Visa CyberSource: AI’s Role Is To ‘Predict’ — Not To ‘Know’

Marvin Minsky, cognitive scientist and co-founder of MIT’s AI lab, said decades ago that computers can only do what we tell them to do.  Scott Boding, vice president of risk solutions product management for Visa’s CyberSource, tells Karen Webster why AI’s potential may be misunderstood by some observers (there’s not likely to be a robot uprising), but how the key tenet holds true in AI-driven payment models: to predict, not to know.

According to Marvin Minsky, cognitive scientist and co-founder of MIT’s artificial intelligence (AI) lab in 1958 (and eventual winner of computer science’s Turing award in the early 1960s for a paper titled “Steps Toward Artificial Intelligence”), “A visitor to our planet might be puzzled about the role of computers in our technology. On the one hand, he would read and hear all about wonderful ‘mechanical brains’ baffling their creators with prodigious intellectual performance. And he (or it) would be warned that these machines must be restrained, lest they overwhelm us by might, persuasion or even by the revelation of truths too terrible to be borne.”

Minksy, however, wrote in that same paper a truism that has echoed across the decades: “A computer can do, in a sense, only what it is told to do.”

Artificial intelligence, at its root, is a way to make search more efficient. In commerce, the searching can span any number of endeavors, like separating good transactions from bad ones, or inferring which customers are trustworthy.

In the latest conversation with Visa on the state of AI, Scott Boding, vice president of risk solutions product management for Visa’s CyberSource payments management platform team, told Karen Webster that, in the decades since Minksy’s paper, “we’ve come a long way” in leveraging technology to create underpinned AI systems that marry complex math with Moore’s Law.

Boding stated that there’s a bit of a misconception in everyday life that the machines can become omniscient or omnipotent — because, at the basic level, while the aim of AI is “to inform better decision-making, the misconception is that prediction means vision or knowledge. In fact, there is just probability. The role of AI is to predict and not to know. It cannot know, because AI is about trying to understand an unknown event, where you try and know as much as possible about an event, and then try to predict the outcome.”

The Challenge In Payments

Knowing and predicting are not easy tasks in payments, a world where there are many unknowns, where transactions are increasingly done across far-flung locales and where parties never meet face to face.

Concerning Visa’s efforts in AI as they apply to commerce, Boding said, “We are trying to use AI as a way to seize and automate a lot of the heavy lifting in fraud detection — a time-consuming task that many merchants are doing manually today. We are trying to use AI to increase efficiencies for our merchant clients. … It is augmentative to human processes.”

Of course, there is no shortage of historical data with which to work and classify. As Boding told Webster, “I like to think about the amount of data required in three different dimensions.”

  1. The volume of data: In eCommerce, for example, the more transactions that can be seen, the better job of predicting that can be done.
  2. The width of data for each transition: This spans additional information, such as email addresses and device information.
  3. Labels: These include outcome information about transactions, such as chargebacks or credit backs.

Labels, when created, offer additional insight, defining whether transactions are good or bad, and building context. Often, though, labeling happens later, and is retrospective in nature.

“We do use past behavior … that we’ve seen over the course of time, … [and] look at trends and patterns that have emerged. And it becomes important to ‘weight’ those patterns differently,” Boding said.

Along the way, human intervention becomes especially critical. “AI can help identify trends, and put them in front of a person to say ‘we need some guidance here,’” he added.

Getting More Granular

That’s high-level detail, but, drilling down a bit, “there’s obviously a big difference between looking at billions of transactions and then being able to make a good decision about an individual transaction. At the end of the day, it boils down to ensuring good customers can make a purchase safely,’” Boding said.

“It becomes paramount to be able to identify trends and patterns of users,” he added, noting that there is a “whole host of different fraud threat vectors that are taking place, and so it becomes critical for merchants to be able to identify groups of people and their shopping patterns. … From an AI point of view, we have to build a system that’s really adept at identifying different types of users and their typical commerce patterns.”

Against that backdrop, Boding noted, AI systems can help determine whether an individual’s activity fits one of the aforementioned behavioral patterns or types. “If we’re seeing a deviation from one of these typical patterns, the system takes an even closer look and submits it for more evidence, and [we] have a better sense of whether this is a legitimate user who has just changed behavior or whether it is a fraudster trying to commit bad acts,” he said.

Visa’s own system is designed to produce a score, and provide information codes that can be used to make decisions for different scenarios and individuals.

In deploying AI to the realm of commerce, said Boding, “seasonality is something that can be very tricky to deal with, and requires being able to profile the merchants as well. The near-term roadmap is one where different AI-driven subsystems — which may operate across verticals or transaction types — can be incorporated into a unified, larger AI system. We also need to have profiles for the merchants, and understand their specific categories in order to understand their unique customers. We work with our airline customers, for example, very differently [from] how we address luxury retail and digital goods fraud management practices.”

He added, “Being able to apply models to detect fraud has allowed merchants to respond to every single transaction at the moment of checkout, boosting confidence and making sure they’re letting good customers make purchases, while protecting themselves against fraud.”

As he told Webster, AI allows “merchants to scale, because eCommerce volumes continue to grow dramatically.”

“This is almost like a Sci-Fi age, when science fiction writers were predicting ‘here are all these amazing things we’re going to have in the future,’” he said of AI. “Now, the hardware and software is starting to enable what the mathematicians had predicted.”



The PYMNTS Cross-Border Merchant Friction Index analyzes the key friction points experienced by consumers browsing, shopping and paying for purchases on international eCommerce sites. PYMNTS examined the checkout processes of 266 B2B and B2C eCommerce sites across 12 industries and operating from locations across Europe and the United States to provide a comprehensive overview of their checkout offerings.