"ChatGPT training booked, box ticked" — that's how it currently goes in many companies. Four weeks later the same three people use AI as before, and the rest never opened their access again. We've seen enough of these trajectories to know: it's rarely the employees. It's the content of the training.
The usual format is a tool demo: here's the interface, here's where you type the question, here comes the answer — impressive, right? That works as entertainment and fails as training, because it doesn't touch the actual hurdles. Afterwards people know where the button is. They don't know what the tool has to do with their work, what they're allowed to type in, and when they can trust the result. Exactly at those three points a training either changes something or it doesn't.
From the trainings we run ourselves and the ones we've readjusted at clients, a structure of five building blocks has emerged. None of them is optional — and neither is the order.
Building block 1: basic understanding — without maths
Nobody in sales needs to know what a transformer is. But everyone needs to have understood three things: a language model computes probable continuations — it doesn't know what's true. It doesn't know your internal data unless you give it to it. And it always answers, even when it has no basis. These three sentences explain ninety percent of the situations that come up later in daily work — from the invented legal clause to the question of why the model doesn't know your own price list.
We do this in twenty minutes with examples instead of architecture slides: the same question asked three times, three answers of varying quality. One question about your own company that the model confidently gets wrong. That produces exactly the right amount of scepticism — respect for the tool, no faith in an oracle.
Building block 2: prompt craft — on your own tasks
The biggest difference between trainings that work and ones that fizzle out: are generic examples practised, or the participants' real tasks? "Write a poem about summer" has never helped anyone in purchasing. The reply to the angry customer email, the summary of the forty-page minutes, the draft for the job ad — those are the tasks that decide whether someone opens the tool again on Monday.
Practically that means: before the training we collect two or three recurring tasks from each department and build the exercises from them. In the training, everyone then works on their own case — with the moves that actually change something: provide context and role, show examples, name the desired format, and follow up on weak results instead of rolling the dice again. Whoever leaves the training with three saved prompts of their own that save work tomorrow will come back.
Building block 3: data protection — what must never go into a prompt
The building block that's missing most often in tool demos — and the only one where a training can prevent damage rather than just create benefit. The rules have to be concrete, not legalistic:
- Personal data — customer names, addresses, HR files — has no place in freely available chatbots. Full stop.
- Credentials, contracts, calculations, source code with business logic: never. Once entered, control is gone.
- A private account is not a company account. The company account with a data processing agreement and no training on inputs is the foundation — the private account on the private phone is exactly the risk the training is supposed to dry up.
- Anonymising is a craft: names out, identifiers out, numbers rounded — and the task still works. You have to practise that once, then it's routine.
The organisational side behind it — DPA, EU processing, vendor checklist — is written up in a separate guide: Using AI in line with the GDPR. In the training itself the short version is enough: these tools yes, this data never, when in doubt ask this person.
Building block 4: recognising limits and hallucinations
Every participant has to have experienced once how convincingly a model talks nonsense. Telling them isn't enough — the effect only sticks when the confident, fluently phrased, completely invented answer appears on their own screen. We deliberately provoke this in trainings: questions about standards, rulings or sources that don't exist, and the model delivers anyway — complete with case numbers.
After that, the rules of thumb are no longer theory: everything that looks like a number, a quote, a legal clause or a source gets checked. The more specific the factual question, the higher the risk. And AI results are drafts — responsibility for what goes out stays with the human who uses it. That last rule isn't a platitude, it's the working principle: the model drafts, the human signs off.
Building block 5: anchoring — what happens after the training
The building block that decides over all the others and is almost always missing. A half-day workshop without follow-up is a firework: nice to look at, forgotten after a week. What actually carries the usage is unspectacular: a named contact person per team that questions may go to. A shared prompt library with the best internal use cases — maintained, not a data graveyard. And after four to six weeks a short second round: what worked, where does it stick, which new cases have come up.
Format and groups: three short sessions beat one big one
On the format, because the question always comes up: we've had the best experience with half a day to start and two short follow-up sessions four weeks apart each — not with the one big training day after which everyone is exhausted and nobody is changed. Twelve participants is a good upper limit; above that, practising turns into spectating. We mix groups by tasks, not by department or hierarchy: the clerk in purchasing and the colleague from accounting have more similar use cases than two people from the same team with completely different jobs. And the leadership circle gets its own session — not because they're more important, but because different questions sit there: what do we permit, who signs off, what do we measure. Put managers and their team into the same practice round, by the way, and you'll reliably get the silence of the cautious — that too is experience from practice.
How to measure whether the training worked
Not by the satisfaction form at the end — that measures mood, not effect. Other things are measurable: how many people actively use the company access after four and after eight weeks? Is the prompt library growing, or has it been frozen since the training? Are the questions in the second round getting more specific — from "what can this do?" to "how do I get the quote into the right format?" And the most honest signal costs no analysis at all: colleagues start sending each other prompts. Then the tool has arrived.
The duty on the side: Article 4 EU AI Act
Since February 2025, Article 4 of the EU AI Act obliges companies that use AI to equip their staff with sufficient AI literacy. Soberly assessed, that's less dramatic than it sounds: there's no prescribed certificate, no number of hours, no official inspection body. What's required is competence that fits role and use — and whoever provides it should be able to document it: who was trained on what, and when. A training with the five building blocks above fulfils the duty along the way; whoever wants the assessment of the whole law will find it in our overview of the EU AI Act for mid-sized companies. And as always with legal topics: this is hands-on experience, not legal advice.
If you don't want to build all this yourself: exactly this kind of training — on your tasks, with your data rules, with follow-up instead of fireworks — is one building block of our AI consulting. But the most important advice holds regardless of who trains: practise on real work. Everything else is cinema.
