Do employees know they are expected to check and override AI?

Mostly not, on the largest measurement so far. IBM’s Institute for Business Value surveyed 1,500 chief human resources officers and 8,800 employees between April and June 2026 and reported on 21 September that 71 per cent of the executives rank the ability to supervise, validate or override AI output among the most essential skills, against 38 per cent of the employees. Only 26 per cent of the organisations had written down which work is human-led, which is AI-assisted and which is AI-executed, and 43 per cent of employees say that when AI fails the blame reaches them. Read together, the three numbers describe a job that has been assigned without being defined. The executives expect a check. The people expected to perform it have not been told, and the work they would need to perform it has not been designed.

What the study is, and who paid for it#

The 2026 CHRO Study is the human resources volume of IBM’s global C-suite series, produced with Oxford Economics. Its press release describes two surveys: 1,500 CHROs and senior workforce executives across 21 geographies and 23 industries, in organisations from 200 to more than 880,000 employees, and 8,800 full-time employees across 28 countries. Both are self-report, neither is peer reviewed, and the publisher sells the software the questions are about. The title, Designing the Thinking Organization, is IBM’s own phrase and this page does not adopt it. What makes the study usable is its size and its design: the same questions put to the people who set the expectation and the people who carry it, which is how a gap can be measured at all. Campus Technology reported the findings on 21 September and Fair Play Talks on 23 September.

The gap: 71 against 38#

On critical thinking and problem framing the two populations nearly agree: 57 per cent of executives and 49 per cent of employees put it first, as Campus Technology reported the figures. The split opens on oversight. Supervising, validating or overriding what the machine produces is named by 71 per cent of executives and 38 per cent of employees, a distance of 33 points. The press release adds that 29 per cent of employees rank judgement among the most important skills for their own near future. One reading is that employees underrate a skill they will need. Another is that they are describing their jobs accurately. An employee who has never been told which outputs they are answerable for, or given the time to check them, has no reason to list checking as part of the work. The estate’s position is that the second reading deserves the benefit of the doubt until the first has been tested, because the executives in this survey had not, on their own account, done the thing that would tell an employee the check was theirs.

Why the gap forms: the work was never divided#

That thing is the division of labour. Only 26 per cent of the organisations surveyed clearly define work across human-led, AI-assisted and AI-executed activities, while 52 per cent of employees say their assigned tasks changed in the past year because of AI. So for roughly three quarters of organisations the tasks moved and the line did not. The study also reports that 36 per cent of executives find unclear accountability complicates AI deployment, and its recommendation is that workflow design should make ownership, validation, exception handling, escalation and override explicit. That is the same list this site has been calling Rules Before Tools since August 2025: which decisions the machine may make, who can stop each one, what people must remain able to do, and how anyone would know if it went wrong. An override that lives in a leader’s expectation and nowhere else is the arrangement who can override an AI system describes, where the authority continues on paper and stops existing in practice. Here it never reached the paper.

What it costs: blame and invisible work#

Two further figures show what the undefined check does to the people holding it. Fair Play Talks reported that 43 per cent of employees say the blame falls on them when AI fails, and that 80 per cent of CHROs believe AI creates invisible work, with 42 per cent of employees saying AI has increased the work that goes unrecognised. The first is the pattern Madeleine Elish named the moral crumple zone in 2019: responsibility for a failure settling on the human nearest the automated system, who had limited control over it. Her paper worked from selected accidents and could not say how often the pattern occurs; this survey is the first large sample to put a number on how often employees believe it has happened to them. The second is Bainbridge’s irony from 1983, examined on the invisible work of oversight: automation removes the routine execution and leaves the exceptions, and the time saved is counted while the checking is not. When 80 per cent of the executives who commission the systems say the work is invisible, the finding is that they can see it and have not yet paid for it.

What changes where judgement is designed in#

The study’s most useful comparison is between organisations that build judgement into the workflow and those that do not. Where it is built in, 62 per cent of CHROs report rising employee confidence in AI-enabled decisions; where it is not, 57 per cent report confidence falling, as Virtualization Review carried the figures on 21 September. The press release also reports 18 per cent lower risk and 20 per cent higher quality from organisations with clearly defined workflows, though both are the respondents’ own estimates of their own outcomes and should be read as such. The direction is consistent with what the estate’s own evidence shows: people trust a system more, and check it better, when the check has a name, a time and a consequence, and the confidence that rises without those is automation bias under a friendlier label. Amit Das, chief human resources officer of Bennett Coleman, is quoted in the release: “Fluency without judgment simply helps an organization make mistakes faster.”

Where this sits in the argument#

The disagreement between the 71 and the 38 is the measurable version of the handover this site keeps finding. Decisions have moved to machines; the capacity to review them has been assumed rather than assigned. Article 14 of the EU AI Act asks for competence, training and authority together, and this survey suggests most organisations have the authority in mind and neither of the other two in place. The decision an organisation controls is the one 26 per cent had taken: write down which decisions the machine may make, who can stop each one, and what the people in the loop must remain able to do, then check that the people named know it. The check on whether they know it is cheap. Ask them what they are answerable for, and compare the answer with the org chart. Where the two differ, the survey says which way. When should I override AI gives the six conditions a written rule should cover, and what management should report to the board asks for the override rate, the number that would show whether anyone is doing what the 71 per cent expect.

What this does not show#

It does not show that employees cannot supervise AI, only that fewer of them than their executives think it is part of the job, and a gap in perceived priority is not a gap in capability. All the figures are self-reported and the report is a commercial publication by a company that sells AI to the organisations surveyed, with the methodology in a press release rather than a technical appendix; response rates and the sampling frames are not published. The 43 per cent who say blame reaches them are reporting a belief, and the survey did not verify any case. The 18 and 20 per cent gains are respondents’ estimates of their own organisations, not measured outcomes. Nothing here measures whether any employee has ever overridden an AI decision, or whether the override was right; no organisation in the study published an override rate, and this page will be updated when one does.