Most organizational AI training is a tour of features. People leave impressed and return to work unchanged, because nothing in the session connected to a task they actually perform.
Adoption is generally blocked by two things that a feature tour cannot address. The first is judgment: knowing which tasks are genuinely improved by AI and which are simply produced faster and worse. The second is permission: whether using these tools is sanctioned, on what data, to what quality standard, and who remains accountable for the output.
Both are addressable, but not with demonstrations. What changes practice is having people bring a real task, attempt it with support present, and leave with one workflow they have already tested — including the part where it failed.
Responsible use belongs inside that competence conversation, not appended as a compliance slide. Accuracy verification, data handling, and disclosure are part of doing the work well, not a separate obligation.