Experiment first, strategy later: How AI is entering the workplace

New AI models have arrived over the past three years faster than most employers could evaluate them. Inside the workplace, the picture looks very different. Research tracking employees in Germany since 2024 finds AI use at work climbing slowly, unevenly, and to a considerable extent without employers’ knowledge.

The Konstanz AI study, run by the Future of Work Lab at the University of Konstanz, has administered the same questionnaire to employees across three waves. Its latest findings, from 1,105 respondents surveyed in May 2026, show workplace AI use rising from 24% to 35% in 2025, then to just 38% this year.

“What I find particularly striking is the difference in speed,” Florian Kunze, Professor of Organisational Behaviour at the University of Konstanz and Head of the Konstanz Future of Work Lab, who led the study, tells HRM Asia. “The AI technology itself has developed extraordinarily quickly, but its diffusion into everyday work is much more gradual. We seem to be moving from an initial experimentation phase into a much slower phase of organisational absorption.”

What separates the two, he suggests, is rebuilding jobs around a tool once it has been acquired. “Introducing a new AI model is relatively easy; changing work processes, developing skills, setting rules, and deciding where AI creates value takes much longer.”

He sees a lesson in how the current debate is framed. “We often look at the technological frontier and assume that organisations are moving at the same speed. Our data suggests that they are not.”

The tools employees brought themselves

Where employers have been slow, employees have not waited. Among AI users in the latest wave, roughly 45% said the tool they reach for most was never officially introduced by their employer.

“Employees are not simply waiting for organisations to develop an AI strategy; they are experimenting themselves,” Kunze says – and he is careful not to cast that as failure. “That has a positive side. It shows curiosity, initiative and potentially a very powerful source of bottom-up innovation. Employees often know particularly well where AI can remove friction from their own work.”

The problem lies in what employers cannot see. “If organisations do not know how their employees are using AI, they cannot adequately address issues such as data security, quality, compliance or unequal access. They may also miss opportunities for mutual learning and to scale applications that are already working well.”

His prescription is not a crackdown. “I would not recommend replacing bottom-up experimentation with a completely top-down approach. The challenge is to connect the two: make existing use visible, which employee-driven applications are worth developing systematically.”

An uneven spread

Access to AI, meanwhile, is following lines that are already familiar. The study covers Germany alone, and Kunze is explicit that its findings cannot be extended elsewhere on this evidence. He expects the underlying mechanism to be widely shared.

“Technologies can arrive almost overnight, while institutions struggle to adapt. Organisations need to develop skills, routines, governance structures and sometimes entirely new job designs. Education and training systems have to adjust as well. That inevitably creates a period in which technological possibilities run ahead of institutional adaptation.”

How wide that lag grows, and how long it lasts, is where he expects countries to diverge. Company size emerges from his data as one of the sharpest determinants: in the latest wave, just 11% of employees in small organisations reported any AI training, and around 10% worked under binding rules for AI use. “Smaller organisations are much less likely to provide AI training or have binding rules for AI use, as they have limited resources compared to large organisations to support such transformations.”

For economies built largely on smaller employers, the implication is direct. “I would therefore expect labour markets with a particularly large SME sector to face a greater risk of uneven adoption unless there are external support structures,” Kunze says. National training infrastructure counts too: “Systems that can update occupational skills quickly may help diffuse AI beyond highly educated knowledge workers, whereas weaker training infrastructure could reinforce existing divides.”

Within his own data, the separation is widening. AI use among office and knowledge workers climbed from around 30% in 2024 to 49% in 2026. In production and more physical occupations, it rose as well, but only from roughly 17% to 25%. By education level, the 2026 figures are even starker: 56% among highly educated employees, versus 21% among those with lower qualifications.

“Both groups are adopting AI, but the gap has not closed; if anything, it has become larger,” Kunze says. Some of that reflects where today’s tools are simply easiest to apply – text, analysis, communication. What worries him is the compounding. “Inequalities can become self-reinforcing. Employees who already use AI gain experience, receive more training and discover more applications, while those without access fall further behind.”

Employees’ own reading of the change carries its own tension. Around 40% expect AI to damage employment more broadly; only 17% believe their own job is at risk. Kunze declines to call that inconsistent. “People may believe that AI will substantially restructure the labour market without expecting their own job to disappear. They may anticipate that other occupations, entry-level jobs or particular tasks will be more strongly affected than their own position.”

Few employees, he observes, have yet met AI as outright substitution. “For many, it is still a tool that changes particular tasks rather than replacing their whole job.” For him, that argues for reframing the question entirely. “This is also a warning against reducing the AI debate to the question, ‘Will AI take my job?” The more relevant transformation may initially be more incremental and should focus on which tasks disappear, which become more valuable, how entry routes into occupations change, and which new skills employees need.”

What actually makes a difference

Asked what distinguishes employers getting ahead of these shifts from those reacting to them, Kunze put visibility first. Organisations “should understand where employees are already using AI, rather than assuming nothing is happening until an official tool is introduced” – which in turn demands “an open culture in which knowledge and experience with AI are openly shared”.

Rules follow, and their purpose matters. “They need clear rules that enable rather than simply restrict use. Employees need to know which tools they can use, what kinds of data they can use them with, and where human verification remains essential.”

READ MORE: A generation is arriving fluent in AI. Nobody has worked out how to train them

On training, he is blunt about what fails. “Training has to become much more systematic. And I would emphasise job-specific training rather than generic courses about AI. A production employee, an HR manager and a software engineer need very different forms of AI capability.” Employers must also attend deliberately to the groups furthest from access, he argues, or “existing inequalities by occupation, education and organisational size may become an AI divide.”

His last point concerns measurement. “The relevant question is now how many employees have access to a chatbot. It is whether AI improves productivity and work quality, which tasks are changing, and whether the benefits are reaching different groups of employees.”

That leads him to a conclusion which lands the responsibility somewhere specific. “The organisations that get ahead will probably not be those that adopt the largest number of AI tools. It will be those that become good at turning experimentation into collective organisational learning,” Kunze says. “The HR functions in organisations, as the promoter of organisational learning and development, have therefore a key responsibility in shaping a successful AI transformation.”