Many teams start with contained GenAI pilots: narrow use cases, limited integrations, and clear human oversight. Once those same systems are expected to power larger workflows, the assumptions behind how the generative layer was designed stop holding up. This session looks at what shifts inside the generative stack as organisations move from experiments to workflow‑level impact. It focuses on the points where prompting, retrieval, evaluation, and model orchestration start to strain as systems become more connected and their outputs drive real work. What this session will cover: - Where early GenAI pilots and “feature‑level” architectures begin to fall short once they’re asked to support end‑to‑end workflows - How prompting and retrieval design change when outputs become dependencies between steps instead of one‑off results - What evaluation needs to look like when generated outputs trigger actions, not just acceptable answers - How teams are reshaping their generative stack before introducing genuinely autonomous or agentic behaviour