AI saves hours a week — the question is how many survive the checking

Google, Google DeepMind and MIT FutureTech surveyed more than 600 scientists and found they save just below seven hours a week with AI — and the same research shows where that time goes: into validating outputs, into a backlog of untested hypotheses, into bottlenecks in experiments and clinical validation. The individual gain is real; the net gain is much smaller, and three other datasets agree. In Europe the median user saves three hours a week, but only half of workers save any time. In Singapore, about half of workers lose at least an hour a week correcting AI output.

Published through Google's AI and Economy ATLAS programme on 15 September 2026, the study analyses 2,600 specialised AI models and surveys more than 600 scientists in the United States and Britain. Nearly half use some form of AI every day and report saving just below seven hours a week. The paper then reports significant time spent validating AI outputs, a growing backlog of hypotheses waiting to be tested, and bottlenecks in physical experimentation and clinical validation. The authors call it an early snapshot, note that the saving is self-reported, and conclude that realising it will need scientific processes redesigned, not just faster models.

McKinsey's State of AI in 2026, published on 25 August, puts the same gap in money. Of 1,719 respondents, 80 per cent say AI has improved their own productivity, yet only 37 per cent attribute at least some EBIT impact to AI, essentially unchanged from a year earlier, and the share of genuine high performers has stayed flat at about 6 per cent. Meanwhile 44 per cent say AI is being scaled across the organisation, up from 38 per cent a year earlier: the tools are spreading faster than the results.

New Relic's 2026 State of AI Coding report, released on 10 June 2026 from a survey of 200 United States technology decision-makers, measures the checking bill precisely. At the moment of review, 94 per cent of leaders rate AI-generated code as higher quality than human-written code. Once it ships, 78 per cent report more incidents, 86 per cent report senior staff spending more time fixing code, and 74 per cent say at least a quarter of AI code needed significant rework over the past year. Code is only the easiest place to see this, because a defect arrives with an invoice attached; the same review burden sits wherever output is plausible but unverified.

Workday's Singapore study, released on 26 February 2026 from a global survey of 3,200 respondents, found about half of Singapore respondents spend at least an hour each week clarifying, correcting or rewriting AI-generated output, with 12 per cent spending two to four hours a week on rework alone. Its ASEAN research, released on 11 August 2026 and conducted by the Harris Poll among 6,100 professionals, adds the second cost: around one in five employees in ASEAN lose more than seven hours a week moving information between systems that do not talk to each other, and only 30 per cent of organisations have embedded AI in the core of the business.

The European Central Bank's blog on 26 August 2026 shows how quickly a headline saving shrinks across a whole population. AI use at work rose from 26 per cent of workers in 2024 to 41 per cent in 2025 and 52 per cent in 2026. The median user reports saving three hours a week, about 7.7 per cent of working time — but only 48.8 per cent of workers save any time at all, which brings the economy-wide figure to roughly 3.8 per cent. The biggest gains are also the narrowest: code generation and debugging saves almost eight hours a week, yet only about 8 per cent of workers use AI for it.

Three habits make the net number bigger. Pick one recurring task rather than rolling out a tool. Write down what a finished result looks like before you start, including who signs it off. Then measure the net, not the gross: hours saved minus hours spent checking. For a small studio the equivalent is already familiar — the script, the shot list, the approved still — and each exists to catch an error while it is cheap.

The rework is not evidence that these tools are disappointing. It is evidence that most people deploying them never decided what a good output looks like, so that decision is deferred to the moment of review, at the highest price and by whoever happens to be holding the file. We would read the Google and MIT finding as the useful one precisely because it is modest: the hours are real, and they arrive as a queue of unfinished work rather than as savings.

We would be careful with the numbers. The coding survey is American technology leaders at large firms, and the Workday and ECB figures are self-reported, so they establish a direction rather than a precise rate. Anyone quoting a single percentage for AI productivity, ours included, is quoting a gross number.