Visa Says AI Can Shop but Trust Still Has to Close the Sale

ChatGPT’s consumer debut may have been only three years ago, but Visa’s leadership said the company is hardly new to artificial intelligence (AI) or turning it into a commercial advantage.

In a conversation with PYMNTS CEO Karen Webster, Andrew Torre, president of Visa’s value-added services division, said Visa has used AI for decades to detect fraud and score risk in milliseconds. Those capabilities now look like prerequisites as commerce shifts toward software agents that can shop and pay on a consumer’s behalf.

For Torre, agentic commerce will not be driven only by better models. It will hinge on whether consumers, merchants, platforms and financial institutions can embed payments, identity, authentication, consent and controls into automated shopping journeys without sacrificing speed or trust. He described the moment as a convergence of fast-changing consumer expectations, accelerating innovation and tightening guardrails around privacy and regulation.

Torre said his household shows how fast AI is becoming a mainstream commerce interface. He now uses generative AI tools as his default search method and sees “three generations” of his family using them. One prompt came from his daughter, who delivered a highly specific wish list: a hard-to-find pair of Nike sneakers in precise green and blue shades to match her university’s colors.

Teachable Moment

The sneakers illustrate the promise of an assistant that can translate intent into a purchase across a fragmented retail universe, and the risk of an agent buying the wrong item, at the wrong price, from the wrong seller. That risk raises a payments question, Torre said: what has to sit behind the scenes so the “search-to-buy” experience is safe, controlled and reversible?

His answer centered on “ancillary services” around authorization, including tokenization, consent capture, identity and risk tools, and the infrastructure to deliver them through modern platforms.

“She goes to a school, a university that has a certain green and a blue shade, and she really wants this cool pair of Nikes, hard to find,” Torre said. With agentic commerce, the agent becomes the shopper that can search for this hyper-specific product with a few simple preferences and parameters, such as brand, size, color or style, and a maximum price, and quickly narrow down the best option. It is made possible in part by the ability to use data capture consent and connecting those experiences to clients through cloud-delivered services, “with either [cloud-based] Pismo services or [processing platform] CyberSource and our acceptance platform.”

That focus on “how” is central to Visa’s pitch to banks, processors and merchants moving from click-and-checkout commerce to agent-driven commerce. Decisions can be made in real time, and prompts can contain “intent clues,” as Webster put it, that reshape the shopping journey dynamically. Torre said clients may be more prepared than they think because they already consume Visa decisioning signals such as risk scores in milliseconds.

“We’ve also had these tools available to them for the last 30 years,” Torre said. “We’ve been using AI for fraud, risk scoring innovation and more. We’ve also been giving them more complicated products as well.”

Flexible Credentials

Torre pointed to Visa’s Flexible Credential and the Pismo platform, while acknowledging a persistent friction: many institutions still run on legacy stacks that make modern deployments harder than they should be. Pismo, a cloud-native issuer processing and core banking platform Visa acquired, is positioned to reduce that complexity and speed the rollout of new credential and risk capabilities.

“One of the constraints we’ve seen is a lot of customers using their legacy, whether it’s in-house or another issuer processing stack. It’s complicated to deploy,” Torre said. Pismo has served as an unlock for clients to access Visa’s services. Torre added, “We’ve been winning deals in Pismo to simply launch some of our new solutions.”

Agentic commerce also changes where fraud can enter the system, he said. “Token provisioning became a vector for fraud,” Torre said. “And now we have token provisioning intelligence where we’re giving them a risk score.”

What’s ‘AI Ready?’

“AI ready” is an operating model, Torre said, as Visa deploys AI internally to write code and simplify product experiences, while embedding it into client-facing risk tools.

“One of our risk solutions, Visa Risk Manager, we have a managed services solution on that where we had people helping write and look at rules,” Torre said. “Now we’re using AI to constantly update and say, ‘Have we seen things that are different? Do I need to rewrite your rules?’ We used to have people do that. Now we can have AI.”

“Every task I have, I start with, ‘Can I solve it with AI?’” Torre said.

Constraints remain. Early agent workflows were “kind of clunky,” he said, though improving quickly. Token rules vary across markets, and in some countries token vaults are domestic, complicating portability.

Visa’s response is what it calls Visa Intelligent Commerce, designed to standardize consent, support disputes and maintain traceability when software is making decisions. Its Trusted Protocol ensures “we can distinguish between agents and bots,” addressing whether an agent acted appropriately and whether it was legitimate, Torre said.

“We don’t think of any innovation without bringing in risk, fraud, security, and then dispute mechanisms,” Torre said, adding that this approach is built into Visa Intelligent Commerce and reflected in open specifications that partners can inspect.

Back to the Shoes

In an agent-driven world, the trust test will often be mundane: the order arrives, but it is wrong. The consumer needs a clear path to challenge the purchase.

“Let’s just say my daughter’s sneakers, they come, and they’re not the right sneakers,” Torre said. “They’re giant, in the wrong colors, and she’s upset … But the fact that I can say, ‘Wait, I didn’t authorize these,’ and they get returned.”

Torre said those proof points depend on capturing consent context so it can be checked later. When Webster asked why Visa is “built for this moment,” Torre pointed to the infrastructure and scale Visa already operates.

“Visa as a whole is ready for this moment,” Torre said, citing “the 4.8 billion credentials” on the network, “the 150-plus-million sellers out there,” “network grade security” and real-time delivery across “over 200 countries and markets.”

Many of the mechanisms that made digital commerce viable, from identity to dispute resolution and consistent rules, are being ported into a conversational interface where the buyer might be software. The winners, Torre said, will let automation move fast while keeping consumers in control—and keeping trust as the product.