Most hospital AI training is a one-hour webinar with a vendor logo on it. Attendance gets logged, nobody's work changes, and six months later the compliance office is still fielding the same question about whether ChatGPT is allowed. Training that works looks different. The difference shows up in what four specific groups of people do on the following Monday.
The useful executive question in 2026 is not whether to use AI. Roughly three quarters of US health systems already run at least one AI application, and industry surveys show the top budget priority has moved from finding new use cases to making the purchased AI perform. The question for leadership is narrower: which workflow, at what measured cost today, checked by whom before the output acts.
After a real briefing, leaders can fill in that one-page answer for any proposal that crosses their desk. They also know the three things that make a pilot drift: no baseline, no named owner, and no rule for who reviews output before it goes anywhere.
What changes in practice is that proposals get shorter and more specific. A director who used to forward vendor decks asks for the baseline first. Governance moves from a policy document to three rules people can recite.
Power users are the 10 to 15 percent of staff (an estimate, not a survey figure) who already use these tools daily, often on personal accounts nobody sanctioned. Training for this group is not an introduction. It is advanced prompting on their own work: the monthly quality report, the payer appeal letter, the policy draft. They learn to chain steps, keep a prompt library, and check output against a source before it leaves their hands.
The more important change is that they learn when not to use the tool. A power user who has watched a model invent a citation in a lab, on a case that looked like their own, stops trusting summaries of documents they haven't read. That habit protects the organization more than any prompt does.
For the general staff session the goal is narrow on purpose. Each person leaves having automated one task in their own job, and knowing three data rules well enough to repeat them without a slide. Ambition beyond that in a half day produces a slide deck, not a habit.
The one win is usually small: a draft of the intake letter they retype weekly, a first pass at a meeting summary, a cleanup of a spreadsheet export. Small is fine. The point is that they did it on their own workflow, on a sanctioned account, and saw the failure modes the same afternoon.
The skeptics are taught last on purpose. Their session opens with cases where the tool was wrong, confidently, on material that looks like theirs. Then the limits: what the tool cannot see, cannot verify, and cannot be accountable for. Only after that does anyone automate anything.
The skeptics are often right. A nurse manager who distrusts an AI summary of a shift handoff is applying good judgment, and the training should say so. What changes is that the skepticism gets specific. "I don't trust it" becomes "I don't trust it for handoffs, I do trust it to draft the staffing memo, and here is how I check it."
Five failure modes account for most of it.
Long acceptable-use policies do not get read. The rules that survive are short and concrete.
If you buy training, measure it the way you would measure any other workflow change. Count the tasks automated in the room. Check at 30 days how many are still running. Ask the compliance office whether the question volume went down. If none of those numbers move, the training was a webinar with better catering.
Sources: Health-system AI adoption (roughly three quarters run at least one AI application); Budget priority shifting from new use cases to making purchased AI perform; Physician EHR time (~5.9 hours per day) as the context for "the tool drafts, a person sends".