Visa is experimenting with agentic artificial intelligence that would add automated execution for consumers to queries on large-language models.
Executives wouldn’t put a timetable on when these products will go live, but they are being tested in internal sandboxes.
Jack Forestell, San Francisco-based chief product and strategy officer, explained how the first iterations of agentic AI will work.
“We formed a point of view that AI and specifically agent-driven AI has the potential to radically transform the way we discover, the way we shop and the way we buy,” he said.
“We think it has the potential to be as transformational as the introduction of e-commerce itself.”
This should lead to more commerce, and hence more payments.
Payments + LLMs
Visa has announced its Intelligent Commerce product, which in Asia is being tested with partners including Ant International, Grab and Tencent, as well as with LLM companies in the US such as OpenAI, Perplexity and Anthropic.
Forestell told DigFin the insertion of AI companies into the payments ecosystem should not impact merchant discount rates (the amount merchants pay to use the payment networks of processors such as Visa), nor the way those fees are divvied up among banks and payments companies.
For now, at least, the makers of LLMs are interested to drive usage rather than seize a slice of MDR. These companies have free versions of their LLMs but are pushing to get individuals and companies to buy subscriptions to more compute and search.
And that’s where Visa expects a new wave of online commerce to start: with someone asking an LLM about a product or an experience, and then being able to easily make the purchase within the query without having to separately track down other websites. Putting together a holiday, for example, won’t require sifting through sites to book flights, hotels, and concerts or restaurants.
This kind of consumption is more effective if the LLM delivers responses that fit the user’s preferences. This means if the LLM has access to data about its users. Some of this would come from the act of querying. But there’s also an open-banking element, in which users could opt to show their card-purchasing history.
The new flow
This immediately raises questions about privacy, safety, reliability, accuracy – for someone’s data, and for their money. Plus the ability of AI agents to do all this properly.
Forestell’s teams have thought through all of these things. This doesn’t mean the service is flawless – it’s experimental – but it’s ready to go live.
Stephen Karpin, Singapore-based president of Visa’s Asia-Pacific business, says this is the latest iteration of a move by the company from organizing itself around products, toward services. “This involves a combination of our infrastructure, capabilities, and APIs to create a ‘Visa as a Service’ stack,” he said, noting this is only possible because the company has spent decades building the foundations of its payments-processing network.
“We are shifting to a services architecture that includes risk management, settlement, credential directors, and now tokenization,” he said. “This is now all scaling.”
Visa reckons there will be massive demand for agentic AI. Forestell noted OpenAI and the like have quickly amassed more than 1 billion active users. Traffic volume from LLM sites to retail and marketplace websites is rising fast.
What’s missing is the ability to pay directly off the back of a query. LLMs don’t have access to credit-card numbers or passwords. And there’s no trust factor: consumers fear agents would misrepresent them or steal their money; banks can’t tell if an agent is legitimate; merchants aren’t confident they’ll be paid.
Forestell says Intelligent Commerce has been engineered for people to buy with AI and be confident the payment is as safe and secure as it is with a credit card.
It’s really about tokenization
This involves three components: upgrading cards to handle agents; enabling agent transactions; and personalization to make LLM queries more relevant.
This AI-prep work is focused on tokenization, which payments companies have applied to other digital payments for a decade or so. This involves replacing the 16-digit number on a plastic card with a unique, cryptographically protected code, like a one-time pass that is linked to an account but if hacked doesn’t provide access to funds or instructions.
This abstraction layer is being refitted for agentic AI, meaning these tokens can interact with agents. This capability is augmented by services that sync payment instructions, and signals that ensure all authorized parties in a transaction have the same access to data for processing, detecting fraud, and handling disputes.
In practice, this means a cardholder would be asked by an LLM to authenticate their agent with their bank. Given Visa has about 14,500 financial institutions in its global network, it’s likely that most users in developed markets or among the affluent will be able to do so.
This authentication involves Visa generating a token specific to a user’s AI agent, and then binding that token to Visa’s agent platform. This means only the user or its AI agent can control use of that token. There is also the possibility of programmability, which at its most basic level means letting the user lock or unlock their agent.
A person wishing to make a purchase with an agent for the first time would load a bank-issued card to generate a token that enables the agent. This can be as simple as tapping a plastic card against a phone. Behind the scenes, Visa obtains a credential from the user’s bank, and passes this back to the token, which is now ready to get to work.
On the cusp
Visa also wants users to opt in to make their purchase histories available to agents. For users this means a more personalized service and better LLM outcomes. For LLMs and Visa, of course, this opens a treasure trove of data upon which to train models and add value to their network partners. If a customer is given more accurate recommendations, that is more likely to mean a sale, and a payment request.
Forestell emphasized that Visa is not sharing user transactional or personal data. It does share metadata about that person, built by Visa’s modeling, which evolves with search and spending decisions. The metadata is shared with LLMs, to help them train their own AI.
Finally comes the actual payment. The agent detects the user’s buy signal, pulls together the order book, asks for a confirmation, and if the user agrees, is now empowered with access to the card account to make purchases.
The engagement triggers a variety of API instructions and records that the agent uses to reconcile the orders and give payment credentials to the various merchants, and then reports back to the user with confirmed purchases and receipts.
In some respects, agentic AI is part of an evolution of digital payments, from the internet to desktop e-commerce to mobile to AI. But Karpin says this shift is of a greater magnitude. “We’re on the cusp,” he said.
