How do I train staff in AI who do not speak German?

Through an honest assignment first, and through guided training in small groups for whoever is left. The obligation under Article 4 of the AI Act calls for measures to promote AI literacy, and a measure the addressee cannot understand linguistically is not one. A training clicked through in a language somebody does not read produces a certificate, not competence.

In a manufacturing company in Lower Bavaria with 180 employees that quickly amounts to 40 to 60 people whose first language is Polish, Romanian, Ukrainian or Turkish, and whose German is enough for the shift handover but not for a text about hallucinations in language models.

This article sets out how many of them need training at all, which routes work without foreign-language content, and how the record comes about along the way.

How many of my foreign-language staff need training at all?

Only those involved with the company’s AI systems. Article 4 has applied since 2 February 2025 and turns on actual involvement, not on headcount. Somebody operating a press with no AI in it is out of scope, whatever their language.

This calculation is worth doing before anyone thinks about translations. Out of 50 foreign-language employees, a clean assignment often leaves 8 to 15, because AI functions in manufacturing today sit mainly in quality inspection, maintenance planning and shift scheduling rather than at every workstation. Guided training is feasible for 12 people; for 50 it becomes a project.

The sentence that does not hold up here is “our foreign colleagues don’t use AI”. Anyone operating an AI-assisted pre-assessment or following its outputs is involved, even if they work the interface purely through icons. So the assignment has to be made at the workstation and not from the HR list.

Which languages is the content available in?

The courses of the PASSION4IT Academy are in German by default. We evaluate content in additional languages on request, which is not a commitment out of the catalogue. Anyone with a larger group sharing one first language should raise it in the first call, because the question can then be costed concretely.

That is the honest state of things. For a company that has to train 12 people next week, the answer to the language question is not translating the catalogue but a different format.

Which routes work without foreign-language content?

The guided small group is the most robust. A supervisor or a colleague with the same first language works through the units together with four to six people, pausing and explaining in the language understood in the room. The Academy units are cut to 20 to 30 minutes and can be paused for this. In terms of time you land at two to three sessions per group.

In many companies that person already exists, because they hold the same role for occupational safety and fire protection. The logistics are established, the time slots are accepted, and attendance appears in the shift plan. Attaching the AI units there means no new process has to be invented.

The second route is the work instruction in the local language for the specific use case. Rather than translating the general AI training, you translate the single page saying what is permitted at this machine, who gets told about an unusual output, and which data must not go into an AI tool. That is the information that matters day to day, and it is short enough to have translated properly.

What does not work is the machine-translated subtitle as the sole solution. With technical terms from the AI Act it quickly turns into something else, and the company ends up with a record it cannot proofread itself.

How does the record come about when the training is guided?

Through group attendance plus the individual record in the platform. A robust record consists of the assignment of which roles work with AI, the training delivered with a content overview and per-person proof of participation, and a date for the review. With a small group, the minutes are added, stating who led the training and in which language it took place.

That addition is the point at which an account convinces or does not. A company that has documented 12 Romanian-speaking employees being guided through two sessions by a foreman in their language has a better basis than one presenting 50 German-language certificates and unable to say, when asked, whether the content was understood. How such a record is built is set out in How do I document AI training for the regulator?.

The licence is booked per user, there is no minimum quantity, and the progress view supplies per-person proof of participation even when the unit was worked through jointly.

What does it cost?

59 euros per person per year for the AI module with the AI licence at list price. For 12 assigned people that is 708 euros a year. Add the working time for the guidance; at two 90-minute sessions for three groups you are at around nine hours for the person leading them.

The comparison that carries the decision is not Academy versus classroom seminar. It is guided small-group training for 12 assigned people against translating an entire learning path for a workforce two thirds of which is not subject to the obligation at all.

Conclusion

The language question comes in the wrong order if it precedes the assignment. First establish who actually works with AI; the group is then usually small enough that guided training in the language understood is the faster route than a translation.

To sort this out for your own company, walk the machines with production management and mark the ones where an AI function is active. After that you can say how many people are affected and whether guidance is enough or a language variant should be evaluated.