Britain has three ways of counting what happens to work. None of them can tell you whether a machine did it

This desk has reported each of these feeds separately over the past month. Put together, against a speech about disruption to work, the shape of the gap is clearer than any one of them shows alone.

What arrives fastest

Every time an employer runs payroll, it files with HM Revenue and Customs. So the state knows within days how many people were paid and roughly what they earned. That is close to real time, and it is the best instrument in the system.

It records that a person stopped being paid. It cannot record why. A payroll file does not have a field for a machine.

What carries the detail

The Labour Force Survey asks households about work, and it is the only source of occupation, hours and whether someone is looking. It is also a survey with known response problems, currently published with a warning attached, and it reports on a rolling three-month window. By the time a change shows up in it, the change is at least a quarter old.

What connects the technology to the work

The Business Insights and Conditions Survey asks firms directly, every fortnight, and has asked about AI since 2023. The most recent bulletin, published on 20 July, drew 38,637 responses at a response rate of 26.7 per cent. It is the only feed that links the technology to the job.

It is voluntary, it excludes six sectors including finance, and it asks businesses to describe their own behaviour.

The three things that are missing

First, a question that used to exist. Employers notifying the state of redundancies had to give an occupational breakdown until the HR1 form went digital on 1 December 2025. That question is no longer asked, so the records of who is being made redundant no longer say what they did.

Second, a figure that was never collected. In July the ONS answered a freedom of information request by saying its Vacancy Survey asks a single question, so it holds no information on how many vacancies are entry-level, or which occupations employers are recruiting for. There is a dataset built from online adverts, and the ONS says it is not a direct measure.

Third, an assessment that was not repeated. The Department for Science, Innovation and Technology published an assessment of AI capabilities and the labour market with the AI Security Institute on 28 January. Nothing has replaced or updated it.

What that leaves

Britain can tell you, quickly, that people stopped being paid. It can tell you, slowly and with a warning, what jobs people say they do. It can tell you, fortnightly and voluntarily, what firms say they are doing with AI.

It cannot tell you which occupations are disappearing, or whether AI is why. On the current arrangements, if somebody asks in five years which jobs went first, the honest answer will be that nobody wrote it down.

One exception worth naming

There is a counter-example, and it is not in employment statistics at all. The NHS commissioned an independent evaluation of ambient voice tools, with its method published before its results, reporting in February 2027. It is the only one this desk has found in any British sector. It shows the thing is possible.