The hidden work of supervising AI

What the survey says

IBM's September 21 release says 80% of surveyed CHROs believe AI creates additional, often unseen work. IBM and Oxford Economics surveyed 1,500 CHROs or equivalent executives across 21 geographies and 8,800 full-time employees across 28 countries during April–June 2026. The 80% is an executive perception measure, not the share of employees observed doing extra work or a measurement of hours lost.

Sources: IBM Institute for Business Value

Our analysis: define the whole task

Imagine a team preparing a customer response. The draft is one step. Someone must decide whether it answers the question, uses the right policy, reflects the customer's circumstances, and can be sent. If a productivity experiment stops its timer when the draft appears, the team has measured an intermediate output. Our recommendation is to measure the completed task, including all work needed before delivery.

This changes the questions a manager should ask. Who owns the final decision? What information does that person need? Which errors are easy to detect, and which would remain invisible until a customer complains? Where can the reviewer stop the process? Writing down those answers makes it possible to evaluate a workflow even when the model changes.

The same approach can make a worker's contribution easier to explain. Keep examples of decisions you improved, missing context you supplied, and cases you correctly escalated. Describe the underlying expertise that enabled each intervention. Counting prompts or generated documents alone will not show whether you can take responsibility for a useful result.

A small audit before a big staffing decision

For one week, log a manageable sample of comparable tasks. Include preparation, generation, review, correction, and handoff time. Record a simple quality judgment using criteria agreed in advance. Where practical, compare these with similar tasks completed using the previous process. Avoid drawing conclusions from the easiest examples alone, and keep confidential customer information out of any shared log.

Treat that audit as a starting point for a discussion, not proof that AI caused every change. Differences in task difficulty and reviewer experience can affect the result. If review consumes the expected saving, narrow the use case or improve the inputs before expanding it. If the process does help, decide explicitly whether the gained capacity should improve service, reduce backlogs, support training, or increase output.