AI on the factory floor: who gets the gains, and who carries the risk?
A robot may take over a task while a worker keeps responsibility when something goes wrong. IBM’s survey of 800 leaders points towards job redesign, but its promises of higher productivity and fewer staff are expectations, not measured results.
The short version
- What happened
- IBM surveyed 800 executives and senior leaders responsible for physical-AI strategy across 11 countries and 13 industries between May and July 2026, alongside 12 interviews.
- Who it affects
- Factory and logistics workers may see tasks reallocated to autonomous machines while remaining responsible for supervision, recovery and unusual cases.
- What remains unproven
- The survey reports executive expectations and self-assessed deployment stages rather than independently audited outcomes.
Keep in mind
The survey reports executive expectations and self-assessed deployment stages rather than independently audited outcomes.
Its cross-sectional design cannot show that wider physical-AI deployment caused revenue, productivity or workforce changes.
The report does not provide a representative worker sample, longitudinal job counts or comparative injury data.
Industry roadmaps include only sectors with at least 50 respondents, limiting subgroup precision.
IBM sells AI, cloud and consulting services and helped produce the study, creating a relevant commercial interest.
Source analysis · 2 linked sources. Primary analysed source date: 13 October 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The job changes before the job title does
Factory workers may find themselves supervising autonomous machines rather than being replaced by them. That is the direction most leaders in a new IBM survey expect physical AI to take. But the report leaves a harder question unanswered: who benefits from the extra output, and who carries responsibility when a machine fails?
Among 800 executives and senior leaders, 55% expected people and machines to work through collaboration or co-creation by 2030, while 15% expected near-full autonomy. Respondents forecast a workforce reduction of 9% alongside productivity growth of almost 50%.
Those answers point to changes in tasks, supervision and skills. They are leaders’ expectations, not observed job losses or measured productivity gains. The study does not support a simple prediction that autonomous machines will remove most factory jobs.
The distinction matters for technicians, operators, safety staff and managers. Physical AI joins software models to robots, vehicles, sensors and other machines that act in the world. A system may inspect equipment, route materials or support maintenance without replacing the whole occupation around that task. Jobs can still disappear, emerge or change as duties are regrouped. The survey records what leaders expect; it does not observe hiring, displacement, wages, injuries or output after deployment, and it cannot tell workers what will happen at a particular site.[1][2]
What IBM studied—and who answered
IBM Institute for Business Value surveyed 800 executives and senior leaders who were responsible for, or directly involved in, physical-AI strategy. Fieldwork ran from May to July 2026. Respondents came from 11 countries and 13 industry groups, in organisations reporting annual revenue from US$100 million to US$180 billion. The regional mix was 37% Asia-Pacific, 36% Europe, 22% North America and 5% Latin America. IBM also conducted 12 in-depth interviews to add operational context to the questionnaire results.
That denominator is useful but it is not a representative census of employers or workers. The report does not turn an executive sample into national adoption rates, and senior leaders may view implementation, productivity and workforce effects differently from frontline staff. The industry roadmap includes only sectors with at least 50 respondents, which avoids presenting very small industry cuts but still leaves limited precision. The report was produced by IBM with the World Manufacturing Foundation and SEMI; IBM sells relevant AI, cloud and consulting services, so its commercial position should remain visible alongside the disclosed method.[1]
Deployment is still concentrated in selected uses
Only 25% of respondents said their organisations were scaling physical AI beyond selected use cases, and fewer than 40% described their organisation as mature in any part of the lifecycle measured by IBM. That gap is more informative than a futuristic demonstration. It suggests that many organisations are still choosing tasks, building simulations, integrating machines with existing systems, training staff and deciding who remains accountable when a system acts unexpectedly.
The technical difficulty is not confined to model accuracy. Among organisations that both build and operate physical-AI systems, 71% said the gap between simulation and reality was a challenge. A robot trained in a controlled digital environment can encounter wear, glare, clutter, unusual objects, changing loads or human behaviour that was absent from training. For workers, the practical safeguards are therefore mundane but consequential: safe shutdowns, clearly bounded authority, local fallback procedures, maintenance access, incident reporting and enough staffing to intervene when the system leaves its validated operating conditions.[1]
The return figures are expectations, not audited performance
IBM reports average expected operational returns of 43% among organisations exploring physical AI, 80% in pilots, 105% in selected deployments, 102% among those scaling and 161% where physical AI was described as core to the operating model. The report explicitly presents these as self-reported expectations rather than projections. The final group is also comparatively small. These percentages do not show realised cash returns, the time needed to earn them, the total cost of integration or the losses from projects that were abandoned.
A credible business case would specify the baseline, investment period, hardware and software costs, integration work, downtime, maintenance, energy, insurance, staff training and safety controls. It would then measure output quality, cycle time, defects and incidents using the same definitions before and after deployment. Without that structure, two leaders can attach very different meanings to operational return. High expectations may identify enthusiasm and capital allocation; they do not establish that a robot or autonomous system will generate the claimed value in another workplace.[1]
A revenue association is real in the sample—but not causal proof
The report also compares organisations at different deployment stages. The 174 respondents reporting selected deployments attributed an average 4.0% of revenue to physical AI, compared with 5.3% among 163 respondents reporting wider scaling. Using unrounded values, IBM gives a difference of 1.24 percentage points and reports t(335)=8.72 with p<0.001. It illustrates that gap as US$248 million for a hypothetical US$20 billion company.
That statistical association does not show that scaling caused the extra revenue. Companies with stronger balance sheets, more automation-ready processes, better data or more favourable markets may be both more able to scale and more likely to report higher contribution. Revenue attribution is itself a management estimate, not an independently audited measure. The US$248 million example is arithmetic applied to a hypothetical company, not a forecast or a guaranteed gain. Longitudinal evidence with comparable baselines and verified accounts would be needed before treating the difference as an effect of deployment.[1]
What workers and employers can do now
Employers can map the change at task level before promising productivity or announcing headcount assumptions. For each proposed system, teams should identify what the machine will perceive, decide and act on; which worker remains responsible; how the system signals uncertainty; and what happens during network, sensor or mechanical failure. Consultation should include operators, maintenance teams, health-and-safety representatives and people whose performance may be scored by the system. Training needs to cover intervention and recovery, not only routine use.
Workers and unions should ask for measured outcomes rather than broad claims about augmentation. Useful indicators include staffing by role, overtime, injury and near-miss rates, downtime, defect rates, work pace, pay, training hours and the number of decisions that still require human approval. Physical AI may remove hazardous or repetitive duties while creating monitoring, integration and repair work, but those benefits are not automatic. A productivity gain can coexist with work intensification or reduced bargaining power unless implementation measures both operational and human outcomes.[1][2]
What remains unproven—and what would change the assessment
This study is strongest as a structured account of senior leaders’ expectations across several regions and industries. It is weaker as evidence of realised economic or workforce effects. The survey is cross-sectional, respondents self-reported their organisation’s stage and expected results, and the report does not establish a counterfactual. It does not provide worker interviews, independently verified financial outcomes, longitudinal job counts or comparable safety data. Commercial involvement does not invalidate the results, but it increases the importance of transparent methods and external replication.
The assessment would change with preregistered longitudinal studies that follow deployments from baseline through scaling, publish total costs and failures, and compare matched sites that did and did not adopt the technology. Independent audits should test productivity, quality, safety, energy use and employment by role, with results reported across countries and company sizes. Worker-reported job quality, autonomy and training access should sit beside management estimates. Until those data exist, IBM’s report is evidence that leaders expect collaboration and large gains—not proof that those gains will arrive or that workforce effects will be modest.[1]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring workforce: compare the finding and its evidence limits.
CFOs expect less than 0.4% AI job loss in 2026—but are they measuring or forecasting?
A 734-executive study found broad but shallow AI adoption and modest expected aggregate employment effects. The figures come from company leaders' attribution and forecasts, not observed economy-wide job losses caused by AI.
Read the next reportExplore the full source and evidence record
What this means for people
- Factory and logistics workers may see tasks reallocated to autonomous machines while remaining responsible for supervision, recovery and unusual cases.
- Technicians and safety teams may need more training in sensors, simulation, machine behaviour and incident response, not simply fewer staff.
- Employers should measure workload, injuries, job quality and pay alongside output so that productivity claims do not hide costs transferred to workers.
Global context
The sample spans 11 countries across Europe, Asia-Pacific, North America and Latin America, but it is an executive survey rather than a workforce-representative study. Labour law, wage structures, industrial relations, infrastructure and access to capital will shape outcomes differently across countries and sectors.
What the evidence does not yet show
- The survey reports executive expectations and self-assessed deployment stages rather than independently audited outcomes.
- Its cross-sectional design cannot show that wider physical-AI deployment caused revenue, productivity or workforce changes.
- The report does not provide a representative worker sample, longitudinal job counts or comparative injury data.
- Industry roadmaps include only sectors with at least 50 respondents, limiting subgroup precision.
- IBM sells AI, cloud and consulting services and helped produce the study, creating a relevant commercial interest.
What to watch next
- Longitudinal studies comparing matched sites before and after physical-AI deployment
- Audited productivity, safety and total-cost measures rather than expected returns
- Job, pay and work-quality outcomes separated by occupation and region
- Evidence that systems remain safe when simulation assumptions fail in real workplaces
Living evidence record
Impact record IAI-0QT3HIA
Evidence stage
Observed
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
13 October 2026
Source trail
2 direct sources across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Related-source reporting disclosure
This record analyses 2 linked source records around the same underlying development. The extra records add method, date or context, but they do not by themselves constitute independent replication of every performance claim or predicted outcome.
Evidence trail
Sources used for this report
Links checked 13 October 2026
Research question, topics and regions
How senior leaders expect physical AI to change operations, productivity and work, and whether organisations reporting wider deployment also report different business outcomes
Topics: Physical AI · Manufacturing · Robotics · Workforce · Productivity · Human-machine collaboration
Regions: Global · Europe · Asia · North America · South America
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.