Teach AI with your own work, not sample exercises.
A typical hands-on AI session hands people a practice file: a fictional company's meeting notes, a made-up customer complaint, a sample spreadsheet with tidy columns. Everyone gets a decent result by the end of the hour. Then they go back to their desks, open the actual quarterly report or the actual intake backlog, and can't see how the exercise applies.
The fix is to run the session on the work itself. Bring the real inbox thread, the real proposal template, the real policy manual. It takes more preparation and a little care with sensitive information, and in our experience it is the main difference between a session people remember and one they forget by Friday.
Practice files teach the wrong lesson.
Sample exercises are designed to work. The text is clean, the task is well defined, and the answer is known in advance. That teaches people that AI is good at clean, well defined tasks with known answers, which is true and not very useful, because almost nothing in an ordinary week looks like that.
Real documents carry the problems people actually need help with. The vendor invoice is a scanned PDF with a handwritten note in the margin. The board packet is forty pages and only six of them matter for Thursday. The intake form has a free-text box where clients write whatever they like. When a team works through those in the room, they learn where the tools hold up, where they slip, and what kind of checking the output needs. They also learn how to phrase a request about their own material, which is a different skill from following a trainer's prompt about someone else's.
The first real result does the persuading.
Something shifts when a person watches AI produce a usable draft of a document they personally write every month. A school district's transportation coordinator who sees a rough version of the parent notice come together from last term's notice and the new route list doesn't need a speech about productivity. She has the evidence on her screen, in her own format, with her own names and dates in it.
When the draft on the screen is a document someone writes every month, nobody has to argue that the tool is useful.
That result also comes with a natural next step. Because the work is real, the draft from the session can be finished and sent. People leave with one task already done and a working pattern they can repeat next week.
Use the session to find the mess.
Training on real work exposes problems that a sample file hides, and that is part of the value. A clinic's front desk team might discover in the session that the referral letters they want summarized live in three different places, or that two people use different templates for the same form. Those are process findings, and they surface fastest when people try to hand the work to a tool.
Keep a running list during the session of what goes wrong and why. Some items will be prompting habits people can fix on the spot. Others will be gaps in how the organization stores or formats its documents, and those belong with whoever owns the process. A facilitator who writes the stumbles down and comes back to them gets far more out of the hour.
Real work, handled safely.
Working on real documents raises a fair question about what goes into the tool. The answer starts with whatever your organization's approved tools and settings allow. If staff are cleared to use a particular assistant with client files under a signed agreement, the session can use that. If they aren't, the session should not quietly expand what's permitted.
Where the material is sensitive, scrub it before the day. Replace client names, patient details, account numbers and addresses with placeholders, and do it in a copy, never the original. Keep the structure and the mess intact: the odd formatting, the long rambling paragraphs, the inconsistent dates. A scrubbed document that still looks like the real thing teaches nearly as well as the original, while a cleaned-up one turns back into a sample exercise.
It also helps to choose material with the lowest stakes that still feels real. Last year's annual report, an expired request for proposals, an internal procedure that's due for a rewrite, a batch of old customer questions with identifying details removed. These carry the texture of the work without putting anyone's private information in front of a room.
A prep checklist for a hands-on session.
Run through this a week or so before the session, with someone from the team that will attend.
- Ask each participant for one task they do at least monthly that involves reading, writing or sorting text. Pick three or four that several people share.
- For each task, collect two or three real examples: the inputs people start from and a finished version they consider good.
- Check which AI tools and settings are approved for this material. Write the answer down and say it at the start of the session.
- Make copies of every document and scrub them. Remove names, contact details, account and case numbers, health or financial details, and anything covered by a confidentiality agreement.
- Open each scrubbed copy and confirm it still looks like real work. If the scrubbing flattened it into something tidy, put some of the mess back with invented details.
- Try each task yourself in the approved tool before the day. Note where the output needed fixing so you can point people to the checking step.
- Prepare a short handout for each task with the starting request you used, the kind of corrections it needed and one variation to try.
- Plan time at the end for each person to start one piece of their own live work, using approved material, so the session ends with something they'll actually send.
- Arrange a follow-up two or three weeks later to hear which tasks stuck and which ones people dropped.
