LLM customization promises better quality and control, but reality is mixed. This talk reviews what works well in practice, what disappoints, and what remains hard even for experienced teams. It contrasts common customization paths across Bedrock and SageMaker, drawing on real trade-offs in data, cost, and operational burden. Along the way, it highlights failure modes, hidden constraints, and sharp edges that rarely make it into blogs. The perspective is practical rather than theoretical, drawn from ML engineers who work on these problems every day. Daniel Zagyva, Sr. Delivery Consultant - AI/ML - AWS; Laurens van der Maas, Delivery Consultant AI/ML - AWS; Aleksandra Dokic, Sr. Delivery Consultant AI/ML - AWS; Mark Andrews, Senior Principal, Product Management - AWS. Recorded at the 5th annual AWS AI and Data Conference 2026 on March 12, 2026 at the Lyrath Convention Centre, Kilkenny, Ireland. Learn more about AWS events: go.aws/events Subscribe: More AWS videos: b...