When Palo Alto Networks deployed agentic AI across its technical support organization, the biggest challenges were not model performance but workflow design, feedback quality, and adoption. A small SME group could not scale training effectively, while broader decentralization produced inconsistent, low-quality feedback that required additional validation layers. The breakthrough came from embedding AI into real workflows and introducing stronger quality control on human feedback loops. In this session, Ameya Kawimandan shares what broke, what was rebuilt, and how those changes reduced time-to-resolution by 25% while improving AI efficacy from 60% to over 72%. Attendees will learn practical lessons on building scalable AI support systems that work with teams, not around them.