TL;DR
The DWP is using a machine learning model, in a “limited live test”, to help staff decide which Universal Credit claims go forward for a Targeted Case Review, The Independent reported on 18 November. Dame Sarah Healey, the department’s most senior official, disclosed the trial when writing to the Commons Public Accounts Committee. The department says a member of staff still makes every decision on a case.
What the letter says
Targeted Case Reviews are an anti-fraud and error programme that reassesses more than a million claims each year, according to Healey. The model supports the agents who choose which cases to examine. Healey wrote that “human consideration remains central”, and that the tool does not decide whether a review goes ahead, what a person is entitled to, or whether payments are changed, suspended or stopped.
She added that the trial runs under the department’s normal safeguards and equality standards, and that the DWP would “consider alternative options” if it underperforms. The criteria the model uses to flag claims will stay confidential, she said, to protect the “integrity of the benefit system”.
A DWP spokesperson said the review programme has already saved taxpayers £1.7 billion by stopping incorrect payments. According to The Independent, the scheme is meant to save £13bn before the decade is out.
The bias question
Earlier in 2026, a study from the Administrative Fairness Lab showed that just 21% of reviews identify any incorrect payment, and that many claimants face repeated reviews, sometimes months apart. Mark Bennett, a law lecturer at the University of Liverpool and a member of the lab, called machine learning a “potential step change” that “comes with clear risks”.
He pointed to an earlier DWP tool for screening Universal Credit advance payments, which sent a disproportionate share of claimants who were older, disabled or from outside the UK for fraud checks. “Given the huge scale of the scheme, the impacts of any such mistakes are likely to be widely felt,” he said.
Looking forward
The trial follows a National Audit Office report that backed the DWP’s AI trials but asked for clearer risk limits. In our view, the confidentiality of the selection criteria is the sticking point: with the department’s own history of skewed flagging, the obvious question for the Public Accounts Committee is how anyone outside the department can check this model for the same problem.