AI in day-to-day operations

The available figures differ substantially, and saying so is more honest than quoting a single percentage.

The Federal Statistical Office reports 26 per cent of German companies for 2025, broken down by size: 57 per cent of large firms, 36 per cent of medium ones, 23 per cent of small ones. The industry association Bitkom reports 41 per cent for 2026.

The gap says less about growth than about what counts as “use”. Someone who once opened ChatGPT to draft an email is inside one survey and outside the other. For your own decision both numbers matter little anyway: whether AI helps depends on your tasks, not on a national ratio.

For tasks that are already named and recurring: summarising official letters, first drafts of routine text, sorting lists, preparing translations.

What these share: a clear start, a clear end, and a result one person can check in under five minutes. That is where adoption fails when it fails — not at the model, but at nobody being able to say when the output is good.

Where no human checks the result, you do not gain time. You gain a delayed source of errors.

Since 2 August 2026, Article 50(4) of the AI Act (Regulation (EU) 2024/1689) applies: anyone publishing AI-generated text that informs the public on matters of public interest must disclose it.

The duty explicitly falls away where the content has undergone human review or editorial control and a named person holds editorial responsibility. Most company sites fall under that exemption — provided the review actually happens and the responsibility is named in the imprint.

Under Article 50(5) a disclosure, where required, must be visible at first exposure at the latest. The footer is not enough. This site labels voluntarily, even though the exemption applies.

From three things, in this order.

First: customer communication without a second pair of eyes. A model invents commitments nobody made, and you find out when someone insists on one. Second: figures from a model without checking the original source. Third: introducing tools before the work step is written down — you then automate a mess and have it faster afterwards.

If something happens twice a year, automation does not pay. Setting it up costs more time than it returns.