Tackling bust-out fraud: why people need data, and vice versa!

Data scientists and engineers tend to have an unwavering belief in the power of data and AI to improve the world we live in. I’m fairly typical in that regard – with one important caveat.

Large amounts of data can be analysed to predict patterns. But only people can determine whether those patterns are relevant, and if anomalies highlighted are significant. The best predictive models are those that manage to integrate all the nuances of past human experience in their data analysis.

Let’s look at a specific example, taken from a previous article I wrote on buyer fraud, or bust-out fraud as it’s commonly known in our industry, and Euler Hermes’ sophisticated machine learning tool.

Bust-out fraud occurs when a buyer places an order for a product, takes delivery, then “disappears” before the seller can collect payment. In many cases, the buyer company never existed in the first place. Sellers are left holding an unpaid invoice, having shipped goods often worth thousands of euros.

Detecting bust-out fraud is part of our day-to-day business at Euler Hermes: we guarantee the payment of our clients’ buyers’ invoices. Part of this involves identifying a suspicious buyer before a client ships a product to them.

We developed an internal, AI-driven product to track large amounts of incoming data relating to buyers – for example credit limit requests, financial data, rejected claims. The product flags up buyers that fall outside the norm: for example, multiple requests by clients to cover the same buyer, or financial statements that look too good to be true.

These buyers are identified by the tool as potentially suspicious. The key is in the word “potential”. Because the tool cannot itself predict buyer fraud with 100% accuracy: for that, we need people with their years of collective experience and expertise.

In fact our buyer fraud product does not make decisions: it is a decision-making aid that saves our people time.

The tool provides a useful way to narrow down the list of potentially fraudulent buyers from thousands to a few ten. A Euler Hermes credit analyst or risk underwriter will then investigate each buyer identified by the tool, first by manually examining the issue raised, then by interviewing the buyer if doubt persists. They make the final decision.

In fact the decision product has a high success rate in detecting suspicious activity – but among the real suspicious cases detected linger a few false alerts. This is because the tool is designed to spot patterns, whereas people are talented at making sense of anomalies and the complex reasons behind them.

The product relies on people in another important sense: the assumptions that we initially fed into the tool, and the improvements we have made since launch are all based on the experience of people in the field.

This point is central to how we develop internal products such as this at Euler Hermes. A large part of my job is talking to our people to understand how they approach decisions and the issues they face.

My goal, and the goal of everyone in our Data Analytics and Artificial Intelligence team, is to take all this valuable experience and distil it into something a machine can understand, and use for people’s benefit. I’m not interested in technology for technology’s sake: it should always exist to help people solve real-life problems.