When separated parents use the Child Maintenance Service’s Collect and Pay service, the government takes the money from one parent and passes it to the other. Some months the money stops. The UK’s Department for Work and Pensions (DWP) estimates that around 8,000 child maintenance cases a month see an unplanned breakdown in payments. Once a case slips into arrears, getting the money moving again takes far more work than keeping it on track.
The clever bit
At the start of every month, a machine-learning model reads the payment history, changes of circumstance and recent letters and calls on about 100,000 cases that are currently paying, then scores how likely each one is to miss an upcoming payment. Each score comes with a short explanation of what is driving the risk. More than 40 caseworkers use the high, medium and low flags to decide whose case to open first, check the file and contact the parent before a payment slips.
The model became part of everyday work in March 2024, run by a new Early Intervention Team. DWP told the National Audit Office (NAO) that the early contact it prompts stops an extra 715 cases a month from moving into arrears and saves staff time equal to 49 full-time employees. In testing, the final proof-of-concept model was right 70% of the time, up from 65% for the first version in May 2022. Building it cost £625,000, and it costs about £460,000 a year to run.
Hooray
In 715 more cases a month, by DWP’s estimate, the maintenance a child is owed keeps arriving, and a caseworker sorts out the wobble early instead of chasing a debt later. The NAO, Parliament’s independent spending watchdog, picked the model out as one of six DWP innovations with lessons for the rest of government.
What comes next
DWP plans a full evaluation once more data has gone into the model in late 2027. The Child Maintenance Service is also looking at other ways machine learning could help it run the service.