How to make AI training stick after the workshop ends

On-Off Group

September 2026

To make AI training stick, pick one real task people do every week, measure how long it takes before the session, train on that live work rather than sample prompts, then check the same number two weeks later and fix whatever blocked people.

You book AI training, people turn up, the feedback is good, and by Monday nothing has changed. Then your boss asks if it did anything.

Most of the work is before and after the session: what you pick to train on, and what you measure two weeks later. Below is the order we run it in.

None of it needs a new budget line. It is one task picked, one number taken, and a check booked in the diary.

Pick one real task people already do every week

The usual starting point is a workshop covering prompting, summarising, drafting, analysis and a bit of image generation. Everyone leaves impressed and nobody has a task they were going to do that afternoon.

Instead, pick the first real task you will use AI on. One task, done by most of the room, at least weekly. Complaint responses. Credit memo first drafts. Turning branch staff feedback into a summary for the weekly meeting. Board pack notes.

It has to be work that already exists, with a person waiting on the output. That is what makes anyone open the tool again the following week.

We wrote about the same problem from the licensing side: Instead of buying AI licences for everyone, train one real task first. Buying licences for everyone and hoping is the expensive version of the same mistake.

Take a baseline number before the session

This is the step people skip, and it is the reason the question "did it work" gets a shrug.

Pick one number attached to the task. Turnaround time on a complaint response. Number of drafts before sign off. Hours a week spent pulling the report together. Take it the week before training, from actual work, not from a survey asking people to estimate.

Ten samples is plenty. Write the number down somewhere your boss can see it later.

In design sprint work with a large insurer, the before number was the only reason anyone could prove the result. Customer journeys that used to take six months or more went live in four weeks, and completion rates on the new journeys rose by 80 per cent. Neither of those numbers means anything without the before figure. Training is no different.

If the number is small when you measure again, you say it is small. That is still better than a shrug.

Train on live work, with their own files open

If the session only uses sample prompts, people learn nothing about their own files. Then they go back to a messy spreadsheet with three tabs of branch data and it does not behave.

So bring the real thing into the room. Real complaint text with names removed. Real report. Real policy document. People build and test a prompt on their own work, fix it, and leave with something saved that they will use that week.

Bringing real work in also changes who should be in the room. Put the people who do the task and the person who signs off the output in the same room. Otherwise the "compliance will not allow this" conversation happens later, quietly, and nobody tells you. Ask the provider to build the session around your own documents and your own sign off rules.

Attendance being compulsory does not help either. If people are only there because they were told to attend, nothing changes on Monday.

Check the first real piece of work after the room empties

A full room and a good feedback score tell you people showed up, but that's it.

The real check happens the week after. Did anyone open the AI tool for the task you picked? Did it help, or did they quietly go back to the old way? The certificate tells you nothing.

So book the check in before the training happens. Two weeks later, same task, same number. Plus five minutes with three people who did the work: what did you try, where did it break, what did you go back to doing by hand.

Expect some of them to say they could not get access to the tool, or were not sure if they were allowed. Those are blockers, not training problems, and you only hear about them if you ask.

The AI prompting training and the two week check are described at onoffgroup.com.

Fix the blockers, then widen it

After the two week check you will have three or four reasons people stopped. Usually the same ones: no licence on their machine, no clear answer on what customer data can go into the tool, a manager who still asks for the old format, or a task that was never really weekly.

Fix those before you train the next group. If the next group hits the same wall, word gets round and the training is the thing people blame.

Only then widen it. Open the next session with the prompt that worked and the person who used it, with their before and after number. People believe a colleague with a before and after figure more than they believe a trainer.

Keeping the skill inside the team afterwards is its own problem, and we have written about it here: UX agency or in-house team: how to keep the skill after the project.

What to do this week

Name the one task the team will use AI on first, and check it happens at least weekly.
Take a baseline number on that task from ten real examples, and write it down.
Ask the training provider to build the session around your own documents, with the approver in the room.
Put the two week check in the diary now, same task, same number, plus three short conversations.
List anything that would stop someone using the tool on Monday, and sort access and data rules before the session, not after.

Questions people ask

What baseline number should we take before AI training?
Take whatever the task already produces: turnaround time, number of drafts, hours per week. One number per task is enough, taken the week before the session, from real work rather than a survey.

How soon should we check whether the training worked?
Two weeks after the session, on the same number you measured beforehand. That is long enough for the first real piece of work to come round and short enough that people still remember the session.

Does feedback score tell us if AI training worked?
No. A full room and a high score tell you people showed up and enjoyed it. The test is whether anyone opened the AI tool for their first real task afterwards, and whether it helped.

What if the number does not move at all?
Say so and find the blocker. Usually it is a tool nobody has access to, a policy nobody has read, or a task that was never the right one to start with.

On-Off Group trains teams, tests products with real customers, finds where a transformation has stalled and builds what gets it moving, for banks, insurers and enterprises in the Philippines, since 2015. How we help with ai training.