AI training for employees that actually sticks

Why most AI training fails

The standard version is an hour in a meeting room, someone demonstrating impressive things on a screen, and a slide of prompt tips at the end. Everybody leaves interested. Nothing changes, for three reasons.

First, the examples aren't theirs. A demo about writing a marketing email means nothing to a bookkeeper. Second, nobody typed anything. Watching someone else use a tool builds no skill at all, the same way watching someone drive doesn't. Third, there's no permission. Staff quietly assume they'll be in trouble if a client finds out, or if it makes a mistake with their name on it, so they don't touch it. Until somebody senior says out loud what's allowed and what isn't, the safest choice for an employee is to do nothing.

Generic online courses have the same problem in a longer format. They teach the tool, not the job. People finish the modules, tick the compliance box, and go back to working exactly as before.

A format that works

Run it by role, not by department and not all at once. Before each session, collect three tasks that role does every week and that a rough first draft would help with. Sales brings a proposal and a follow-up email. Admin brings an invoice chase and a client update. Operations brings the SOP nobody has written.

In the session, everyone has a laptop open and does the work. You demonstrate one task, then they do the next two on their own files while you walk the room. Bad output is the most valuable thing that happens: it's where people learn what context they left out. Finish by having each person write down the one thing they'll use this week, and where it will go wrong.

Then the part everyone skips. Two to three weeks later, half an hour with the same group: what did you actually use, what did you drop, what annoyed you. That follow-up is where a training day turns into a habit, and it costs almost nothing.

What to actually cover

• Context, not prompt tricks. The single biggest quality jump comes from giving it the audience, the constraints and an example of good output. Everything else on the prompt-tips slide is noise next to that.
• Judgment and checking. Teach people that it's confident when it's wrong, that facts and figures must come from your own systems, and that whoever sends the email owns the email. This is the part that keeps you out of trouble.
• Data rules, in one page. What may be pasted in, what may not, and which account to use. Client identifiers, payroll, health information and anything under a confidentiality clause need a clear line drawn. Business and enterprise workspaces keep your content out of model training by default; personal accounts do not unless someone changes the setting.
• Where it doesn't belong. Regulated advice, final legal wording, hiring decisions, anything client-facing that nobody reads first. Naming the limits gives people permission to use it everywhere else.
• Their three tasks. Not a tour of the features. Three jobs from their week, done properly, with the shortcuts written down afterwards.

If you want a starting shortlist of tasks per function, the ChatGPT use cases for business guide is written for exactly that.

Measuring whether it stuck

Attendance and satisfaction scores tell you nothing. Four things worth tracking instead: how many people opened the tool in the last seven days, how many of the tasks named in the session are still being done that way after 30 days, the time on one specific named task before and after, and a quick quality spot check on the output going out the door.

On a business or enterprise workspace, the admin console gives you the usage half of that directly. The rest is a ten-minute conversation with each team lead. If usage in a role has collapsed by week four, the answer is almost never more training. It's that the tasks chosen for that role were the wrong ones.

Where I fit

I don't run classroom courses. AI Consulting is A$1,195 per month + GST, with a call each fortnight, clarity on what to build in Claude or ChatGPT and practical guidance on how to do it. No lock-in.

In practice, owners and managers use that time to work out which roles to train first, which tasks to put in front of each one, what the data rules should say, and what to do when adoption stalls in one team. Sessions run over video, so where you are in Australia makes no difference. The point is to decide what to do, in what order, and what to avoid, before you spend money on a programme that teaches the wrong people the wrong things.