Introducing artificial intelligence into a workplace does not automatically make employees more engaged, productive or satisfied. A new study of 695 German knowledge workers suggests that the difference may lie partly in how people change their jobs around the technology. Workers who viewed their AI systems as supportive were more likely to report proactively developing their roles, but they were also more likely to report avoiding demanding aspects of work. Those two patterns were associated with markedly different outcomes.
Published on 9 October 2026 in Frontiers in Psychology, the peer-reviewed research moves beyond the familiar question of whether employees use AI. Instead, it examines what employees do with the opportunities and pressures they perceive after AI enters their work. Its central finding is nuanced: positive perceptions of AI were linked to stronger job satisfaction, self-rated productivity and engagement overall, while different kinds of employee-led work redesign pointed in opposite directions. The research is observational and cannot establish that AI caused those outcomes.
Why the way people adapt matters
Organizations often assess AI investments through adoption rates, time saved or the number of tasks automated. Those measures may miss changes in employees’ experience of their jobs. A system that handles routine information processing could free someone to learn a new skill or collaborate with colleagues. It could also prompt someone to withdraw from difficult assignments, narrow their responsibilities or become less connected to meaningful work.
Researchers describe employees’ self-initiated adjustments to their roles as job crafting. In this study, approach crafting referred to changes that expand resources, learning and constructive challenges, such as seeking advice, building skills or taking on interesting responsibilities. Avoidance crafting referred to reducing burdensome demands, including making work less mentally intensive. Neither behaviour is necessarily good or bad in every circumstance, but the distinction helps explain why the same technology can accompany different workplace experiences.
The authors, Fabian Willemsen, Charlotte Hopp, Susanne Mütze-Niewöhner and Verena Nitsch of RWTH Aachen University, investigated whether these two forms of job crafting statistically connected perceived AI support with five outcomes: job satisfaction, perceived productivity, work engagement, intentions to leave and burnout symptoms.
How the researchers studied 695 workers
The team conducted an online survey in June 2024 using a panel of adults aged 18 to 65 living in Germany. Participants had to be employed in information-intensive work and already use an AI system in their job. Of 768 completed surveys, 73 were excluded following data-quality checks, leaving 695 analysed responses. The participants averaged 38.64 years of age, and 53.1% held at least a bachelor’s degree.
Men accounted for 62.9% of the sample, women for 36.7%, and non-binary respondents for 0.4%. About half, 49.8%, reported holding leadership roles. These details matter because the group was not a representative cross-section of all employees. It was concentrated in information and knowledge work, where AI systems may be more readily integrated into daily tasks than in frontline, manual or informal employment.
The AI tools themselves were varied. Data analysis and business analytics systems were the largest category at 33.1%, followed by generative AI at 27.2%, conversational AI at 13.8%, image and document recognition at 13.7%, and prediction and planning systems at 12.2%. The researchers also distinguished employer-led deployment from employees’ own initiative: 59.0% reported organization-led AI adoption, while 41.0% had adopted a tool themselves.
Participants answered established or adapted questionnaires on perceived AI support, the two job-crafting approaches and the five work outcomes. The researchers then used structural equation modelling to examine their relationships simultaneously, accounting for age, gender, education and whether AI adoption was employer-led. They calculated bias-corrected confidence intervals from 5,000 bootstrap samples. These are statistical associations within a model, not measurements of what would happen if a company introduced AI in a randomized experiment.
AI support was associated with both proactive and avoidance behaviour
The first result complicates a simple success narrative. Perceived AI support was strongly associated with approach crafting, with a standardized coefficient of 0.626 (p < 0.001). It was also positively associated with avoidance crafting, although more weakly, with a coefficient of 0.296 (p < 0.001). In other words, respondents who regarded AI as helpful tended to report more of both kinds of job adjustment, not only the proactive kind.
A standardized coefficient expresses the modelled change in an outcome, in standard-deviation units, associated with a one-standard-deviation difference in the predictor, after accounting for the other specified variables. A coefficient of 0.626 is not a 62.6% improvement in work performance, nor does it mean 62.6% of employees benefited.
There is a plausible workplace explanation for the two-direction pattern. AI can create room for learning and new responsibilities while also allowing employees to sidestep demanding tasks. The study did not directly test why individual workers chose either response. It therefore cannot determine whether avoidance represented sensible workload management, disengagement, or a mixture of both.
Proactive job redesign showed the strongest links to better outcomes
Within the statistical model, approach crafting was positively associated with job satisfaction (β = 0.708), perceived productivity (β = 0.717) and work engagement (β = 0.777). All three relationships had p-values below 0.001. It was also negatively associated with turnover intentions (β = -0.171, p = 0.021) and burnout (β = -0.243, p = 0.001).
The opposing pattern appeared for avoidance crafting. It was negatively associated with job satisfaction (β = -0.161, p < 0.001) and perceived productivity (β = -0.101, p = 0.016), but positively associated with intentions to leave (β = 0.416, p < 0.001) and burnout (β = 0.195, p < 0.001). Its association with work engagement was not statistically significant (β = -0.054, p = 0.188).
The difference is especially striking for turnover intentions: avoidance crafting had a considerably stronger positive association with thinking about leaving than approach crafting had in the opposite direction. That does not establish that reducing demanding work makes employees resign. Someone already dissatisfied with their role could both avoid difficult tasks and consider another job.
What the combined model actually found
When both pathways were considered together, perceived AI support showed positive total associations with job satisfaction (β = 0.373), perceived productivity (β = 0.394) and work engagement (β = 0.465), all statistically significant at p < 0.001. The total association with burnout was smaller and negative (β = -0.117, p = 0.011). The total association with turnover intentions was not statistically significant (β = 0.024, p = 0.548).
That last result is important. It would be misleading to say that supportive AI was shown to make employees less likely to leave. Although the approach pathway pointed towards lower turnover intentions, the avoidance pathway pointed in the other direction. In the combined analysis, those tendencies did not produce a statistically clear overall relationship.
The researchers also found that the combined indirect association with burnout, through both job-crafting pathways, was not statistically significant (β = -0.095, p = 0.086), even though the overall total association with burnout was significant. The distinction matters because a significant overall relationship does not prove that the proposed intervening mechanisms fully explain it.
All direct associations between perceived AI support and the five work outcomes became statistically non-significant in the model after the two crafting pathways were included. The authors interpret this as a pattern consistent with statistical mediation. Because every variable was measured at one point in time, it is more accurate to call this an associational mediation model than proof of a causal chain.
How employees described changes after AI adoption
Additional survey responses provide useful context beyond the main model. 60.0% said AI had changed how they performed existing tasks, and 27.1% said it had changed collaboration with colleagues or clients. 42.9% reported improved working conditions, while 12.1% reported deterioration. Because respondents could select multiple types of change, those percentages do not add up to 100.
Meanwhile, 35.8% reported that AI had eliminated tasks or task components, 26.2% said it had added tasks or components, and 34.4% reported new learning or development opportunities. These are self-reported experiences rather than independently verified changes in job descriptions or productivity records. Still, they illustrate why a single measure such as adoption can conceal different forms of workplace change.
Research Today recently reported a related distinction in education: using AI in everyday life does not necessarily mean using it for learning. The workplace study raises a parallel question: does using AI translate into more meaningful and sustainable work, or simply a different distribution of tasks?
What managers and workers can take from the findings
For managers, the research suggests that AI implementation should include the design of work, not merely software deployment. A useful question is whether employees gain room to exercise judgment, learn new skills and collaborate, or whether they experience a narrowing of responsibilities. Training may be more valuable when it helps workers use AI to improve their role rather than simply to complete an existing task faster.
For employees, the findings highlight the potential importance of agency. An AI assistant might help someone develop an analysis, seek feedback or explore a more challenging assignment. Another worker might use it mainly to reduce demands. Neither behaviour should be judged without understanding workload, health, managerial expectations and the quality of the tasks being removed.
Employers also need to distinguish legitimate experimentation from unmanaged use of sensitive information. The study notes that self-initiated AI adoption can create governance and security questions. Research Today’s coverage of AI-based cybersecurity detection offers a separate example of why technical capability and real-world organizational safeguards should not be conflated. The present survey did not evaluate security incidents or compare formal AI policies.
What the study cannot establish
Several limitations substantially constrain interpretation. First, this was a cross-sectional survey. It cannot show whether feeling supported by AI led employees to redesign their work, whether proactive employees were more inclined to view AI positively, or whether another factor influenced both. Statistical mediation cannot resolve the direction of causation without stronger longitudinal or experimental evidence.
Second, productivity, engagement, burnout symptoms and other outcomes were self-reported. The study did not measure output per hour, verified absenteeism, actual resignations or medically diagnosed burnout. Shared response tendencies can inflate correlations even when researchers take steps to reduce common-method bias.
Third, the sample was drawn from an online panel of German AI-using knowledge workers, with a substantial proportion in senior or leadership positions. Findings should not automatically be extended to South African workers, manufacturing operators, informal workers or employees who have not adopted AI. Workers who independently chose AI tools may also differ from those required to use them.
Finally, the model fit was acceptable but not perfect, and some relationships were sensitive to controlling for other variables. The authors call for studies that follow workers over time, compare organizations and incorporate objective performance measures. Those designs would help test whether encouraging particular forms of job crafting actually changes employee wellbeing or retention.
The bottom line
The study offers a more useful question than whether AI is good or bad for workers. In this sample, perceived AI support accompanied both proactive and avoidance-oriented changes to jobs. Proactive changes were associated with stronger satisfaction, engagement and perceived productivity; avoidance-oriented changes were linked to less favourable outcomes on several measures. The evidence is not causal, but it suggests that organizations should pay close attention to how work changes after AI arrives, rather than treating adoption as the finish line.
Source Information
Original study: Perceived AI support and work outcomes: testing the mediating role of job crafting in AI-supported work.
Authors: Fabian Willemsen, Charlotte Hopp, Susanne Mütze-Niewöhner and Verena Nitsch.
Journal: Frontiers in Psychology, Organizational Psychology, volume 17.
Publication date: 9 October 2026.
Study design: Peer-reviewed, cross-sectional online survey of 695 AI-using information and knowledge workers in Germany, analysed using structural equation modelling with 5,000 bootstrap samples.
DOI: 10.3389/fpsyg.2026.1824712.
