Who pays for the AI rework?

The two figures arrived on the same day and read as one story. Workday's research, reported on 2 October 2026, put the saving at 3.6 hours a week and the repair at 3.4 of them; on the same day IBM's Institute for Business Value found 68% of Canadian HR chiefs calling the ability to supervise and override AI output the most essential skill, while only 29% of employees ranked judgement as important. One says the hours come back. The other says the people who have to catch the mistakes do not think catching them is their job. Our view is that the rework is neither waste nor a technology problem: it is work that was never assigned to anyone, and until it is named and costed, an organisation cannot tell a saving from a delay.

A productivity problem is one you solve by buying something better. This one survives an upgrade, because it is not about how good the output is — it is about who is accountable for it. Checking is the part of the work that AI moved rather than removed, and a correction that catches an invented figure or a citation to a document that does not exist is not overhead. It is the system working.

Our objection is narrower and more practical than the headline suggests. The checking is real work performed by real people with real weeks, and in almost every plan we have seen it is invisible. It sits with managers and directors, which is to say with the people least able to absorb it. That is not a flaw in the tool. It is a gap in the plan.

Workday's pair of figures says the time exists. It cannot tell you whether the trade was worth making, because it adds together two very different things: a correction that improves a draft and a rewrite that should never have been needed. Both are counted as 3.4 hours, and only one of them is a problem.

The finding we would actually act on is the gap between the two sides of the desk. Employers have named the scarce skill — supervising, validating and overriding what the model produces. Employees, in the same study, ranked judgement low. That mismatch is not cynicism and it is not laziness; it is a job description nobody wrote. Asking people to catch what a machine got wrong, without saying it is now part of their job, is how good intentions turn into quiet resentment.

Three things, none of them technical. Name the checker, in writing, for each output that leaves the team — a name, not a role. Put rework on the baseline: how long did this take to check last month, and is that number going up. And train the judgement rather than the prompting; reading an output against its source and knowing when to reject it is a practice, not a trick, and it is the cheapest thing on the list to teach.

The best illustration of the underlying fault we have seen is small and unglamorous. GovTech described on 2 October 2026 a layer that gives government AI users official definitions before they ask, and its worked example was a query returning 142 resolved cases where 37 had closed automatically with no verified resolution — so the correct answer was 105. The rework there came from a word nobody had defined. Deciding what the words mean before the model answers is cheap. Discovering the disagreement afterwards is not.

Singapore's firm-level survey points the same way, and we would rather quote it plainly than dress it up: of the minority of firms actually using AI, about seven in ten report better worker productivity, while only a small fraction have integrated it into core processes. The gain shows up first. The redesign mostly has not happened.

We are sceptical of the exact pair. Self-reported hours are a soft measurement, and 3.6 against 3.4 invites a precision the method cannot carry; we would not plan a project around the difference between them. Nor are we sure the two categories travel between countries and industries as neatly as they are reported.

We would also resist the framing that checking is a tax to be minimised. Some of it is the price of the speed, and a studio that shipped unchecked output would have a far worse problem than slow work. The honest position is that we do not know the right ratio — and that nobody who has not measured their own rework knows theirs either.

Our position is simple: the work moved, and nobody moved with it. Read the figures as a diagnosis of an unassigned job rather than evidence that AI was a mistake, and the fix becomes unexciting and cheap — name the checker, measure the checking, teach the judgement, look at the number again next month.

We are unimpressed by the word savings, because it is usually counted on the way in while the corrections are counted on the way out, and the two are almost never put side by side. Until an organisation does that, it is not measuring productivity; it is measuring enthusiasm.

We would rather see a low, clear standard that people can meet honestly than a high one that only well-advised teams can prove. And we would rather have this argument early, while the numbers are still small enough to argue about, than discover in three years that the checking was never anybody's job at all.