Executive Summary

Key Figures

Source: Konstanz AI Study 2026, third wave (May 2026, 1,105 employees) — press release PI Nr. 83/2026

The Future of Work Lab at the University of Konstanz has followed employees in Germany with the same questionnaire since March 2024. 2,019 people answered the first wave, 1,024 the second in May 2025, and 1,105 this third one. The panel provider Bilendi assembled the sample by calling back earlier participants and adding new respondents, and the funding came from the German Research Foundation's Cluster of Excellence "The Politics of Inequality." Professor Florian Kunze, Carolina Opitz and Elena Gerdiken wrote the results report.

The word fixed in the report's title is slowed. Use of AI at work went from 35% to 38%, which is 3 points in a year. Outside, models turned over several generations in the same stretch. The press release said this is not the stage at which one can speak of AI being anchored across working life, and German HR outlets ran the story under headlines about no upheaval yet. The survey's headline stops there.

Before reading 3 points as a trend, the sample deserves a look. The third wave refilled itself by calling back about 600 participants from the second and adding roughly 400 new ones, and the occupational mix moved in the process. Production and manual work went from 28% in 2025 to 37% in 2026, while office and knowledge work fell from 62% to 55%. The research team nonetheless records that use rose by 4 points in each of the two groups. Four points in each group, three points overall. That looks related to a sample where the lower-using group grew heavier. The report does not draw that connection; this article does.

One more question sits in the same survey. Employees who said they use AI were asked whether that most-used tool was introduced officially by their employer. 55% said it was. The press release put the result this way. AI spreads in many cases not through organizations introducing it, but through employees carrying it into the company themselves.

The report never writes 45% itself. The base and the item are stated clearly enough that turning the figure over carries little risk. The 55% comes not from all employees but from those who say they use AI, each answering about the single tool they reach for most. Once people who mix company tools with personal ones enter the count, the number touching an unofficial route grows further. Converted to all employees it comes to 45% of 38%, roughly one in six. That multiplication is not in the report.

The report does not leave the 55% standing alone. It sets another value beside it. 24% of all respondents say their company has AI tools of its own making, which leaves three in four with none. That does not rule out external tools the company licensed, but the research team ties the two values together and describes formal and informal routes running side by side. The pattern already has a name. A 2025 working paper from Germany's ifo Institute, cited by the research team, calls it dual diffusion: employer-driven and employee-driven adoption opening at the same time. A count that follows one route only loses the other half from the statistics.

Employees bringing a tool in first is not bad news by itself. New technology has always spread that way, and the people doing the work commonly try something before the organization grants permission. The question is what sits on that route. The survey found its answer in smaller organizations. 11% say they have had AI training, 10% say binding rules for use exist. Nine in ten use the tools with no training behind them and no sentence to follow.

The press release and the English coverage stop here and say only that larger organizations do better, but the results report carries the figures for every size band. The research team split organizations into four. 49 employees or fewer is small, 50 to 249 is medium, 250 to 1,000 is large, and above 1,000 sits the last band. In that order, the share reporting binding rules for AI use runs 10%, 24%, 29% and 31%. Training is 11% in small organizations, about 26% in large ones and about 29% above 1,000 (the report gives no figure for the medium band). Having an in-house AI tool follows the same climb, at 11%, 22%, 29% and 36%.

The top of the ladder is 31%. Even where more than 1,000 people work, fewer than one in three say their company has settled rules for AI use. The report attaches its own caveat at this point: such rules are not yet established across the board even in larger organizations. Reading the rules gap as a small-company matter is hard for that reason. Small organizations are merely the lowest, and the highest does not reach a third.

Usage rates themselves also diverge sharply by occupation. 49% of people in office and knowledge work use AI, against 25% in production and manual work. By education the gap is nearly threefold, 56% among the highly educated and 21% among those with less education. AI is not arriving at the same speed for everyone, and Gerdiken described the distribution this way: "Employees in knowledge-intensive occupations, those with higher levels of education and those working in larger organizations are currently benefiting most from the new opportunities and advancement potential that AI offers. Without targeted support measures, there is a risk that existing inequalities will become further entrenched."

Kunze, who leads the study, summed the situation up: "In many organizations, the major AI revolution has not unfolded as a planned transformation process so far. Instead, many employees are experimenting with AI, while employers are still lagging behind when it comes to establishing clear policies, providing training and implementing secure AI solutions." The recommendations the research team published follow that same order: clear rules for safe use, retraining attached to the work itself, support shaped for small and medium firms, and AI applications that reach occupations outside office work.

A distinction matters here. No rules means no prohibition, which is not the same as freedom. From where an employee stands, nobody has said what may go in and what may not. If something goes wrong, the structure also leaves whoever made the call alone with it. The 11% training figure points at companies that have not laid a path, not at careless people.

The phrase shadow AI conjures an employee using something in secret. This survey shakes that picture as well. A question asked for the first time in the third wave covers how openly people talk about their own AI use. 9% of AI users said they hold back or feel uncomfortable about disclosing it. Split by occupation, 51% in office and knowledge work tell colleagues and managers about their use, against 28% in production and manual work, and the reluctant share splits 7% to 16%. Companies are in the dark not because many people hide, but because nobody has made a place for saying it.

Taken alone, one survey can read as a peculiarity of the German labor market. Around the same time, two surveys built on entirely different designs pointed the same way. Their samples and questions differ, so the figures should not be lined up and added or subtracted. Direction is the thing to read here, not the size of the numbers.

The survey PagerDuty published on June 11 was run by Wakefield Research in April among 1,250 office professionals in Australia, Japan, the UK and the United States. All of them work at companies with revenue above 500 million dollars, and IT and technical roles were excluded. Among those who had used an AI tool, 66% said they used one while believing company policy did not allow it. In organizations above 1,500 employees the figure rises to 72%. 88% had put work-related information into a public AI service, and within that group 34% had entered customer information and 31% sensitive business documents.

The report Teramind released on June 17 comes from another direction. It pairs a survey of 300 corporate information security leaders with behavioral records, and it puts 67% of AI use inside companies on personal accounts running on company devices. Executives came in at 69% on putting speed ahead of security, frontline staff at 37%. That runs opposite to the assumption that rule-breaking belongs to junior employees.

The three surveys form a single picture when set side by side. Konstanz asked academically about the route of adoption, PagerDuty asked office workers whether they had gone against the rules, and Teramind asked security staff what they can see. Three different questions, and the answers point to one place. The tools are already inside, and the company does not know what those tools took away.

Data lineage is the record of which data came from where, passed through what, and ended up used in what. It serves as the evidence when the source of training data comes into question, and when someone checks how far customer information traveled. A tool the company introduced has a place to attach that record. Terms for handling data go into the contract, access logs accumulate, retention periods and deletion procedures get settled.

A file that went in through a personal account skips that place. The document pasted in to summarize a meeting, the name and phone number entered to polish a reply to a customer, the block of code: once they cross the company boundary, nobody at the company can confirm where they were stored, how long they remain, or whether they were used for training. This is design rather than accident. That route never carried anything that keeps a record in the first place.

Employees themselves worry about it too. The Konstanz survey carries a question on this. Concern about what happens to the information entered into AI tools ran highest in medium-sized organizations, at 27%. The research team wrote that the value cannot be read in one direction. Questions may only arise once responsibility has become clear, or the figure may signal unease about control. Under either reading, one in four people already wonder where the information they typed in went.

That is why lineage comes first when we talk about AI-Ready Data. Putting data in order is preparation for model performance, and before that it is the work of reaching a state where you can say what is where right now. The Konstanz survey shows the distance from that state widening faster than adoption itself. Use rose 3 points, while nearly half of that use came in through a route that records nothing.

Writing one more rules document does not close this gap. With a list of prohibitions attached, the 66% in the PagerDuty survey still stands: people go on using tools they believe are not allowed, and the use only moves somewhere less visible. A list is what comes first: which AI tools are in use at this company, what data goes into each of them, and where that data stays and for how long. Four checks give a rough position.

• Can you name the AI tools used for work right now? Not only the ones that went through approval, but everything actually in use.
• What kinds of data go into each tool? Can you keep the tools that take customer information, contracts and source code marked separately?
• Is there a way to confirm where that data stays? Down to whether a clause about not training on it sits in the contract, and whether anyone has checked that clause.
• Is there somewhere for an employee to ask before using a new tool? With nowhere to ask, silence answers in place of a rule.

These questions are not an insight Pebblous arrived at alone. The first item of the recommendations carried in Section 2 already lists four things for a company to settle: which tools may be used, which data should not be entered, where human oversight has to remain, and how AI-generated results get verified. The second of those points at the same place as the second item in the list above. Deciding what must not go in and knowing where what went in has gone are the front and back of one question. The research team added that rules of this kind belong not to the category of restriction but to the conditions for using AI safely. Researchers in organizational behavior and a company that works with data came in through different doors and arrived in the same room.

Two things are worth holding onto. This survey asked 1,105 employees in Germany, and the two surveys attached in Section 3 are materials published by companies that sell security products. None of them says what the figure is for Korean companies. The structure of tools arriving ahead of the company, on the other hand, travels across borders easily. Whether you can name the AI tools in use at your company, together with where the data went, is the same question anywhere.

Thanks for reading this far. The figures this article cites can be checked by anyone in the Konstanz press release and the English coverage. We would be glad to hear how your organization keeps track of the AI tools in use, and whether that list carries the destination of the data alongside each name.

Academic Surveys & Institutional Sources

• 1.Kunze, F., Opitz, C., & Gerdiken, E. (2026). "Konstanzer KI-Studie 2026: Verlangsamtes Wachstum der KI-Nutzung am Arbeitsplatz, aber nicht für alle Beschäftigte gleichermaßen." Results report, September 2026, University of Konstanz (KOPS). The full 20-page report. The size-band figures and the Section 4.4 material in this article were read directly from it.
• 2.Arntz, M., Baum, M., Brüll, E., Dorau, R., Hartwig, M., Matthes, B., Meyer, S.-C., Schlenker, O., Tisch, A., & Wischniewski, S. (2025). "Low barriers, high stakes: Formal and informal diffusion of AI in the workplace." ifo Working Paper No. 422, ifo Institute. The study the Konstanz report cites as the source of the dual diffusion concept.
• 3.Universität Konstanz. (2026-09-08). "KI-Nutzung am Arbeitsplatz: langsam, informell und ungleich." Press release PI Nr. 83/2026 (idw).

Press Coverage

• 4.Phys.org. (2026-09-18). "Workplace AI adoption remains slow, informal and uneven, survey finds."
• 5.Haufe. (2026-09-08). "Konstanzer KI-Studie: Kein Umbruch, aber größere Ungleichheit."

Industry Surveys (Vendor-Published)

• 6.PagerDuty. (2026-06-11). "PagerDuty Report Finds Two-Thirds (66%) of Office Professionals Have Used Unauthorized AI Tools at Work." Fieldwork by Wakefield Research, 1,250 office professionals, April 2026.
• 7.Teramind. (2026-06-17). "Shadow AI Behavior Report." Survey of 300 corporate information security leaders plus behavioral-record analysis.