TL;DR
Britain’s Office for National Statistics is embedding AI in how it produces official figures, with a single planned tool forecast to free up 7,500 hours annually. The rollout comes as its core budget shrinks by nearly 10% in real terms and the repair of its jobs data drags on. One classifier has already pushed coding accuracy from 71% to 80%.
Automating while under repair
Two years of turnaround work followed the loss of confidence in the agency’s labour market figures, which forced retreats from subject areas including health and crime. Its own assessment this month was blunter still: rebuilding that survey is crowding out improvements everywhere else. AI arrives in the middle of that, pitched openly as a way to recover capacity.
James Benford, who took charge of economic statistics in June last year, was direct about the motive. “AI is mainly a productivity play. We’re trying to do more things with our workforce … and find big efficiency gains that can be reinvested.” Quality gaps remain open, he said, alongside fresh GDP standards, a census in 2031, and rebuilt survey and register systems. Census preparation draws an extra £100m a year; the core budget has no such shield.
What the tools actually do
A classifier running on Google’s enterprise language model has spent the past year sorting survey answers into occupation and industry codes, among the first such uses inside any national statistics body. The agency’s lead data scientist, Andrew Banks, put the saving at roughly 350 hours annually across a pair of surveys, with accuracy up nine points to 80%.
The larger prize lands early next year: a scanner for the 90,000 or so receipts households submit to the spending survey that underpins inflation weights. Processing one household’s paperwork currently takes three hours. Banks expects seconds, and a conservative 7,500 hours freed, with the spare capacity going into a bigger sample. A third tool asks a follow-up when a respondent’s job description is ambiguous, standing in for the interviewer who would once have probed. Of 1,000 people trialled, 43% were coded more accurately, and none abandoned the survey as a result.
Looking forward
Standard AI tools now reach 5,000-plus of the agency’s 5,980 staff, and coding assistants run daily for 120 data scientists. Benford was careful to say headcount is not the target.
The circularity deserves a moment. These are the figures the Bank of England, the Treasury and everyone arguing about AI’s economic effect rely on — a debate Resultsense has covered from both sides in recent weeks. The producer is now also a user. It lands as Andy Burnham’s government leans harder on public-sector adoption, with a task force under Lord Patrick Vallance and an AI minister lifted to cabinet rank. Of the deployments underway across Whitehall, this is the one least able to absorb a bad outcome: degrade quality here and the loss is not one department’s efficiency but the reliability of the numbers everything else is steered by.