With its main HQ in Munich, Allianz group is a multinational financial services company offering products and services to 100 million customers in more than 70 countries. With estimated combined annual turnover of €4bn, its 2000-strong Benelux (Belgium/Netherlands/Luxembourg) operation has a dedicated focus on the insurance market in retail, the SME space and mid-sized companies, primarily selling insurance through brokers.
As a result, says Dr. Jan Doumen, strategic team lead for Customer & Broker Information and Insights, Allianz Benelux needs to have an in-depth understanding of both that part of its value chain as much as its end-customers. To help, the company has set up a special cross-border ‘data office' of just under 40 people, of whom 14 are employed as data scientists, and of which he is also head of its ‘School of Expertise'.
The idea: help derive useful insight to aid the company's sales and marketing teams. And a key application of such help is to combat fraud, which is a problem both in underwriting and when a customer makes a false claim on their policy.
We have a history of trying to identify fraud in the traditional way, but fraud is getting more and more complex. Take a fraud ring; if you have a car accident involving two vehicles, you might look at the case and say, this looks normal, these are two independent people just happening to be driving on the road at the same time and got involved in a crash.
Best way to spot a fraud ring? Find the social connections
But what if you then spot that the driver of the second car lives in the same address as the first one, or that there is some other link connecting what at first appeared as two random individuals? Clearly, there is now doubt this is a legitimate claim. The problem for an organisation like Allianz Benelux is that historically, spotting that suspicious connection has proven extremely difficult. Not impossible, note; just difficult-Doumen and his people have been on the case here for some time. But what he says is making his life much easier is a different way of working with data, one based on graph software (the supplier in this case being Neo4j).
Looking at cases, in particular claims cases, from a connection perspective allows you to identify in an easy and a quick way these fraud rings. It's something that if you have to do that, on a relational database, would normally take a long time to figure that out, or you would have to have someone diving very deep into the case to understand it. But we've found that if you have a graph presentation of this kind of claim situation, you immediately see it; you see that the people are actually connected more than just by this incident.
The data team still needs, obviously, to carry out a full investigation to see if there really has been an attempt at wrong-doing here, but graph is delivering a list of all the cases human colleagues should be looking at, he says.
Beating fraud is ultimately about making sure the people from whom we collect premiums actually get the service that they're expecting, and at the right price. If you have people who are trying to cheat the system, you make premiums more expensive for other customers. Detecting fraud and removing it therefore gives you the possibility of lowering your charges for honest customers.
However, that's not the only application of graph in the company, it turns out. Equally important for Doumen and his team, is what he calls the possibility to have a truly 360 degree view of your customer, or achieving full ‘customer-centricity'. The reason is our old business IT friend: silos.
When you look at PNC (primary and noncontributory) business and life insurance, they're frequently very disconnected. You have customers that have car insurance and legal protection insurance but the information is in separate systems. One part of our organisation is trying to sell life insurance but don't realise that the customer you're trying to sell to already has a product with us. So how can I use that to help this customer? Can I offer a more appropriate solution?
‘€2 million worth of operational profit value'
For Allianz Benelux, that can only really become possible if the broker has a complete view of the customer and their contracts, but also her overall life situation such as where they live, family situation and so on. Thus customer-centricity, finding better products better tailored for the market-and here again graph as a way to easily visualise social connections has proven useful.
How did we end up with a relational database in the first place? It was mostly because of a lack of enough computing power and a lack of a way to represent data the way that it's presented in your head. Everything you do with information is how one bit is connected with the other pieces, so having graph databases allows you to do it in a more naturalistic way, and now computers are sufficiently powerful to support that. We strongly believe in storing business data in a graph way; I'd even go so far as to say that within five years most of our information will get stored in a graph format, even our data warehouse.
There's another reason Allianz Benelux is so convinced it needs to use graph more going forward; its application to the fraud and customer-centricity issue has already, it has said on the record, contributed to the identification of €2 million worth of operational profit value. Doumen says that is probably an under-estimate.
For fraud, once we were able to identify possible frauds, we were able to hand them over to the fraud team and we were able to recover so much money. That's one way that we get to these kinds of numbers. And we've also built an application that allows you to search the customer database in a very user-friendly way; you can click on a customer, you can see her claims, you can see her connections, and so on. That's now the basis of a front office application so when you call the contact centre we can very quickly figure out who's calling and their status with us in one view.
Next steps for graph at this part of the Allianz Group will include exploration, Doumen concludes, of advanced analytics and Machine Learning.