ONS turns to AI to fix faulty UK jobs data

The Office for National Statistics (ONS) is using AI across its statistics production as it works to replace the faulty labour market survey behind Britain’s unreliable jobs figures.

Speaking to the Financial Times, the stats bods said it expects to save thousands of hours a year through its new AI tools.

It is midway through a turnaround after long-running internal problems damaged its jobs data and several other core economic outputs, and has scaled back work in areas including health and crime to protect its central economic statistics.

It also plans to cut its total outputs by 10% to concentrate resources on the figures that matter most to government and markets. Earlier in August, it warned that the effort to replace its broken labour market survey was increasingly squeezing its capacity to invest in improvements elsewhere.

The ONS received an extra £100 million a year to prepare for the 2031 census, but its core funding will fall by almost 10% in real terms over the next two years. This has made the move to AI something of a necessity.

“AI is mainly a productivity play,” said James Benford, ONS director-general for economic statistics, who told the paper that the agency is trying to do more with its existing workforce and free up efficiency gains it can reinvest.

Benford joined in June 2025 to help steer the turnaround and has pointed to a queue of outstanding quality problems, alongside new GDP standards, the census, the transformed labour force survey and the statistical business register.

The ONS became one of the first national statistics agencies anywhere to use AI in the production of official figures when it introduced a tool built on Google’s enterprise large language model last year.

That tool reads survey responses and assigns job and industry classifications, work that previously ate up analyst time. Andrew Banks, ONS lead data scientist, put the saving at roughly 350 hours a year across two surveys, alongside the rise in industry classification accuracy from 71% to 80%.

The same approach now feeds into the transformed labour force survey, the replacement for the broken jobs data. The system adds a follow-up question about a respondent’s job on the fly, mimicking the back-and-forth an interviewer would normally use to pin down the right industry code.

A trial with 1,000 people classified 43% of those who received the extra question more accurately, and completion rates held steady. Benford described the digital-first nature of the survey as one of the trickier parts of the project, and said the AI is designed to reproduce a conversation that would otherwise involve a human interviewer.

Now read: Britain’s productivity is quietly recovering – but it’s not AI driving it