While committees are still negotiating the framework for AI, the matter has long been settled among employees. What remains are two gaps, one in competence and one in responsibility.
This piece works through German law and German survey data. The mechanism it describes is not specific to Germany.
Practice moves faster than policy. In a survey of 1,037 predominantly desk-based employees in June 2026, just under 48 per cent reported using AI tools at work that their employer had not provided. Among them, 15.7 per cent had entered strategic information and 12.9 per cent customer data. In spring 2025, Boston Consulting Group found among a good 10,000 respondents that 54 per cent would turn to AI without their employer’s consent if in doubt. Legally, the ground is clear. In January 2024, the Hamburg Labour Court held that merely permitting the use of ChatGPT via private accounts, with no employer access to the data, triggers no codetermination at all.
What gets taught is operation. In spring 2026, the German TÜV Association surveyed around 500 companies with 20 or more employees. 56 per cent use generative AI in everyday work, while 27 per cent have trained their staff for it. That figure has more than doubled since 2024, when it stood at 12 per cent. The direction is right, but the gap remains: in every second company using the technology, nobody has prepared the people who use it.
What those courses actually cover is the real question. Look at the training on offer and you find tool instruction, prompting techniques, and data protection rules. A survey showing how many curricula deal with checking a result does not exist on either search route. That is telling, because verification is precisely where the critical competence lies. A language model generates text that sounds plausible, regardless of whether it is true. Whoever does not understand how that output is produced can neither judge its value nor spot its errors.
Checking costs more than producing. It is a question of time and cost. A draft appears in seconds; checking it takes longer than drafting it by hand would have taken. Research into automation bias has documented this pattern for years: the more useful and efficient people find a system, the more readily they adopt its flawed recommendations. Generative systems reinforce this, because fluent phrasing mimics competence.
The scale of this tendency was measured in the largest survey on the subject to date. For the global Trust in AI report by KPMG and the University of Melbourne, more than 48,000 people across 47 countries were surveyed between November 2024 and January 2025, over a thousand of them in Germany. Among German employees who use AI at work, 65 per cent say they have relied on an AI result sometimes to very often without checking whether it was accurate. Worldwide the figure is 66 per cent. At the same time, only 45 per cent of German respondents feel capable of using AI applications appropriately, and only 20 per cent have ever had any training in them.
Those affected feel the strain. A survey of 1,230 employees in October 2025 found that regular users gain an average of 5.7 working hours a week. At the same time, two in three users say the effort of checking and correcting results has risen, and more than half now experience their own work as more error-prone or less certain. Time gained and extra workload are two sides of the same coin. As occupational psychologist Patrizia Thamm puts it: if saved time is immediately filled with new tasks while expectations of personal performance keep climbing, the result can be what is known as techno-stress.
There is another side to the story. In a qualitative study by the ZHAW at the end of 2025, thirty specialists and managers who use generative AI daily reported that they experience the technology as predominantly relieving, even alongside rising pressure to be productive. Those who use it daily have learned how. That sample therefore shows what is possible, and not what is widespread.
The second gap is responsibility. Who is liable when an unchecked result causes damage? Externally, the courts have begun to answer the question. In May 2026, the Hamm Higher Regional Court held that a company is directly liable for misleading statements made by its AI chatbot. The chatbot was “(merely) a technical means” over which the defendant “held sufficient power of control”. No attribution to a third party, no detour through interferer liability: it was the company’s own conduct. The judgment is not final; an appeal to the Federal Court of Justice has been allowed. The Kiel Regional Court ruled along similar lines in 2024, and in July 2025 the Cologne Local Court held that a lawyer breaches professional duties by filing a pleading containing invented AI citations.
Internally, the matter remains unresolved. A published German employment court ruling apportioning damage between employee, manager and employer does not exist on either search route. That leaves managers in a difficult position: they are accountable for results whose derivation they can no longer assess professionally, while the internal apportionment of damage is measured by fault, instruction and training. Precisely the training that is missing in every second company.
What follows is unspectacular and unfamiliar. First, organisations must define which results have to be checked before they leave the organisation, and how that check is to be documented. Second, checking must be budgeted into schedules, otherwise the time gained becomes a trap. Third, the question of why somebody adopted a wrong result belongs to the working conditions rather than to the person, because under time pressure anybody adopts.
While formal approval procedures drag on, the matter is already being settled without them. Employees are not waiting for the works agreement. They are already working, and they are doing so with tools whose results nobody taught them to check.