Keeping Payments at Mastercard on AI’s Leading Edge — without Falling Off
In an age of technology that can "reason," Mastercard's AI head Greg Ulrich must balance skepticism and an open mind. Here's how the payments network fits rapid AI development with careful deployment.
By Steve Cocheo, Senior Executive Editor at The Financial Brand
If artificial intelligence’s rapid pace of change periodically leaves you breathless, don’t feel too bad — you’re in good company.
Mastercard’s Greg Ulrich spends his days steeped in AI and data analytics development, but even he finds the pace of change "staggering." The staff at the payments company’s AI and data innovation hub must keep tabs on goals and applications of current technology to meet the needs of issuers and merchants, while also monitoring how the technological environment rapidly evolves. Nothing holds still.
"Eighteen months ago, people were saying that the largest large language models were going to be able to do everything you needed in any ecosystem," says Ulrich, chief AI and data officer. "Then, in our area, people said you needed a vertically focused LLM because there was no way that even a large LLM would be smart enough to understand the financial ecosystem."
So financial companies on the leading edge of this technology developed such focused varieties — only to see their efforts eclipsed by newer developments. And then the discussion switched back, with people arguing that the costs of the latest broad models were too high and that the financial services business needed narrow, vertically focused industry specific models, again.
"The discussion keeps going back and forth on these themes," says Ulrich, "not because people don’t understand them, but because the technology continues to evolve and we’re still learning what’s working and what’s not."
When ‘Reasoning AI’ Meets Large Language Models
Ulrich notes that six months ago "reasoning" GenAI models began to come out, the idea being that they would take time to "think" more deeply about queries and tasks put to them. While these reasoning models were once separate from mainstream GenAI, now providers are starting to merge the technology with the latest iterations of their products: Ulrich points out that Anthropic, producer of Claude, recently announced that going forward it will be combining the two AI variants in the same ecosystems. OpenAI, he adds, recently introduced GPT-4.5 and said that future releases of GPT would incorporate that firm’s reasoning model.
"We keep evolving how these things work, not just their actions, but also their error rates," says Ulrich.
Now, here’s the conundrum for expert and layman alike. One of Ulrich’s favorite AI studies compared LLMs and reasoning models, specifically regarding two factors: how often they said "I don’t know" and how often they gave bad information.
The LLMs studied rarely said they didn’t know, according to Ulrich, but "they got a bunch of things wrong."
On the other hand, reasoning models gave correct answers more frequently — but they said they couldn’t answer the query more often too.
For reasons like this, Ulrich says as Mastercard adopts and adapts the latest AIs, it strives to keep a "human in the loop." While the company has many automated tools for evaluating AI performance, he says that a constant has been human testing and monitoring.
Mastercard and its users are dealing in money — and Ulrich admits to some apprehension that keeps him focused on new developments.
"The idea of AI being falsely positive or confidently wrong is the most dangerous risk, as I look at these developments," says Ulrich. "An LLM or other model is the same as a person. The most dangerous thing is when you have someone who believes they’re right when they are dead wrong."
Read more:
Mastering the AI Possibilities … Carefully
Mastercard has been using AI for years, especially in the anti-fraud arena. Ulrich has been in his current post at the company for nearly a year. Previously he has held other jobs at Mastercard, most recently EVP for strategy, corporate development and M&A.
Ulrich believes getting the state of the art of AI to the point of confidence and trust is critical as the financial services industry moves towards agentic AI. He doesn’t foresee an "aha" moment, but more of a gradual adoption as people see what can be trusted.
As a parallel example, he points to Waymo self-driving car service.
"It’s a very niche thing," says Ulrich. "It evolved and evolved and evolved. Now, in San Francisco, it has the same market share as Lyft, and it continues to grow there." Note that Waymo, owned by Google’s parent company, Alphabet, began developing self-driving technology more than a decade ago in 2009.
Wide scale adoption of agentic AI will be very experimental and iterative, Ulrich believes.
"When the model takes action on your behalf, the bar becomes higher," says Ulrich. "People are going to want to be appropriately cautious, particularly enterprises where trust is a major component of their brands and reputations."
Much of what’s been done with GenAI in financial services involves improving productivity and efficiency internally — helping customer service teams, legal staff and engineers, says Ulrich.
Those are all more controllable and containable than AI that directly deals with customers, he explains.
"You’re going to see more trepidation as we go out there with agentic, because these things really have to be foolproof," says Ulrich. "You don’t want agentic AI to book your trip to Orlando and find out that it ended up booking too expensive a flight, or that the dates are wrong. Customers will have a very low tolerance for that. They should have a very low tolerance for that."
"Hallucinations" remain one of the risks of AI and the degree of that risk will vary according to the nature of use, says Ulrich. He adds that the perception of the rapid adoption of new forms of AI across the board needs tempering. He says a good deal of what the consumer reads or hears about dwells on the "right-hand side of the bell curve." The center of the curve — where most enterprises sit right now — by definition, is much more conservative, still trying things out.
Read more: Can GenAI Restore the ‘Humanity’ in Banking that Digital Has Removed?
Where the Consumer Feels Mastercard’s AI Investments
Mastercard rarely interacts directly with consumers using its payments services, at least not in a visible way. As a payments network operating as a B2B2C — or a B2B2B2C, depending on how you count merchants and issuers — its activities are below the surface.
"Every time a consumer taps, dips or swipes a card, or pays online, our ecosystem is working to determine if the transaction is legitimate or not," says Ulrich. "AI has been a component of that for years."
In milliseconds, Mastercard AI scores a transaction by consumer, location and many other factors to give issuers guidance used for accepting or denying the transaction.
Ulrich explains that what GenAI has brought to this function is the ability — in those same milliseconds, mind you — to evaluate each consumer’s transaction in the context of the entire Mastercard network of merchants, issuers and others. Ultimately, the decision to process a payment lays with the issuer.
"We’re never going to get to perfection in this world, but everything we do to make that experience safer for the consumers, and more seamless for them too, is important," says Ulrich. For Mastercard, this is a competitive tool , as much as a loss-control function. Ulrich explains that competitive forms of payment that have sprung up around the world often don’t have the anti-fraud features that established payment networks like Mastercard can provide.
The other end of the process — dispute resolution — is also being streamlined using GenAI, according to Ulrich. He says this is reducing some dispute handling by weeks.
Another focus for Mastercard’s AI development is customer onboarding, assisted by AI tools. In addition, Mastercard has been developing forms of GenAI aimed at improving search functionality on ecommerce sites.
This is called "Shopping Muse." Ulrich says the AI is a chatbot that provides shopping recommendations much like an attentive salesperson in a retail store. Right now, the bot is a standalone function, but he says ongoing development will help to integrate it more natively into sites as well as into a suite of personalization solutions offered by the company.
At the outset Ulrich’s function was referred to as a "hub," and the company’s business units are the spokes. Innovation ideas for AI applications flow in both directions. But "running around and breaking things" isn’t an option for a major payment company.
"Putting in the appropriate [control] processes while you keep your foot on the accelerator of innovation is the real challenge," says Ulrich. "I’m a strong believer in this technology’s capabilities."
