AI That ‘Thinks’ Like A Fraud Analyst

Fraud, chargebacks and false positives can impact revenue, and even cause firms to lose customers. The data is out there to aid in intelligent risk decisioning, but not all data is created equal, and analysis must happen within milliseconds.

To that end, digital fraud prevention company Kount announced in June that it launched the next generation of artificial intelligence (AI) in fraud prevention with a feature called Omniscore. The AI-driven solution helps digital businesses enhance their efforts to combat payments fraud and reduce chargebacks.

As has been reported, Kount’s AI and Omniscore use both supervised and unsupervised machine learning that leverages a universal data network spanning 12 years and 6,500 customers across 180 countries and territories. Combined with additional calculations, Kount’s fraud prevention solution closely emulates the decision-making process of an experienced fraud analyst.

In an interview with PYMNTS, Rich Stuppy, chief customer experience officer at Kount, said that in general, fraud prevention exists as “one of the few places where you have a technology, people, and a process coming together to solve a problem — and that problem is not a static problem. It is constantly changing, constantly evolving.”

That’s because businesses themselves are changing, and the digital customer journeys that those businesses promote are changing as well. At the same time, he said, fraudsters are quick to change, too, with attacks being waged in new and creative ways.

In laying out an effective roadmap in fighting fraud, Stuppy said, companies must have a fully formed strategy in place focused on how the customer journey unfolds, how those customers transact and how fraudsters could potentially extract value from or corrupt that journey.

In payments, he said, the key focal points for companies lie in whether they are getting chargebacks or whether friendly fraud is in evidence, and whether such occurrences are within an acceptable range. Yet the considerations, he said, can be different across different types of businesses.

Companies that traffic in physical goods and have a relatively high cost of goods sold will see lost goods (i.e. pilfered by fraudsters) and chargebacks hitting results, Stuppy said. As a result, they may not be focusing on what’s happening until chargebacks are so high that “you’ll be in trouble with the card schemes.”

He noted that other companies, where products and services are digital in nature, may seek to “accept as much business as you can while steering clear of penalties and excessive chargeback programs” put in place by card schemes.

Yet, Stuppy said, some individuals have a “misguided concept about machine learning — that all data and all data types are valuable.” He offered an example of a larger company that has 50 million addresses in a database. Machine learning algorithms don’t work with strings like addresses — they need numbers.

But there is value if companies use techniques to take all of those addresses — all 50 million addresses, in this hypothetical case — and compress them into a single value that retains historical information and what Stuppy termed metadata (millions of data points) tied to that address.

“Then it can become a single small valuable component in one of those models,” he said. Extending that concept to transactions, he said, “If you can figure out a way to strip all that information out and process it in a fraction of a second, across tens of thousands of decisions, across millions and billions of data points, you can get a really healthy understanding of the relative safety or the relative risk of a given transaction.”

The Balancing Act

In finding the right balance between fighting fraud and triggering false declines, which can translate into lost revenue and customer relationships, Stuppy maintains that machine learning and AI can be effective tools in analyzing transactions.

He pointed to the concept of supervised machine learning, where signals exist, comprised of millions of transactions, stretching back across historical timeframes, classified as “good or bad” events.

“In the payments fraud space, it’s challenging to get that signal,” said Stuppy, “because if you rely on chargebacks it can take days, weeks and even months to get that signal. And by that time, the horse is out of the barn.”

Part and parcel of Kount’s AI is the combination of supervised and unsupervised machine learning combined with additional calculations. The model detects and examines both anomalies across a massive amount of data as well as learns from historic data, ultimately delivering a score within 250 milliseconds, at volume and speed that no team of humans alone could handle.

Kount’s solution allows businesses to focus on their desired outcomes and set thresholds as to what transactions should be approved or declined. It offers up a “one-two punch. No matter what your business is doing, you are going to be protected.” Users can customize analysis and actions, determining, for example, to reject the ideal percentage of risky transactions whether that is 0.5 percent or 10 percent. Every business is different.

“There’s real money on the line,” Stuppy said of fighting fraud. “It’s a challenge and it involves a lot of moving parts. And if you get it right, it can generate a tremendous amount of value.”