Machine learning saves £4.4M in UK.gov work and pensions fraud detection
(2025/10/27)
- Reference: 1761558312
- News link: https://www.theregister.co.uk/2025/10/27/dwp_machine_learning_savings/
- Source link:
The UK government's Department for Work and Pensions (DWP) has saved £4.4 million over three years by using machine learning to tackle fraud, according to the National Audit Office (NAO). However, the public spending watchdog found the department's ability to expand this work is limited by fragmented IT systems and poor cross-government data standards.
In its October 22 [1]report , the NAO praised the DWP's efforts but urged it to go further.
UK pensions dept hands Softcat £250M for Microsoft subscriptions [2]READ MORE
"DWP should build on its existing use of data analytics to explore how these emerging technologies may help to detect and prevent fraud and error," said report director Laura Brackwell.
The challenge is significant. The NAO said the [3]DWP's IT systems are not fully integrated , preventing staff from accessing complete claimant information. The department is working to develop an application to provide a single view and told the auditor that scaling up would need cross-government data standards to enable inter-departmental data sharing.
Denmark, which has introduced interoperable IT systems and government-wide data standards, has around 100 anti-fraud machine learning models.
[4]
DWP's current machine learning work focuses on Universal Credit, which is replacing a number of legacy benefits. Since May 2022, a model has flagged potentially fraudulent hardship payment advance claims for human review rather than automatic rejection.
[5]Irony alert: UK.gov Work dept hires IBM to aid AI projects
[6]UK Home Office doubles down on Oracle with £54M cloud contract
[7]Sopra Steria bags £115 million legacy extension from UK pensions department after delays to replacement ERP project
[8]How sticky notes saved 'the single biggest digital program in the world'
The NAO found fairness issues. Applicants aged 45-plus and non-UK nationals were more likely to be flagged but less likely to have claims refused. The DWP assessed this for only one of nine protected characteristics under equality law – age – as it lacked sufficient data on the others. Despite this, the model is three times as effective than random sampling and will remain in use while being improved.
Four additional machine learning models are in development, all focused on Universal Credit. These are targeting undeclared self-employment income, financial assets, undisclosed partners, and general fraud detection.
[9]
The NAO recommended the DWP standardize claimant data formats, engage with cross-government data initiatives, and extend anti-fraud efforts to other benefits, particularly Pension Credit, which had the highest overpayment rate in 2024-25.
Context matters because in 2024-25, the DWP distributed £291 billion to 23 million people – more than the UK spent on healthcare and triple its defense budget. The £4.4 million saved works out to roughly two pence per Briton annually. ®
Get our [10]Tech Resources
[1] https://www.nao.org.uk/reports/tackling-benefit-overpayments-due-to-fraud-and-error/
[2] https://www.theregister.com/2023/04/19/dwp_collars_softcat_for_250m/
[3] https://www.theregister.com/2024/09/18/dwp_business_processes_procurement/
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/publicsector&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aP9QxtBdhFCnASkDJNKyYwAAAUQ&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[5] https://www.theregister.com/2025/10/02/uk_pensions_and_benefits_department/
[6] https://www.theregister.com/2025/10/08/home_office_oracle_cloud/
[7] https://www.theregister.com/2025/08/08/sopra_steria_gets_115_million/
[8] https://www.theregister.com/2025/05/16/universal_credit_commons_committee/
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/publicsector&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aP9QxtBdhFCnASkDJNKyYwAAAUQ&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://whitepapers.theregister.com/
In its October 22 [1]report , the NAO praised the DWP's efforts but urged it to go further.
UK pensions dept hands Softcat £250M for Microsoft subscriptions [2]READ MORE
"DWP should build on its existing use of data analytics to explore how these emerging technologies may help to detect and prevent fraud and error," said report director Laura Brackwell.
The challenge is significant. The NAO said the [3]DWP's IT systems are not fully integrated , preventing staff from accessing complete claimant information. The department is working to develop an application to provide a single view and told the auditor that scaling up would need cross-government data standards to enable inter-departmental data sharing.
Denmark, which has introduced interoperable IT systems and government-wide data standards, has around 100 anti-fraud machine learning models.
[4]
DWP's current machine learning work focuses on Universal Credit, which is replacing a number of legacy benefits. Since May 2022, a model has flagged potentially fraudulent hardship payment advance claims for human review rather than automatic rejection.
[5]Irony alert: UK.gov Work dept hires IBM to aid AI projects
[6]UK Home Office doubles down on Oracle with £54M cloud contract
[7]Sopra Steria bags £115 million legacy extension from UK pensions department after delays to replacement ERP project
[8]How sticky notes saved 'the single biggest digital program in the world'
The NAO found fairness issues. Applicants aged 45-plus and non-UK nationals were more likely to be flagged but less likely to have claims refused. The DWP assessed this for only one of nine protected characteristics under equality law – age – as it lacked sufficient data on the others. Despite this, the model is three times as effective than random sampling and will remain in use while being improved.
Four additional machine learning models are in development, all focused on Universal Credit. These are targeting undeclared self-employment income, financial assets, undisclosed partners, and general fraud detection.
[9]
The NAO recommended the DWP standardize claimant data formats, engage with cross-government data initiatives, and extend anti-fraud efforts to other benefits, particularly Pension Credit, which had the highest overpayment rate in 2024-25.
Context matters because in 2024-25, the DWP distributed £291 billion to 23 million people – more than the UK spent on healthcare and triple its defense budget. The £4.4 million saved works out to roughly two pence per Briton annually. ®
Get our [10]Tech Resources
[1] https://www.nao.org.uk/reports/tackling-benefit-overpayments-due-to-fraud-and-error/
[2] https://www.theregister.com/2023/04/19/dwp_collars_softcat_for_250m/
[3] https://www.theregister.com/2024/09/18/dwp_business_processes_procurement/
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/publicsector&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aP9QxtBdhFCnASkDJNKyYwAAAUQ&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[5] https://www.theregister.com/2025/10/02/uk_pensions_and_benefits_department/
[6] https://www.theregister.com/2025/10/08/home_office_oracle_cloud/
[7] https://www.theregister.com/2025/08/08/sopra_steria_gets_115_million/
[8] https://www.theregister.com/2025/05/16/universal_credit_commons_committee/
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/publicsector&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aP9QxtBdhFCnASkDJNKyYwAAAUQ&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://whitepapers.theregister.com/
To add some further perspective
Like a badger
1) Overpayment and fraud of £9.5bn is 3.3% of total benefits paid - I'd reckon a well run system should achieve something of the order 0.3-0.7%
2) DWP identify £4.5bn of overpayment and fraud stopped currently...so
3) Manual processing stops about a third of benefits error and fraud, and...
4) Machine learning has saved one-thousandth of the amount saved by manual processes
And finally, total benefits error and fraud is about one fifth of the error and fraud in tax collection of £47bn in 24/25 (which doesn't include tax avoiding big tech firms).
A footnote: The link "DWP's IT systems are not fully integrated" leads to a story that is not about the operational systems that are implicated in overpayment and fraud, but refers to back office support services systems like HR, procurement, accounting.
Here's [1]fraud detection fucking up for people in Northern Ireland.
Admittedly, this was HMRC (which administers the benefit in question) and it's not clear if machine learning has been used. But I'm sure DWP won't make any these kind of mistakes. Nah, who am I'm kidding?! Of course they will accuse legitimate users of fraud.
[1] https://www.theguardian.com/uk-news/2025/oct/26/ni-parents-caught-in-uk-crackdown-lose-child-benefit-after-travelling-via-dublin