IBM’s new workforce study finds a gap between the AI oversight skills executives prioritize and the work employees say they actually experience.
The clearest finding in the IBM AI invisible work study is not a productivity percentage. Eighty percent of surveyed CHROs say AI adoption creates extra work such as validating recommendations, correcting mistakes, adding context and managing exceptions. Forty-two percent of employees say AI increases their workload or produces work that goes unrecognized.
How the study was conducted
The IBM Institute for Business Value worked with Oxford Economics on two surveys between April and June 2026. One collected responses from 1,500 CHROs and equivalent senior workforce executives across 21 geographies and 23 industries. The second surveyed 8,800 full-time employees in 28 countries.
Executive respondents worked for organizations ranging from $250 million to $123.7 billion in annual revenue or budget and from 200 to more than 880,000 employees. That is a broad sample, but the results remain self-reported survey evidence. They identify patterns and perceptions, not a controlled estimate of AI’s causal effect on productivity.
The oversight-skill gap
| Finding | Reported result | What it suggests |
|---|---|---|
| Supervise, validate and override AI | 71% of CHROs call it essential | Oversight is becoming a core job skill |
| Employee priority for judgment | 29% rank it as important | Training messages may not match daily incentives |
| Concern about skills erosion | 60% of employees | Automation anxiety includes capability loss |
| Invisible work from AI | 80% of CHROs | Validation and correction need explicit ownership |
| More or unrecognized work | 42% of employees | Productivity claims may omit human repair effort |
| Limited or no AI in HR | 72% of organizations | HR is asked to govern tools it may not use deeply |
What invisible AI work looks like
- Checking whether a summary omitted a customer commitment.
- Comparing generated analysis with the source spreadsheet.
- Rewriting a plausible answer that uses the wrong policy.
- Adding context that the model could not access.
- Handling exceptions that fall outside an automated workflow.
- Documenting why a human overrode a recommendation.
- Taking responsibility when an AI-assisted decision fails.
These tasks are easy to miss because they are dispersed across many roles. A dashboard may record faster drafting without recording how long a reviewer spent validating the result. Our coverage of Workday’s AI variance analysis shows why a generated explanation still needs an owner who can verify the numbers and business context.
Accountability often stays with the employee
IBM reports that 43% of employees say blame falls on them when AI goes wrong. That can create a bad incentive: use the tool because management expects speed, but carry the personal risk when its output fails. A responsible workflow gives employees authority to question, pause or override the system without being punished for slowing the process.
The study says 76% of employees feel safe questioning or overriding AI recommendations when HR shares responsibility for deciding which decisions remain human-led. The figure falls to 43% when HR is merely advisory. This is an association from the survey, not proof that changing HR’s role alone will cause the improvement.
Measure completed work, correction work and risk together
- Define the unit of completed work before adding AI.
- Measure review and correction minutes, not only generation speed.
- Track how often people reject or override the output.
- Record severity, not only the count, of errors that reach users.
- Separate time saved from work shifted to another team.
- Ask employees whether tool use is optional, expected or enforced.
- Reward safe escalation instead of treating it as resistance.
- Revisit staffing only after quality and risk metrics stabilize.
Critical thinking is a workflow property
Training people to be critical is not enough if the interface hides sources, the deadline leaves no review time or managers reward acceptance speed. Organizations need source visibility, confidence boundaries, sampled audits and clear override paths. The Atlassian structured-content analysis illustrates a related principle: better inputs can improve answers, but users still need evidence and governance around the final result.
Limits of the reported business-performance comparisons
IBM says organizations with mature HR AI capabilities are nearly twice as likely to report a higher number of positive business indicators. It also reports risk-reduction and quality differences for organizations that classify workflows as human-led, AI-assisted or AI-executed. These comparisons are observational and based on IBM’s segmentation. They should not be presented as proof that one governance practice directly caused a specified business gain.
Primary sources
Checked September 22, 2026. Survey findings and business comparisons are company-reported and do not establish causation.
