Many teams start with contained generative AI pilots: a narrow use case, a couple of integrations, and a human checking every output before it goes anywhere. Once that same system is expected to carry a full workflow instead of a single feature, the assumptions behind how the generative layer was originally built stop holding up. This session looks at what actually shifts inside a generative stack as a project moves from a contained pilot to something a real workflow depends on. It focuses on where prompting, retrieval and evaluation start to strain once outputs stop being one-off answers and start becoming inputs the next step relies on. What this session will cover: - Where a feature-level generative AI pilot starts to break down once it is asked to carry an entire workflow end to end - How prompting and retrieval design need to change once an output becomes a dependency for the next step, not just a one-off result - What evaluation has to measure once a generated output feeds a downstream step, not just when it produces an acceptable answer on its own - How teams are rebuilding their generative stack for reliability before layering any further automation on top of it