AI training for managers: where to start

The usual story goes like this. The team gets a one day prompting workshop, everybody leaves pleased, and three weeks later nothing has changed except that two people now paste emails into a chatbot and nobody knows what data is going in. The workshop was fine. Nothing changed because it taught a tool rather than the work.

This piece is about what a manager has to settle before the team starts at all.

Three things worth remembering

1. 1.Start by mapping where the time actually goes, not by choosing a tool.
2. 2.Pick one process, not ten. Ten means none of them gets finished.
3. 3.Before anything goes live, decide who checks the output. Without that, adopting AI is just moving responsibility onto software.

Why a one day prompting workshop is not enough

Because it answers a question the team never asked. People leave able to write instructions and return to a job where nobody has decided what those instructions may be used for, on which data, who checks the result, and what happens when the result is wrong. A skill without a process does not change the process.

The second reason is simpler. An hour of training does not compete with habit. Unless the change is built into how work is handed out and handed back, the team will be back to the old way within a week, because the old way is safer.

Step one: map where the time actually goes

For two weeks, ask the team for one thing: a note whenever a task takes longer than it should and repeats. This is not about measuring to the minute. It is about producing a list. It usually comes to fifteen or thirty items, and the same things end up at the top: writing similar emails, retyping data between systems, summarising long documents, preparing the same reports, hunting for information that already exists somewhere.

That list matters more than any workshop, because it shows where change is worth attempting at all.

Step two: pick one process

A good first process has four properties:

• •it repeats at least a few times a week
• •it has a predictable shape, so you can describe what a good result looks like
• •mistakes in it are visible and reversible
• •it does not involve data that must not leave the building.

That last point rules out more candidates than people expect, and it is better checked at the start than three weeks in.

Step three: settle four things before the team begins

What data may be entered. Write it in one paragraph and put it where people work. A three page policy will not be read.

Who checks the output. For every process, one named person is answerable for what leaves the building. Not the team, not the system. A person.

What happens when the output is wrong. Where it gets reported and what changes so it does not recur. Without this the same errors repeat and the team stops trusting the whole idea.

What we do not hand to a machine. Decisions about people, performance assessment, recruitment, anything touching somebody's circumstances. That is also the area where the rules are strictest.

What a manager needs to know about the rules

You do not need to be a lawyer, but two things concern you directly. Since 2 August 2026 the transparency duties in Article 50 of the AI Act have applied, which among other things means a person should know when they are talking to a machine. The AI literacy obligation also already applies, although Regulation (EU) 2026/1744 softened it and postponed the core requirements for high risk systems.

The practical conclusion for a team is short: a customer facing chatbot should say it is a chatbot, and the team should understand the tools it uses. I have written about the detail separately, in the piece on the AI Act after 2 August 2026. EU AI Act after 2 August 2026: what companies must do

A twelve week plan

Weeks 1 and 2. Map the tasks and choose one process. Settle the four points above. Measure the starting point, for example how long the weekly report takes today.

Weeks 3 and 4. A live workshop, but on your own tasks rather than slide examples. The team leaves with a working method for that one thing.

Weeks 5 to 8. Run the process with the named checker in place. Improve the method weekly. This is the stage missing from one day workshops, and the only one where a change actually sticks.

Weeks 9 to 12. A second process, if the first one works. Write the rules down so a new joiner can read them and apply them.

How to tell whether it is working

Measure the time on a specific task before and after. Measure how many outputs need correcting. Measure how many people are still using the new method after eight weeks, because that is the real adoption figure.

Do not measure the number of queries sent to a tool, or the number of people trained. Both rise on their own and neither tells you whether the work is faster.

The most common mistake

Starting by buying a tool for the whole company. A licence for a hundred people is easy to purchase and hard to defend six months later when twelve people use it. The reverse order works better: one process, one group, a measurable result, and only then scale.

If you would rather run this with somebody from outside

I run AI adoption in organisations: a readiness audit, a tailored adoption roadmap, live workshops including hands on prompt work, team training and support while the processes go in. Up to twelve weeks, from 9,900 PLN. I work on your tasks rather than slide examples, and I leave written rules behind that stay in the company afterwards.