When I give this keynote on agentic workflows, I usually start by disappointing the room. Nobody wants to hear that the first step is not the model, the platform or the prompt. It is a whiteboard. Map the process first, exactly as it runs today, including the messy manual bits, and only then decide what an agent should touch. Every failed automation I have seen in the last two years failed at that step, not at the technology.
The Rule I Give Every Team
Automate the repeatable, not the big batch. If a task runs the same way every week with the same inputs and a clear definition of done, it is a good candidate. If it is a large one off project with judgement calls at every turn, an agent will produce confident nonsense and you will spend longer checking it than doing it yourself.
So the sequence I teach is simple: write the process down step by step, remove the steps that exist only out of habit, define what a correct output looks like, and only then hand a single step to an agent. A broken process handed to AI is still a broken process, just faster and more expensive.
We run this mapping exercise live with teams in our workshops, and you can see how the same thinking shapes product builds in our case studies.



