The Unclaimed Hour

The Unclaimed Hour is the capacity AI creates that nobody decides how to use. Every automation returns time. In most organisations no one owns the question of where that time goes, so it is absorbed silently into more of the same work, and the gain disappears without anyone being able to say when. Where nobody decides, drift decides.

The question for a leadership team is therefore not how much time AI saves. It is who has claimed the hour.

SuperSkills uses the term to describe this pattern. No claim of first use is made: the phrase is not documented here with a dated first publication.

Why the time vanishes#

Three mechanisms, and none of them requires anyone to behave badly.

Nobody owns the surplus. Automation is justified on efficiency, delivered by a function, and the time it returns arrives with individuals. No one is accountable for what happens next, and unowned capacity behaves the way unowned anything behaves.

The work expands to fill it. If a report took a day and now takes an hour, the default is more reports rather than a shorter week or a deeper report. Nobody decides this either. It follows from queues that were always longer than capacity.

The saving may not be real, and nobody checks. This is the uncomfortable one, and it has to be put carefully. In the METR randomised trial, experienced developers were measured 19 per cent slower with AI tools while believing they were about 20 per cent faster. METR withdrew that speed figure as a current signal on 24 February 2026, so it is early-2025 evidence and not a claim about today. The forty-point gap between what those developers experienced and what they believed is untouched by the withdrawal. That gap is the part that matters here: an organisation planning against self-reported hours saved is planning against a number nobody has measured, in either direction.

Seven and a half seconds against fifty-five#

The paragraph above leans on a gap between belief and measurement, and until recently the sharpest version of that gap was a figure its own authors have withdrawn. It has since been measured directly, on a much larger sample, by people who set out to measure exactly it.

Yu and colleagues ran three preregistered studies through Prolific, 2,691 participants in total, on 24 short tasks drawn from a published taxonomy of what people ask AI to do. Every task was designed to be finishable in under five minutes unaided. Their sentence is the one to carry: "On average, people predicted AI assistance to save time by 55.7 seconds when it only saved 7.5 seconds." Graded entry.

Three things inside that are worth more than the headline.

The error is one-sided. People were well calibrated about how long the task would take them alone, 99 seconds predicted against 93.7 actual. The whole of the miscalibration sat on the assisted side, where they predicted 43.3 seconds and took 86.2. So what people misjudge is the cost of working with the model, rather than anything about their own capability.

Most of that cost is the prompt. Writing the instruction took longer than reading and processing the answer, 48.7 seconds against 37.6, and 41 per cent of prompts were the task instructions copied and pasted straight back in. On the easiest variants the whole exchange ran 10.0 seconds slower than simply doing the task, 70.2 against 60.2.

Use begets use. Participants who had been given AI on an earlier block chose it on 44.5 per cent of later tasks, against 27.7 per cent of those who had not. The habit entrenches faster than the calibration improves.

The honest limits are the authors' own. Every task was under five minutes, so nothing here speaks to a week of professional work, and their stated focus is "human miscalibration rather than AI capability". Two of the three studies compare groups rather than the same person's forecast against their own behaviour. And this paper is a working paper with no stated venue, published one day before the same seven authors' peer-reviewed CogSci paper, so the two are one research programme and must not be set beside each other as though they were independent corroboration.

For this page the use is narrow and firm. An organisation building a case on hours returned is using a number whose own subjects overestimate it by a factor of seven on tasks lasting a minute and a half. That says nothing about whether the hour exists at scale. It says a great deal about whether anyone asked before spending it.

Time absorbed rather than converted#

The aggregate picture is consistent with time being absorbed rather than converted.

Humlum and Vestergaard found precise null effects on earnings and hours two years after ChatGPT across roughly 25,000 Danish workers, ruling out effects larger than 2 per cent, alongside substantial task reorganisation. Work changed. The measurable return did not appear.

Acemoglu's modelling puts total factor productivity gains at no more than 0.66 per cent over ten years, revised below 0.53 per cent. That is not the profile of a technology delivering large recoverable time savings to the economy, whatever it delivers to an individual on a given afternoon.

What the evidence does not show is that the hour is being wasted. Absorbed and wasted are different. Some of it goes into more work of the same quality, some into slack that people badly needed, and nobody has measured the split. Anyone claiming to know the proportions is guessing.

What the hour does not get spent on#

The unclaimed hour matters because of what it does not get spent on.

The capabilities this research argues are appreciating, judgement, verification, the ability to notice when a confident answer is wrong, are all built by practice, and practice needs time that somebody has deliberately protected. If the returned hour is absorbed into throughput, the organisation has converted a capability opportunity into volume and will not notice until it needs someone who can tell a plausible answer from a correct one.

That is capability debt accruing through a mechanism nobody is watching, because time absorbed into more work looks like productivity on every dashboard anyone runs.

There is a sharper version for leaders. If you cannot name what the saved time became, you did not save time. You changed the composition of the work, which may be fine. It is a different claim from the one in the business case.

Claiming it#

• Name the beneficiary before you deploy. Who gets the hour, and for what? Written down, in the business case, alongside the efficiency number.
• Protect a portion for practice. The single highest-return use, and the one that never survives contact with a deadline unless it is scheduled.
• Measure what it became, not that it existed. "We saved 200 hours" is unfalsifiable. "Those hours went to X" is checkable, and usually reveals that nobody knows.
• Do not trust self-reported saving. The METR finding is enough to retire the practice on its own. See how to measure AI adoption properly.
• Consider giving some of it back. Slack is not waste. An organisation with no slack cannot absorb a shock, train anyone, or think.

Related SuperSkills research#

On the pattern underneath it, drift versus design. On what erodes when practice stops, capability debt and the missed reps. On measuring properly, usage theatre and how to measure AI adoption properly.

Key sources

• Yu, S., Cheng, M., Jabbar, A., Sucholutsky, I., Collins, K. M., Jurafsky, D. and Hawkins, R. D. (2026). The efficiency-gain illusion. arXiv 2605.22687, working paper.
• METR (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.
• Humlum, A. and Vestergaard, E. (2025). Still Waters, Rapid Currents.
• Acemoglu, D. (2024). The Simple Macroeconomics of AI.
• Brynjolfsson, E., Li, D. and Raymond, L. (2023). Generative AI at Work.

About this research#

Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company.

How this research works · Reviewed quarterly · Found an error? Tell me and it is corrected on the page.