After AI quickly produces a summary, someone still needs to verify figures, restore omitted information and decide whether it is ready to deliver. In a survey published on 21 September 2026, 80% of participating senior HR leaders said AI would increase less-visible work such as checking recommendations, correcting errors and handling exceptions. IBM's study
Companies evaluating a pilot can record this work alongside generation time. Measuring only how quickly a tool produces an answer misses the people-hours needed to complete the task.
Survey responses are not a test of employee ability
In the survey, 71% of senior HR leaders valued the ability to oversee, verify and override AI outputs, while 29% of employees identified judgment as important. These figures describe how two respondent groups rated skills; they do not mean that only 29% of employees can exercise judgment. Another 60% of employees said they were concerned about skill erosion. That self-reported concern does not prove that AI has caused a decline in ability. IBM data and methodology
IBM's Institute for Business Value and Oxford Economics surveyed 1,500 senior HR leaders and equivalent executives, and 8,800 full-time employees, between April and June 2026. The leaders' organizations reported annual revenue or budgets from US$250 million to US$123.7 billion. This sample differs from an individual Taiwanese SME and is useful for framing management questions, not diagnosing a particular workforce.
Record the work from input through delivery
The following is a suggested pilot design, not an outcome promised by the study. For a document-summary task using company-approved files with unnecessary personal data removed, record four durations for each trial: preparing the input, generating a draft, human review, and revision through delivery. Also record error types and reasons for returning a draft. This helps distinguish time saved by the tool from added review effort.
Reviewers should have access to the source material and authority to reject content they cannot verify. Decide in advance who handles inconsistent amounts, missing sources or information outside the permitted scope, rather than leaving accountability until an error occurs.
Training can use the same log. Ask participants to identify which statements are supported and which need evidence, then have a manager compare their judgments. Training outcomes can then be shown through review and delivery records, not merely a polished AI output.
Whether to expand the pilot depends on the task's quality requirements, data constraints and observed results. Do not infer department-wide savings from one successful use.
Source checked: 26 September 2026.