Why AI productivity gains come with rework

The two numbers arrive together and should be read together. Workday research reported on 2 October 2026 found that employees in the United Arab Emirates save an average of 3.6 hours a week using AI, and spend 3.4 of them correcting, clarifying or rewriting what the tools produce. Almost all of the time comes back. Our view is that this is not a technology problem and not waste. 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. The checking lands on the people least able to absorb it: managers and directors do most of the validating. IBM found the same gap from the employer's side, and Singapore's official survey shows the same shape — the gains are real where AI is used, but very few firms have redesigned a process around it.

Workday reported the findings on 2 October 2026, as the company opened a Dubai office. The time arithmetic is close to a wash: 3.6 hours a week saved, 3.4 hours spent fixing. What the report does not do is separate useful rework from waste, and that is the part we would press on. A correction that improves a draft is not the same as a rewrite that should never have been needed, and the headline figure treats them alike. The survey is more encouraging elsewhere, and we would rather say so plainly: the efficiencies are not mostly being taken as cost cuts, but directed at handling more work, at training and reskilling, and at workload and well-being. What the report cannot tell you is whether the trade was worth making, because it counts the two sides of it as the same thing.

On the same day, 2 October 2026, IBM's Institute for Business Value published a study of chief human resource officers and employees, and its Canadian results show a clear mismatch. Sixty-eight per cent of Canadian CHROs said the ability to supervise, validate and override AI outputs was the workforce's most essential skill; only 29% of Canadian employees ranked judgement as important. The finding we would keep is that organisations which clearly define which workflows are human-led, AI-assisted or AI-executed report lower risk and better quality. That is Workday's point said from the other side of the desk, and it is the one we would act on: rework is not a tooling problem; it is an unassigned job.

Singapore has no matching survey of hours saved, and it is worth saying so rather than borrowing the Emirates' numbers. What it has is a firm-level picture, and it points the same way. The Ministry of Manpower's Artificial Intelligence Survey found that most firms had not begun adopting AI, and that of those which had, only a small fraction had integrated it into core business processes — the rest were still planning or piloting. Among firms actually using AI, about seven in ten reported improved worker productivity. Workday's own earlier Singapore study found the same shape among its respondents. The pattern holds: the gain shows up first, the rework shows up next, and the redesign mostly has not happened.

Singapore's government has put a name on one cause of rework. On 2 October 2026 GovTech described a context layer it is building so that AI tools querying public data receive official definitions, agency rules and pre-tested queries before they answer, with the definitions stored as version-controlled code rather than as a document somebody updates by hand. Its own worked example is the cleanest illustration of rework we have seen: a query that counted every record marked resolved returned far more cases than the policy-correct answer, because a share of them had closed automatically without a verified resolution. That fixes ambiguity in the data and says nothing about the other causes of rework. But it is a useful template for any team: decide what the words mean before the model answers, and keep that decision somewhere changes are visible.

We would treat the 3.6 against 3.4 as a diagnosis rather than a verdict. The pair says the hours exist; it does not say the work is worthless. A correction that catches a wrong figure or an invented citation is the system working, and a studio that shipped AI drafts unchecked would have a worse problem than slow output. Our complaint is that the checking is unbudgeted. It lands on managers, it is invisible in most plans we have seen, and it is the reason an AI pilot can look like a success in month one and quietly cost more by month six.

The IBM finding is the one we would act on, because it names the fix. If validating AI output is the scarce skill, then training it is the highest-return spend available — not prompt courses, but the unglamorous practice of reading an output against a source and knowing when to reject it. That is also cheap, which is why we are unimpressed that it keeps being deferred.

We are sceptical of the word savings. What AI often does is move time from producing to checking, and checking is a different job with a different cost. Until an organisation puts a baseline on rework and reviews it, the savings will keep being counted on the way in and the corrections on the way out, and nobody will be able to say whether the trade was worth making.