Status
Under active build with a dedicated team working directly with our UK motor claims business. The first release assembles the entity graph from claims documents; the networks it surfaces will be assessed by claims investigators before the next stage is scoped.
FLAGSHIP PROGRAMME
Organised fraud network detection
Summary
Organised fraud is designed to look ordinary one claim at a time. Each individual claim passes review because, on its own, there is nothing wrong with it. What gives a fraud network away is the relationships across claims: a repair shop, a phone number, a vehicle or a witness that keeps reappearing across cases that should have nothing to do with each other. But fraudsters leave a trail across the documents they file themselves: statements, invoices and correspondence that are read one claim at a time, and never against each other.
OXFIDA is building a system that reads those documents, links the people, vehicles and businesses named across them, and surfaces the connections that no individual claim assessment can see.
The problem
Claims fraud detection works well at the level of the single claim, and our UK motor business has scored claims that way since 2021. But a model that examines one claim at a time cannot detect coordination between claims, because the evidence of coordination is not inside any of them.
Networks can also be identified once they are known, by tracing outward from a confirmed case. Harder still are the networks nobody has identified. A model can only be trained on fraud that was caught, and organised networks are typically caught late, after years of claims.
Solution
Extracting names from documents and linking them is not new, and rule-based link analysis has been available to insurers for years. However, language models can now read a statement or a report much as an investigator would, following meaning across documents. And graph learning methods can learn what ordinary connection structure looks like across a claims portfolio, so the system surfaces structure that does not fit instead of waiting for someone to write a rule describing it.
The system resolves the same person or business when they appear under different spellings across claims, and assembles what it finds into a graph of who and what is connected to whom. Because it looks for contradiction between sources and for structure that does not belong, rather than for patterns someone has already documented, it does not depend on confirmed examples of past fraud.