Summary

• IBM surveyed 1,500 CHROs and 8,800 employees about how AI is changing work, skills, and accountability.
• Seventy-one per cent of CHROs call supervising, validating, and overriding AI essential, while only 29% of employees prioritise judgement.
• Eighty per cent of CHROs say AI creates hidden work including checking recommendations, correcting errors, and handling exceptions.

Companies adopting artificial intelligence are finding that automation creates a new category of human work around checking, correcting, and overruling machines, according to a large workforce study from IBM.

The IBM Institute for Business Value surveyed 1,500 chief human resources officers and equivalent executives alongside 8,800 full-time employees, finding a substantial gap between the capabilities employers think AI requires and the skills workers themselves are prioritising. Seventy-one per cent of CHROs identified the ability to supervise, validate, and override AI output as essential, whereas only 29% of employees ranked judgement as important.

At the same time, 60% of employees said they were worried that AI was eroding their skills, with critical thinking most frequently identified as an area of decline. Among employees who held those concerns, three quarters said some erosion had already begun.

The findings complicate a familiar productivity argument around workplace AI. Generative systems can remove portions of routine work, but somebody still has to determine whether an answer is correct, recognise an unusual case, provide context the model lacks, and take responsibility when a recommendation becomes a business decision.

Automation creates work around the automation

IBM’s research gives that additional labour a measurable shape. Eighty per cent of CHROs said AI adoption creates “invisible” work for employees, including validation, error correction, context provision, and exception handling, while 42% of employees said AI either increases their workload or produces additional work that goes unrecognised.

Those tasks are easy to omit from the original business case for a deployment because they appear downstream of the automated process. A model may produce a draft in seconds, for example, while the time required to verify factual claims, ensure compliance, correct tone, or investigate an abnormal result appears elsewhere in the workflow.

The balance varies considerably by use case. Low-risk repetitive tasks may tolerate extensive automation, whereas a financial decision, safety process, customer complaint, medical workflow, or regulated document can require a human to understand why an output should be trusted before it is acted upon.

That places judgement closer to the centre of technology implementation than many skills programmes assume. Prompting and tool familiarity can help employees interact with AI, but they do not replace the subject expertise required to spot an answer that is plausible, fluent, and wrong.

IBM found a related accountability problem when systems fail. Forty-three per cent of employees said blame falls on them when something goes wrong with AI, while 41% of CHROs believed staff might not feel safe challenging or overriding automated outputs. More than a third of CHROs said unclear accountability complicates AI deployment.

Work design becomes part of AI control

The research suggests organisations are beginning to differentiate more explicitly between work that remains human-led, work assisted by AI, and work executed by AI. IBM reports that organisations defining those boundaries clearly also report lower risk and improved quality, although the study relies on self-reported organisational outcomes rather than controlled performance measurement.

That distinction remains useful because deploying an assistant without deciding who owns the final judgement leaves an accountability gap inside the process. When humans are expected to intervene but lack authority, time, training, or access to the evidence needed to challenge the system, a nominal human checkpoint can become little more than a final click.

The organisation of AI programmes can make that problem worse. Nearly half of the organisations surveyed do not involve the CHRO when AI strategy is defined, while only 28% of CHROs said HR and IT have a joint roadmap supported by a shared operating cadence. At the same time, 73% reported difficulty coordinating consistently across the executive team.

That separation is hard to sustain once AI changes job design rather than merely providing another software tool. Decisions about which tasks disappear, which require checking, how performance is measured, who may override a system, and how employees are trained sit across technology, operations, risk, and workforce management simultaneously.

IBM’s own data shows that HR departments have work to do as well. Seventy-two per cent of organisations reported limited or no use of AI inside the HR function, while CHROs gave relatively low assessments of their teams’ AI literacy, performance measurement, and change-management capabilities.

The gap is therefore not simply between workers and technology. Organisations are introducing AI into operating models while many of the functions responsible for skills, accountability, governance, and measurement are still deciding how their own roles should change.

As companies move from experiments into routine use, the productivity calculation will have to include the human effort required after an AI system has produced its output. Automation can remove work, but validation and exception handling remain labour even when they are absorbed quietly into somebody else’s job.