Video
Status
In production since 2021 and in daily use by the triage and fraud handling team. Our Head of Claims estimates a cumulative return of approximately £2.6 million by August 2026. Two extensions have completed development and go live this year, allowing the system to surface claims that resemble ones already under investigation, and tracing links from a known syndicate to the other claims connected to it.
FLAGSHIP PROGRAMME
Motor claims fraud scoring — UK
Summary
Our UK motor claims team has built its triage around a machine learning system developed by OXFIDA. It scores the day's claims in an overnight run, so handlers start each morning with the highest-risk cases at the top of the queue. The same run retrains the model on the decisions handlers made the day before. It has been in daily use since 2021, and our Head of Claims estimates that it has generated a cumulative return of approximately £2.6 million as of August 2026.
The problem
Before 2021, escalation ran on a fixed set of rules and written criteria applied to every claim. Rules catch fraud that has already been seen and documented, but do not scale to patterns not seen before. Adding a new rule can take months of losses, and fraudsters can still reverse-engineer them and submit seemingly valid claims that sit within the rulebook.
Solution
The system is trained on claim history, claimant behaviour, vehicle data and external checks, and returns a priority risk score in the application the triage team already works in. Handlers decide what should be escalated to the special investigation unit.
Fraudsters change their methods faster than any specification can be rewritten, so the team retrains the model on its own decisions every night. Every judgement a handler records goes back into training in the same overnight run. Five years in, the model has learned from the team's decisions rather than from a specification written at the start, which is why it tracks how fraud moves instead of becoming a second rulebook.