Training for all staff: what we'd recommend and how to roll it out

Not a course. Two hours of hands-on practice on each person's own work, a buddy for a month, and a monthly hour where people show what they did. The structured programmes mostly train people on tasks they don't have.

By Matthew Steele4 min read28 October 2026

A small training session in a meeting room, four staff with laptops working on their own tasks while a colleague crouches beside one of them to help
Illustrative scene, not a customer's premises.

Why the generic course fails

It teaches prompting tricks on examples nobody has, to people who'll have forgotten them by the time the real task turns up. It happens once. There's nobody to ask the week after. And it treats a sales rep, a bookkeeper and a warehouse supervisor as if they had the same job. Completion rates look fine, and usage a month later is a third of the seats. We've seen it often enough to be blunt about it.

The rollout that works

When What Who
Week 1 Leadership uses the tools themselves for a fortnight before anyone else gets trained. Visibly. Owner, directors
Week 2 The one-page rule: what never goes in, who to ask. Sent to everyone. AI lead
Weeks 3 to 6 A two-hour session per team, on that team's real tasks, run by a colleague who already uses the tools. AI lead plus a user from each team
Weeks 3 to 10 A buddy per team – the person who ran the session, available for the month after it. Buddies
Monthly, ongoing The show-and-tell: 10 minutes each on what I did with it this month. Everyone
Day 30 after each session Ask what are you using it for, and count the answers. AI lead

Leadership going first isn't for show. A team trained by a manager who's never used the tool learns that it's optional.

What goes into the two hours

Everyone brings three real tasks from their own week. For the first hour they do those tasks with the tool, together, with the buddy walking round. The second hour covers 4 things: the never-list and why it exists, how to brief the tool the way you'd brief a person, how to check an answer and when not to trust one, and where to ask when you're stuck. No slides about what AI is. Nobody needs them, and the people who do will pick it up from the task.

Four levels of needFour steps rising from left to right. Everyone: a two-hour session on their own tasks, the one-page rule, a buddy for a month. Regular users: the show-and-tell and shared briefs. Builders, two to five people: a structured course on building and testing agents. Owners: the owner's card. Training the whole company as if it were the top step is the expensive mistake.EveryoneTwo hours on their own tasks.The one-page rule. A buddy.Regular usersThe weekly show-and-tell.Briefs that worked, shared.BuildersTwo to five people.A structured course onbuilding and testing agents.OwnersThe owner's card: the sample,the rules, the off switch,the monthly line.Who needs whatEveryone gets the first step. A handful get the top two. The generic course for all is training the whole company on the top step's problems.
There are four levels of need. Everyone gets the first one, a handful get the top two. Training the whole company as if it were on the top rung is the expensive mistake.

The builders' track

A few people – the 2 to 5 from who may build live tools – need something different: how an agent is put together, how to write and run test cases, how connections and permissions work, and what the three tiers mean in practice. That's worth a structured course, paid for, and worth the builders' time away from their desks. We'd choose one that makes them build and test something real rather than watch videos about it, and we'd make the owner's card and the 5-case acceptance test part of what they come back knowing.

The agent training the question mentioned

We can't speak for whoever was asked that at the event. What we can say is what the training for the people who build and own agents should cover, and it's the track above: building, testing, permissions, ownership. The important part is that it's for a handful of people, not everyone, and that it ends with each of them owning something that's live.

How you know it worked

Not completion. At 30 days, ask each person what they used the tool for last week and count the answers by team. A team where half the people have a real answer is trained. A team where nobody does needs the session again, run differently – or has no tasks the tool helps with, which is also worth knowing.