Does supportive AI improve work outcomes?
A survey of 695 German AI users linked perceived support to better self-reported work outcomes mainly through proactive job redesign. Because every measure came from one cross-sectional questionnaire, it cannot show that AI—or job crafting—caused higher productivity or wellbeing.
The short version
- What happened
- The study analysed 695 German information and knowledge workers who already used AI; 768 completed the survey and 73 were excluded for speeding or straight-line responding.
- Who it affects
- Workers should not be expected to repair weak AI implementation through invisible, unpaid job crafting.
- What remains unproven
- The survey was cross-sectional, so the temporal and causal direction between perceived support, job crafting and outcomes is unknown.
Keep in mind
The survey was cross-sectional, so the temporal and causal direction between perceived support, job crafting and outcomes is unknown.
Every central measure, including productivity, came from the same respondent at the same time; no objective output, quality or system-use data were collected.
The sample included only people already using AI and was skewed toward senior, leadership and degree-qualified information workers.
German panel data collected in June 2024 may not transfer to other labour markets, current systems, frontline roles or tightly managed work.
The model did not directly measure job loss fears, technology stress, workplace surveillance, leadership or social support.
Source analysis · 1 linked source. Primary analysed source date: 9 October 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The direct answer: the survey shows a pathway, not an AI effect
This study does not prove that supportive AI improves work. Among 695 employees in Germany who already used an AI system, people who perceived that system as more supportive also reported more approach-oriented job crafting—actively changing tasks, relationships or thinking to make work more useful—and better job satisfaction, engagement and perceived productivity. The statistical model placed most of those favourable associations along the approach-crafting pathway rather than as a direct link from perceived AI support to outcomes.
The distinction matters for managers and workers. Buying a model or switching on a workplace assistant is not the same as improving a job. The paper is more consistent with a work-design interpretation: employees may report better experiences when a system feels useful and they have room to reshape work around it. Yet the reverse explanation is equally possible. People in satisfying, autonomous jobs may be more likely to see AI as supportive and to describe their own adaptations positively. A one-time questionnaire cannot determine the direction.[1]
Who was surveyed and what kinds of AI they used
A panel provider surveyed Germany-based employees in June 2024. Eligibility was restricted to people aged 18 to 65 who mainly performed information-intensive work and already used AI at work. Of 768 respondents who finished, 73 were excluded under the study's quality rules for speeding or straight-line answers, leaving 695. The mean age was 38.6; 255 participants were women, 437 men and three non-binary. The sample was senior-skewed: 52.9% described their role as senior or higher, 49.8% had leadership responsibility and 53.1% held at least a bachelor's degree.
The AI category was data or business analytics for 230 people, generative AI for 189, conversational AI for 96, image or document recognition for 95, and prediction or planning for 85. Organisations had introduced the system for 410 respondents, while 285 said adoption was self-initiated. That 41% is not an estimate of unsanctioned AI use across Germany; it is the share within a selected panel of existing workplace AI users. ICT, manufacturing, public administration, education and health were prominent, with business, administrative and STEM occupations dominating.[1]
What workers said had changed
Participants could select several changes. Sixty per cent said AI changed how they performed existing tasks, 27.1% reported changed collaboration, 42.9% reported improved working conditions and 12.1% deterioration. Tasks or task components had been added for 26.2% and eliminated for 35.8%, while 34.4% reported new learning and development opportunities. Because multiple responses were allowed, these categories overlap and do not add to 100%.
Those figures describe experiences among current users, not objectively observed task change. They also reveal why a single positive-or-negative verdict is inadequate: the same technology can remove work, add work, improve conditions and worsen them for different people—or for one person at different moments. A responsible implementation should therefore measure workload, autonomy, error correction, monitoring and skill use separately. An average adoption score can conceal whether gains for one group depend on hidden review or repair work by another.[1]
How the model separated approach and avoidance crafting
All constructs were measured on five-point self-report scales. The perceived-support items asked whether AI relieved unpleasant tasks, handled supplementary work, enabled focus on meaningful work, supported creative problem-solving or improved performance. Approach crafting covered efforts to expand resources or take on useful challenges. Avoidance crafting covered efforts to reduce or escape difficult aspects of work. Reported reliability was 0.83 for perceived support, 0.87 for approach crafting and 0.86 for avoidance crafting.
The researchers fitted a parallel-mediation structural equation model, controlling for age, gender, education and whether adoption was organisation-led or self-initiated, and used 5,000 bootstrap draws for indirect effects. Perceived AI support was associated with approach crafting at a standardised coefficient of 0.626 and, unexpectedly, with avoidance crafting at 0.296. Because the data were cross-sectional, the authors explicitly describe these as associations rather than causal mediation. The analysis cannot show that support occurred first, crafting second and outcomes third.[1]
The favourable and adverse pathways partly offset each other
Approach crafting was positively associated with job satisfaction, perceived productivity and work engagement, with standardised coefficients of 0.708, 0.717 and 0.777. It was negatively associated with turnover intention at minus 0.171 and burnout at minus 0.243. Avoidance crafting showed the opposite pattern for four outcomes: minus 0.161 for satisfaction, minus 0.101 for perceived productivity, plus 0.416 for turnover intention and plus 0.195 for burnout. Its link with engagement was not statistically significant.
All direct paths from perceived AI support to the five outcomes were non-significant once both crafting pathways were included. Total indirect associations were significant for satisfaction, perceived productivity and engagement, but not for turnover intention or burnout. The paper also reports a negative total association with burnout even though the combined indirect estimate missed its significance threshold, illustrating how model components can give a more complicated picture than a headline. None of these coefficients is an objective productivity gain or a clinical estimate of prevented burnout.[1]
What employers and workers should do with this evidence
Employers should not make workers carry the entire burden of redesign. Job crafting can be useful when people have autonomy, time, training and permission to change a process. It can become unpaid adaptation when staff must invent workarounds, check unreliable output and absorb new responsibilities while performance targets stay fixed. Deployment plans should specify which tasks change, who remains accountable, how review time is counted, what monitoring is prohibited and how workers can report deterioration without risking their role.
For workers and representatives, the practical measures are concrete: collect before-and-after evidence on workload, error correction, discretion, pace, learning and meaningful task content; separate self-rated productivity from audited output; and examine differences by role and seniority. The senior and leadership-heavy sample may have had more room to craft jobs than junior, precarious or tightly managed workers. An AI tool that supports a manager may intensify scheduling or surveillance for frontline staff, a distributional effect this survey was not designed to test.[1]
Funding, disclosure and what would change the assessment
The work was conducted within the AKzentE4.0 project, funded by Germany's Federal Ministry of Research, Technology and Space under funding code 02L19C400; RWTH Aachen University supported open-access publication. The authors declared no commercial or financial conflict. They also disclosed generative-AI use for language polishing, academic-expression refinement and sentence structure. Those declarations improve transparency but do not resolve selection or common-method limitations.
Stronger evidence would follow workers over time before and after an implementation, compare teams or sites with different work-design support, and include objective quality and workload measures alongside confidential worker reports. Randomised or carefully matched interventions could test whether giving employees autonomy, training and formal redesign time changes outcomes. Representation from junior, manual, customer-facing and algorithmically managed roles is essential. Until then, the finding is a useful warning against technological determinism: perceived support and worker agency travel together, but causation and real productivity remain unproven.[1]