Britain's official productivity figures cannot see the machine at all, and the two methods the statisticians use disagree by a factor of five

If AI were delivering the gains its buyers expect, the place it would eventually appear is output per hour worked. That is the number, more than any other, that decides whether a country gets richer.

The Office for National Statistics published its most recent estimate on 19 May, covering the first quarter of this year. Output per hour worked was 0.4 per cent higher than the same quarter of 2025. Output per worker was 0.1 per cent lower. Gross value added rose 1.1 per cent over the year while hours worked rose 0.7 per cent, and output per hour finished the quarter about 2.6 per cent above where it sat before the pandemic.

For scale: British productivity grew at roughly 2 per cent a year for most of the post-war period, and has grown far more slowly since the financial crisis. A reading of 0.4 per cent is not a technology arriving. It is the same weak trend the country has been running for fifteen years.

Set that beside what workers say. In survey after survey, the majority of AI users report meaningful personal time savings, with one recent study putting it at roughly eleven hours a week. If a third of a working week were genuinely being freed across a large share of the workforce, something ought to be visible in the national figures by now. It is not.

There are three honest explanations and the evidence does not yet separate them.

The first is timing. Historically, general-purpose technologies take one to two decades to show up in productivity statistics, because the gains require the surrounding work to be redesigned rather than merely augmented. On this reading the number is not wrong, only early.

The second is that the saved time is being consumed by new work created by the same tools, a case made elsewhere in this edition.

The third is measurement, and it is the one that should trouble anyone who plans to quote a productivity figure this autumn.

The ONS publishes its Quarter 1 estimate on two different bases. The measure built on the Labour Force Survey gives output per hour at 0.4 per cent and output per worker at minus 0.1 per cent. The measure built on administrative data, principally PAYE Real Time Information, gives 2.1 per cent and 1.6 per cent for the same quarter and the same economy.

The two disagree by a factor of five on the headline, and on direction for output per worker: one has it falling, the other has it rising by more than one and a half per cent.

The reason is known and the statisticians say so plainly. Response rates to the Labour Force Survey collapsed and have not fully recovered, and the ONS notes that historically low responses continue to affect annual growth rates in the headline productivity data. It has also published a correction to the seasonal adjustment on its output-per-hour-by-division dataset, caused by an inconsistency around the 2020 break week. That is a statistical agency doing its job in public. It is also a warning label.

None of this is a scandal, and it is worth saying so before the point is overread. Every statistical agency is dealing with falling survey response, and the ONS is more transparent about it than most. Absence of precision is not evidence of anything in particular.

But the consequence is concrete and it has a date on it. The Chancellor confirmed on Friday that the Budget will be held on Wednesday 28 October. The fiscal forecasts underneath it rest on assumptions about productivity growth, and the Office for Budget Responsibility has begun factoring AI-driven productivity into those forecasts. The next productivity estimate lands on 18 August, the labour market release the same week.

So a forecast covering the next five years will be built on a series whose two versions currently disagree about whether output per worker went up or down last quarter.

Two details worth holding on to, both from the ONS release. Information and communication made the largest positive contribution to productivity growth relative to the 2019 average, driven by output rising faster than hours. Human health and social work activities made the largest negative contribution. The sector most associated with the technology is pulling the average up. The sector employing more British people than any other is pulling it down, and roughly one and a half million people work in it.

What would change the picture. If the August release shows the two methods converging, the measurement explanation weakens and the timing explanation strengthens. If they continue to diverge, then anyone citing a single productivity number in the Budget debate, in either direction, is citing a choice of method rather than a fact about the economy. That includes this paper, and we will hold ourselves to it.