Visa bets on AI to boost credit cards, not replace them

For years, predictions about the future of commerce have often ended with the same outcome: credit cards would eventually disappear. They would be replaced by digital wallets, or cryptocurrencies, or buy-now-pay-later services. More recently, artificial intelligence has been cast as the next game-changer, with AI systems expected to take over everything from product discovery to purchase completion, as Emily Price writes.

New Initiatives

If an AI assistant can compare prices, find the best deals, and complete a purchase on your behalf, what role is left for the traditional payment network? Visa, however, believes the answer is simple: the network has a crucial role.

At the Visa Payments Forum 2026 this week, the company unveiled a suite of new initiatives in AI, tokenization, and stablecoins designed to ensure that Visa cards remain the preferred payment method as commerce transforms with AI.

The company is, in fact, betting that the AI-powered shopping boom could boost the role of credit cards, rather than diminish it.

AI's Limitations in Spending

AI helps people shop, but it doesn't 'spend' much money yet. The idea of AI agents managing purchases on behalf of consumers has generated significant excitement in the tech sector. But Visa points out a significant gap between consumers using AI for product discovery and actually letting it spend money.

Jack Forestell, Visa’s Chief Product and Strategy Officer, told Fast Company: “I was in one of our overseas offices a few months ago and asked the attendees, ‘Have you used AI for shopping?’ Everyone raised their hand. Then I asked, ‘Have you used AI to complete a purchase?’ Not a single hand went up. And this was a room full of payments experts.”

Shoppers' Lack of Trust in AI Managing Their Money

Forestell added: “We’ve been very focused on the concept of agent commerce, or AI-powered commerce, for over a year. We assumed that would eventually lead to agents helping not just with product discovery, but also with payments and beyond.” But the problem isn't that AI can’t help people shop; it’s that consumers aren’t yet convinced they should trust it with their money. The real barrier, then, is trust.

Technology Designed for the 'Smart' Era

So, the company isn't trying to replace AI transactions with its card system. Instead, it’s developing technology designed to make AI-powered purchases as secure and protected as traditional card transactions.

“I think consumers will want to feel that they have control over the process, and that they are protected,” Forestell said.

Consumers already know that if something goes wrong with a purchase, they can turn to their card issuer and payment network. Visa believes this protection becomes even more important when automated systems are making transactions.

Why is Visa Making Credit Cards Smarter?

A significant part of Visa’s recent announcement focuses on a technology consumers rarely think about: payment tokens. Today, many online transactions use tokens instead of card numbers. These digital credentials help protect payment information while allowing transactions to flow through existing card networks.

Visa is now adding more information to these tokens, including data about who initiated the transaction, its source, and whether an AI system was used. Forestell says this additional context is necessary to ensure the security of purchases made by AI systems. “The system doesn’t necessarily tell you, for example, if this was a purchase made with an AI system, or who that system is? Or what its level of trust or assurance is? Those are the kinds of variables we’re adding to the tokenization process.”

Improving Fraud Detection

The goal is to help banks better understand what’s happening behind the scenes of a transaction, which improves fraud detection and reduces false declines. For consumers, this could mean fewer legitimate purchases being misclassified while still maintaining strong fraud protection.

AI-powered shopping still needs a payment system. While many discussions about AI-powered commerce assume entirely new payment channels will emerge, Visa is betting on the opposite.

AI Agents Will Pay According to User's Financial Controls

The company announced a partnership with OpenAI that enables AI agents to initiate Visa payments within user-defined permissions and controls. It’s also launching an Agent Directory, a registry of approved merchants and AI agents, along with tools to help merchants determine if their websites are ready for AI-powered shopping.

This strategy makes sense, because AI agents still need a way to pay. Every transaction requires authorization, needs fraud controls, and may require dispute resolution if something goes wrong. These are areas where traditional payment networks already operate at scale.

Payment Reliability During Smart Shopping

The future may look familiar. Forestell believes some of the first AI-powered shopping experiences will involve routine purchases that consumers don’t enjoy making themselves.

Travel is another area where AI could gain traction. “We’ve had travel agents for over a hundred years for a good reason. Travel is an incredibly complex process that requires planning, exploration, and careful research,” Forestell says.

But in either case, the transaction still needs reliable payment data. That’s why Visa doesn’t seem worried that AI will make cards irrelevant. Quite the opposite: while AI may radically change how consumers shop, discover products, and make decisions, Visa is positioning itself so that the payment process continues to flow through the same infrastructure that consumers know and trust.

The future of shopping may be AI-powered. But if things go Visa’s way, the future of payments will still look a lot like credit cards.

* Fast Company

In the age of artificial intelligence, infrastructure security is changing from periodic management of updates to a continuous race against vulnerabilities that can be discovered and exploited at unprecedented speed. Cisco believes that organizations can no longer deal with networks and data centers with the logic of spaced maintenance or annual updates. In an exclusive interview with Al Sharq Al Awsat on the sidelines of the Cisco Live 2026 conference in Las Vegas, USA, Tom Gillis, Senior Vice President and General Manager of Cisco’s Infrastructure and Security Group, says that AI not only accelerates attacks but also imposes a new model for operating and securing infrastructure. He adds that the data the company monitors clearly shows this shift, explaining that the period between discovering vulnerabilities and exploiting them “has moved from months to minutes.” The problem is not only the speed of attackers, but advanced AI models, he says, now have a much higher ability to find vulnerabilities, putting organizations in front of more weaknesses, less time to deal with them, and infrastructure that is still managed at a slow pace in many cases.

End of the Annual Update Logic

For decades, organizations have dealt with infrastructure using a design, verify, then deploy approach, and then avoiding change for as long as possible. Gillis describes this model by saying that teams would build a design for a data center or network, verify it, deploy it, and then try to “not touch it for as long as possible.” This logic was acceptable in a world where a few vulnerabilities appeared during the year, and attackers needed months to exploit them.

But this equation is no longer valid. If an organization updates its infrastructure once a year, while AI tools can find and exploit vulnerabilities in a short period, the gap between the speed of attack and the speed of defense becomes dangerous. The problem is not the existence of vulnerabilities alone, but that the model for dealing with them is still based on old assumptions.

Gillis calls for transferring the operating philosophy known in cloud environments to traditional infrastructure. In the cloud, organizations typically do not rely on large, spaced-out changes, but on small, frequent modifications that are constantly tested and can be quickly rolled back if something goes wrong.

In Gillis's opinion, this model should not be confined to cloud applications; because networks, data centers, branches, and distributed environments need the same flexibility. He adds that updating once a year becomes a “big step” from one state to another, and often carries the risk of disruption. Small, sequential changes, on the other hand, make the organization more adaptable, allow for faster fixes, and reduce the fear of large updates.

Temporary Shield, Not a Substitute for Patching

Within this transformation, Cisco is introducing 'Live Protect' technology as a tool to bridge the gap between the moment a critical vulnerability is announced and the moment a full patch is applied. Gillis is careful to clarify the limits of this technology, saying that it is “not a patch” and does not eliminate the need for updates, but rather acts as a “temporary shield” or “compensating control” that prevents the exploitation of a known vulnerability until the final solution is applied.

The importance of this concept is that it addresses a practical problem facing most organizations; because some systems cannot be easily stopped. Disabling an e-commerce platform, a banking service, a telecommunications network, or a government infrastructure could cause significant losses or disruption. At the same time, leaving a critical vulnerability exposed until a maintenance window becomes available is no longer a safe option.

Gillis explains that 'Live Protect' can be deployed on components such as switches, routers, and firewalls without the need to restart them. This is a key point; because it shifts protection from a heavy procedure awaiting a maintenance window to a faster response, used as a bridge between the announcement of the vulnerability and the application of the full patch.

He believes this capability is “transformative” for organizations that cannot afford downtime but also cannot ignore vulnerabilities. In a world where vulnerabilities are exploited more quickly, the requirement becomes reducing the time between knowing the threat and containing it, even if the final patch has not yet been applied.

Initially, customers may prefer to deploy these shields manually to test them and build trust. But the direction Cisco is heading, he explains, is that when the company announces a vulnerability, it can offer the patch, and in appropriate cases, offer the temporary shield, making its deployment immediate when customers trust this mechanism.

From Infrastructure Attacks to Continuous Protection

Gillis points out that Cisco's thinking about these controls did not come out of nowhere, but from monitoring attacks that focused on the infrastructure itself, including campaigns like Volt Typhoon and Salt Typhoon. These attacks, he explains, targeted environments such as telecommunications providers, where updating the infrastructure is not easy or fast.

In such cases, the problem is not only knowing the vulnerability or having the patch, but the ability to deploy it without disrupting sensitive services. Therefore, 'compensating control' becomes essential. It does not solve the problem at its root, but it gives the organization a relatively safe time to act. As the time between discovering and exploiting a vulnerability shrinks, this bridge becomes an essential part of the new defense model.

Infrastructure Needs Cloud Flexibility

What is striking about Gillis's proposal is that he does not present security as a separate layer added to the infrastructure, but as a new way of operating it. The network or data center cannot remain a static system that the organization fears touching, while threats change at machine speed.

Gillis points to the use of a 'digital twin' as a software model that monitors infrastructure behavior and represents it using AI. When an organization wants to apply a change in settings or an update in operational components, this change can be tested within the model before being transferred to reality.

The more the technical teams trust that the change will work without disruption, the faster they can move. Gillis believes that the combination of small changes, pre-testing, and digital twins can make infrastructure “easier to manage, more flexible, and more dynamic,” fitting what he calls the post-AI world.

Smart Agents... A Two-Way Security Problem

Gillis’s discussion does not stop at traditional vulnerabilities; the emergence of AI agents adds a new layer of complexity. These systems do not just answer questions; they can perform steps, access tools and data, and interact with applications within the organization.

Gillis sees the problem as two-way. The first relates to protecting agents from the external world, i.e., preventing them from being manipulated, poisoned, or driven to execute malicious commands. He explains that agents are “complex and intelligent software pieces,” but they lack human judgment. Therefore, they must be protected at various stages: during construction, before launch through testing and trial attacks, and during operation by monitoring inputs and access to tools and resources.

Gillis likens smart agents to “teenagers” who have good intentions but may make “stupid mistakes” if appropriate controls are not put in place. This analogy clarifies the nature of the risk: the agent may not be malicious, but it may act unexpectedly, misunderstand the goal, or exceed the scope of the task.

The other, more difficult, direction relates to protecting the organization from the agents themselves. An organization may want a smart agent to access internal systems to accomplish a specific task, but it does not want to grant it open-ended privileges. Gillis gives an example of an agent that processes expense reports and needs access to the travel system, calendar, and perhaps images of receipts or credit card data to match expenses. This is a useful and specific task, but the organization does not want this access to translate into the ability to buy things unrelated to the task.

To illustrate the point, Gillis uses a simple example: an agent asked not to buy a Porsche might choose a Ferrari, and if prevented from doing so, might look for another option. The problem, he explains, is not the name of the forbidden item, but the intent to act. Therefore, fixed rules alone are not enough. The organization needs systems that understand context and intent, and can distinguish between an agent completing an expense report and one deviating from its original purpose.

Why Humans Are Easier Than Agents

In response to a question from Al Sharq Al Awsat about what is easier to control, Gillis considers controlling humans to be “much easier” than controlling AI agents. Humans have judgment, context, and an awareness of consequences. An employee might misuse a card or system, but they know there will be accountability and someone will pay the price. An agent, Gillis says, “doesn’t care” about these consequences in the same way.

The challenge within organizations is no longer just granting privileges to a human user, but managing software entities capable of moving within systems and taking multiple steps at high speed. Therefore, organizations need a new model of identity, governance, and monitoring that treats agents as actors within the digital environment, not as simple tools.

Gillis points out that monitoring agents will, in turn, require AI in terms of interaction between the agent and the application, or between the agent and another language model, which can be complex and fast. Therefore, organizations need systems capable of analyzing behavior, detecting deviations from the norm, and assessing intent, not just searching for specific words or commands.

Quantitative Readiness Before the Full Danger is Clear

Gillis addresses the issue of quantum computing and the risks of 'harvest now, decrypt later.' The idea here is that parties may collect encrypted data today that they cannot decrypt currently, but they are betting that future quantum computing advancements may make it possible to decrypt some encryption systems.

Gillis does not provide a definitive timeline for this shift, but says everyone recognizes the concern, yet “no one knows the timeframe” precisely. The risk may be near, or it may require an additional generation or two of technicians. But Cisco’s position, as he explains it, is that organizations should not wait for complete certainty before acting.

He adds that the company is working to introduce quantum-threat-resistant algorithms into the infrastructure, so that this protection becomes an integrated and lightweight part of the systems. He points out that Cisco aims to have these algorithms embedded in data center infrastructure by the end of the year, with gradual expansion later.

The importance of this file is that it relates to long-term data, especially government, financial, health, or industrial information that may remain sensitive for many years. If it is collected today in encrypted form that may become breakable in the future, the risk is not only for today, but for the lifespan of the data itself.

Security as an Operating Method, Not an Additional Layer

What Gillis presents in his conversation with Al Sharq Al Awsat is a broader vision of infrastructure security in the post-AI era. It is no longer just a matter of buying additional security tools, but of changing how the systems themselves operate. Attacks are faster, vulnerabilities are more numerous, smart agents are entering work environments, and quantum computing requires early preparation.

In this context, security becomes a continuous process, not a periodic project. Updates must be smaller and more frequent. Temporary protection must bridge the risk gap without causing disruption.