Pick a workflow with visible friction
The strongest starting points are repeated often, painful enough that people want them fixed, and contained enough that you can observe the result. Think document intake, proposal preparation, recurring reporting or enquiry handling.
A glamorous project with unclear ownership is usually a worse first move than a boring workflow that consumes hundreds of hours a month.
Make the human role explicit
Good implementation does not hide the human. It decides where people add judgement, where they approve actions and where the system can safely proceed on its own.
That clarity makes adoption easier because the team understands what has changed and what has not.
Optimise for the second project
The first deployment should leave behind reusable patterns: how you connect data, how you manage permissions, how you test outputs and how you measure value.
When those patterns exist, the next workflow should be faster. That compounding speed is a much better signal of AI maturity than the number of pilots running.