Do not start with the biggest ambition.
The first workflow should not be the whole transformation program. It should be a recurring piece of work where the pain is visible and the owner is close enough to help.
Good candidates repeat often, cross at least one handoff, and already create delay, rework, trust problems, or failed AI experiments.
Look for drag you can observe.
A monthly report, a payroll exception, a customer triage queue, or an internal handover is usually better than a vague AI initiative.
You want a workflow where people can point to the current state and say: this is where it breaks.
Start where recurrence, pain, ownership, and AI leverage are visible.
Make the first workflow teach the next one.
The output should not be a one-off fix. It should leave behind a map, decision logic, AI boundaries, and transfer notes the team can reuse.
That is how one workflow becomes a pattern for making AI useful in the next one.