TL;DR

A National Audit Office report published on Wednesday 21 October finds that the Department for Work and Pensions is getting real results from small-scale innovation, including machine learning and generative AI. DWP estimates one predictive model keeps an extra 715 child maintenance cases a month from falling into arrears. The NAO wants the department to give staff clearer limits on acceptable risk and to share lessons more widely.

What the auditors looked at

The NAO examined six examples of innovation across DWP, which pays benefits and pensions to around 23 million people in Great Britain. The NAO highlights two that use AI, and a third that automates the reading of customer post.

The most detailed figures come from the Child Maintenance Service. A machine-learning model estimates how likely each case is to break down, so caseworkers can step in before payments stop. By DWP’s own estimate, an extra 715 cases each month now avoid slipping into arrears, and the model saves staff time equal to 49 full-time posts.

A second tool, GAIL (Generative AI in Learning), was built to cut the time it takes to produce training material for DWP staff. A third, the Whitemail Insights and Vulnerability Scanner, reads roughly 25,000 items of customer post a day that are not standard forms, helping staff spot potentially vulnerable people and decide more quickly what to do.

The remaining three were a project to rethink benefits administration, text-message nudges for Carer’s Allowance claimants, and jobcentre trials assigning work coaches to particular sectors.

The auditors’ caveats

The NAO credits DWP with adjusting its approach to risk for each project and with piloting ideas before scaling them. But it notes the limits the department works within: benefits must still arrive on time for millions of people, and the personal data DWP keeps must stay protected.

Its recommendations focus on culture. Staff need practical guidance on where the boundaries of acceptable experimentation lie, and DWP should spread lessons more consistently, including from projects it dropped, so teams do not repeat mistakes. Gareth Davies, head of the NAO, said the department should now give staff “the confidence to test new ideas safely”.

Looking forward

The child maintenance figures are DWP’s own estimates rather than independently measured outcomes, which is worth bearing in mind. Even so, the report gives an auditor’s view of where AI is paying off inside a large UK department. It makes a useful companion to our report on GDS’s pilots on legacy systems. In our view, the NAO’s point about sharing lessons from abandoned projects applies well beyond DWP: knowing why an AI pilot was stopped can be as useful to the next team as a success story.