Can an AI determine which potentially fraudulent motor insurance claims are worth contesting? Yes, says Clyde & Co’s Damian Rourke
Imagine you’re a claims manager for a motor insurer who’s been given 100 potentially fraudulent first notifications of loss to review. The question you’ve been asked is a simple yet stark one: of these 100 cases, how many is the insurance company likely to successfully repudiate and how many will it have to pay? It’s a question that demands all your insight, knowledge and years of experience. On the face of it, it seems to be a question best suited to human intelligence rather than machine intelligence. But new technology means that is changing. Rapidly.
The idea of using artificial intelligence (AI) and algorithms to detect fraud has been around for a long time. It’s been used in banking for several years and insurers have long been aware of the concept, but our sector has been held back by one major problem: the quality of the data insurers hold. Data is the lifeblood of AI. The more data an algorithm—a set of rules—can crunch, the more it learns. And even though insurers store vast quantities of data, the formatting is inconsistent, and its storage varies from business to business. This has held back our efforts to most effectively utilise the power of AI to detect fraud.
One other important difference between banking and insurance in this area is that banks detect fraud after the fact. They identify a fraudulent transaction and then work backwards to trace its components. Insurers seek to repudiate before the claim is processed. In a very real sense, the first notification of claim is just the opening gambit. The fraud is not successful until the insurer has paid the claim.
Around two-and-a-half years ago, we decided to explore the use of AI in fraud detection to see if we could overcome these problems. Working in conjunction with Clyde & Co’s in-house data lab and experts at University College London, we set out to develop an algorithm that could predict the outcome of an insurer’s attempt to contest a potentially fraudulent claim. That prediction would be simple and binary: win or lose.
Traditional wisdom dictates that this is a hugely complicated task. Each case is usually judged on its own merits. There are so many unknowns at the stage of first notification of loss. Will the insurer be able to connect various parties via their use of social media? Will CCTV or dashcam footage be located? These are all unknowns at the point at which the claims team has to make its initial decisions, albeit those gaps may be filled in later.
With these concerns very much in our minds, we took six months to create the algorithm and test it using data from historical first notifications courtesy of an insurer with whom we piloted the project.
The biggest challenge we faced was getting the data into the correct format. Once we’d accomplished this, we fed historical data to the algorithm and compared its predictions with the real results of these cases. The correlation was little better than flipping a coin in the first instance. So, using more historical data, we trained the algorithm until we improved the successful prediction rate to just short of 90%.
This approach using AI could have major benefits for an insurer’s claims team. If, for example, you have thousands of potentially fraudulent claims to manage, there’s a limit to the resources you can bring to bear on the problem. Therefore, a win-lose indication for each case on day one would offer claims teams a clear steer as to where to deploy their