliability of AI-generated responses. Key takeaways: → Identifying hallucinations and measuring uncertainty across retrieval and generation → Designing risk-based controls to determine when models should answer, escalate, or abstain → Applying these techniques to educational recommendations and other evidence-grounded AI applications