It is easy to make confident claims about what AI will change. I would rather build evidence through real users and real outcomes.