The Invisible Work AI Creates — and Who Ends Up Owning It
IBM surveyed 1,500 CHROs and 8,800 employees: 80% of leaders say AI creates invisible work, and 43% of employees say they get blamed when it goes wrong. Here is how to time both halves of a workflow before calling it improved.
A workflow moves to AI. The first draft gets faster. The review queue gets longer soon after, and nobody put that second part on the timeline.
What is "invisible work" in an AI-assisted workflow?
It is the validating, correcting, context-adding and exception-handling that an AI output still needs — work that existed before AI arrived, now concentrated onto whoever reviews the output, and rarely counted anywhere.
IBM's Institute for Business Value surveyed 1,500 CHROs and senior workforce executives, plus 8,800 employees, for its 2026 CHRO study. 80% of the CHROs said AI creates this kind of invisible work. 43% of the employees said they are blamed when AI goes wrong.
IBM sells AI technology and consulting, so read the incentive alongside the finding. The sample is global, not Czech or Slovak, and it describes a pattern across many organisations, not a guarantee about any one rollout.
Why does this break the ROI story before anyone measures it?
Because most AI timing comparisons stop at generation, and the invisible work happens one step later.
A workflow gets timed at the point where AI produces a draft — an email, a summary, a first pass at code. Faster. Everyone moves on. What doesn't get timed: the person who checks that draft against the source, catches the exception the draft missed, and owns the outcome if it was wrong. That step existed before AI arrived. AI changed who does it and how visible it is, not whether it exists.
Who ends up doing the invisible work, and why does that matter?
Usually whoever is closest to the output when something goes wrong — and that is rarely the person who decided to roll AI into the workflow.
A support agent using AI to draft replies still owns the reply that goes out. A developer using AI to write code still owns the pull request. A finance analyst using AI to summarise a variance still owns the number that reaches the board pack. If the review and exception work is not named as a role with time attached to it, it lands on whoever is standing closest to send — and 43% of IBM's surveyed employees said they carry blame for exactly that.
How do you make invisible work visible before calling a workflow improved?
Time both halves of the workflow, not just the half AI touches.
Before changing a workflow, write down what the reviewer checks, how long that check currently takes, and who owns the exception path. After the change, measure the same three things again. If generation time fell but review time rose by nearly as much, the workflow has not improved — the cost moved, and the report should say so.
This is the same discipline behind every AI Workflow Audit: map the whole handoff, not the part that photographs well. Saved time is released capacity for the team, not an automatic euro figure on a spreadsheet — and it only becomes either once someone assigns the freed hours to something.
The question worth asking this week
Pick one process AI already touches. Ask the person doing the review: what do you check, how long does it take, and what happens when you find something wrong? If nobody can answer in under a minute, the workflow was never fully mapped — only the part in front of the model was.
The first draft got faster. Whether the workflow got better depends on what happened to the person checking it.