A big friction in using artificial intelligence to shop is payment: You’re still managing a card.

For agentic commerce to become a frictionless consumer experience – the hankering for a hamburger to its delivery with minimal human action – the payment still holds technical barriers companies are working to overcome.

“What the real goal of agentic is, is to have agents that will act and execute on your behalf,” says John Kain, the head of worldwide financial services market development for Amazon Web Services, the tech company’s cloud computing business.

To further this goal, AWS and Visa, the largest U.S. card network, announced in December that they’re cooperating to develop tools to build “network agnostic agentic workflows” for the payments aspect of AI commerce.

“The goal is straightforward: provide a secure, scalable foundation for building the next generation of intelligent commerce solutions,” the companies said in a blog post.

AWS is Amazon’s fastest-growing and most profitable business segment. The cloud unit’s operating income rose 14% last year, to $45.6 billion, with sales rising 20% to $128.7 billion, Amazon said last month in its annual report. Kain, a former JPMorgan Chase executive who’s based in New York, discussed AWS’ agentic commerce work in an interview Thursday.

Editor’s note: This interview has been edited for clarity and brevity.

PAYMENTS DIVE: Where does AWS focus its efforts in agentic commerce? 

JOHN KAIN: Our real focus is very much on the payments execution layer. Commerce is certainly interesting but fundamentally the most important layer of that is being able to do secure payments at scale in a way that enables agentic commerce as a whole. And so we were very focused on making it easier for our customers to build applications that could actually integrate agents into the existing payments rails.

What’s your specific offering in the agentic commerce space? Software? Consulting services? 

It’s the platform and the partners. From an agent platform perspective, what we’re giving our customers is the ability to use a variety of models from different model providers, whether that’s OpenAI, Anthropic, our own models, plus a wide variety of other model providers that are out there, so customers have choice in being able to pick and tailor models for their specific applications. We do that in an environment that’s secure, so both the inputs and outputs and any training you do for the model is restricted to your own environment. And now, with the emergence of agentic commerce, or agentic frameworks as a whole, it’s being able to build agents leveraging some of that generative AI technology, but also putting in the control and governance frameworks that allow organizations to build agentic frameworks at scale. 

Much of the agentic discussion today focuses on consumer retail. Longer term, what are the dominant agentic use cases?

I think what’s actually going to happen is that most of the services that we use today are just going to improve dramatically from a user-experience perspective. If you see where generative AI is being applied, certainly customer support is one of the biggest areas, the ability to communicate with your customers, divine intent, and then route them to the right agent. And in this case, most likely an agentic agent.

What are AWS customers deploying to battle fraud?

Traditional machine learning has been used heavily in the fraud space. Mastercard and Visa both have traditional machine learning solutions on top of AWS that will do real-time fraud detection in just a few milliseconds. So being able to do fraud detection in a more real-time manner, particularly as the payments infrastructure as a whole moves to be more real time in nature, we think that more and more transaction volume will move to a real-time payments rail. But because of the finality of settlement, that means that fraud is becoming a more important component of both the decision on which rail to use, and then, if you’re using real-time rails, how to prevent and detect that. That is almost exclusively a data and AI challenge.

Is fraud risk hindering the adoption of real-time payments?

No, I wouldn’t say that. I would say, though, it’s an area where firms were investing a lot of research and energy even before they started adopting real-time rails.

Is the AI arms race between criminals and corporations evenly matched if both sides have the same tools?

Firms that have large datasets and large experience with their customers at scale fundamentally have an advantage. At the end of the day, what makes AI effective is the quality of the data you have that is actually going into it. And so if you have really good insights about your customer and what their typical behaviors are, and a history of years with them, the odds of you being able to detect fraud are far better than someone who doesn’t have that data, right? That history of relationship actually becomes really meaningful from a fraud perspective. Firms that have a long set of experience in the area from a data perspective are fundamentally better positioned than even new entrants or fraudsters.