AI agents are starting to shop: what Visa, Mastercard, and Amex are building, and what a brand must do to be chosen
American Express and Mastercard announced tools for payments made by AI agents on the same day; Visa already measures whether a site can be understood and used by an agent. When software does the buying, the brand is chosen based on what the software reads, before the person sees anything.
In brief
- American Express and Mastercard announced new tools for payments made by AI agents on the same day. Visa had opened its network to agents in April 2025, and in March 2026 Santander completed the first payment in Europe carried out by an agent from start to finish.
- The three networks are working on the same three points: the agent has to be registered and verified, the purchase intent has to be recorded, and the card is replaced by a token. Amex promises to cover errors made by registered agents; Mastercard is testing a score that tells the bank whether an agent made the payment.
- For a brand, the news comes before the payment: Visa has a score that measures whether a site can be understood and used by an agent, and Amex wants its benefits “discoverable on AI platforms.” An agent chooses by reading, and it chooses before the person sees anything.
To buy on someone’s behalf, an AI agent has to do three things: understand what the person wants, find who sells it, and pay. The last one, paying, is what the companies that handle the world’s cards are working on. American Express introduced two AI agents for its business customers, and Mastercard announced new services the same day to recognize payments initiated by an agent. Behind the two announcements are a year and a half of work by the three networks, and a consequence for anyone who sells: when software does the buying, the brand is chosen based on what the software reads.
Index
Two announcements on the same day, a year and a half in the making
For now, American Express’s agents are office tools. The first texts the employee via SMS or Slack when a corporate card expense needs details, suggests a category and business purpose, and submits the expense for approval. The second answers the finance team’s questions about company spending and produces charts in seconds. Amex is announcing them for the fall and says the features are in development and timelines may shift. The same day, Mastercard added to Agent Pay, its agentic payments program, services that recognize agent-initiated payments.
The matching dates cap a sequence that began in the spring of 2025. On April 29, 2025, Mastercard unveiled Agent Pay, with tokens built for agents, and with Microsoft and IBM as its first partners. The next day Visa opened its network, 4.8 billion cards and 150 million merchants, to developers building agents, with OpenAI, Anthropic, Perplexity, Microsoft, and Mistral among its partners. On April 14, 2026, American Express released a kit for agent payments with its cards. Three networks, three different paths, one idea: agentic commerce, meaning commerce in which software buys on a person’s behalf, needs a checkout that trusts the software.
How an agent pays
Read side by side, the three networks’ documents describe the same mechanism in three parts. The first is the agent’s identity: for Mastercard, agents must be registered and verified before they can pay, Amex has a registration service, and Visa keeps a directory of agents and merchants it has verified. The second is intent: Amex records what the customer asked to buy, to use it in authorizations and disputes; Mastercard says consumers have complete control over what the agent is allowed to purchase. The third is the token: the agent pays with a substitute credential, the way phones do today with contactless payments, and the card number stays out of the conversation with the chatbot.
In Europe the mechanism has been tested live. On March 2, 2026, Banco Santander and Mastercard completed Europe’s first payment by an AI agent, executed end to end on the bank’s live infrastructure, within limits and permissions set in advance. The two companies specify that it was a controlled environment and that the service is not commercially available. It is the detail to keep in mind when reading about agents that buy: the technology works, and mass use has yet to begin.
Who pays when the agent gets it wrong
An agent can buy the wrong thing, buy it twice, or buy from the wrong seller. American Express has answered with a promise: if the customer authorizes a registered agent and the agent sends Amex the authenticated purchase intent, Amex will cover charges resulting from agent error. The promise is written in the future tense and points to terms that have yet to be published, but it shifts part of the risk from the customer to the network.
Mastercard is working on the other side, the bank that has to approve. The first service being tested in the United States is a probability score that a transaction was initiated by an agent. Mastercard’s example is a trip: the agent books the flight, hotel, and transportation with a single approval, and three purchases in a few seconds from three different merchants, seen by a fraud system, look like a stolen card. The score tells the bank that this is normal behavior for an agent. Among the partners in this work is Cloudflare, which brings the signals it collects on the web: it is the same company that measured non-human traffic overtaking human traffic in 2026.
Trust is the bottleneck
The networks are investing because consumers do not trust agents yet. On September 9, 2026, Visa published a Harris Poll survey of 2,065 U.S. adults: 72% have used an AI assistant, but only 23% trust AI to pay on their behalf. Trust rises when a familiar brand steps in: 61% would trust Visa to handle their agent’s payments, 68% among those ages 18 to 34. The survey was paid for by Visa and Visa comes out as the most trusted brand, so that 61% should be read with caution; the 23%, on the other hand, describes a limit that applies to everyone.
The figure matches what was measured among consumers in 2025, when three in five were already using AI to research products but only 17% were willing to delegate the purchase. Mastercard estimates that by 2030 one in ten consumers will routinely use an agent to buy: a company estimate, with no method behind it, but consistent with today’s numbers. Delegating the purchase remains the slowest part of the change, and the networks know it.
The site has to be readable by an agent
The part that concerns brands comes before the payment. On June 10, 2026, Visa introduced Agent Score, a readiness score for agents: it tells a merchant whether an agent can move through the pages, understand what is written there, and complete a purchase. It is the first time a payment network has measured a site’s readability by software and treated it as a condition for selling.
Two months earlier Visa had launched a service that, with a single integration, accepts agent payments coming from the major protocols, including the UCP Target used to multiply its visits from AI. Among the features listed is one that concerns sellers: making catalogs discoverable on AI platforms, with descriptions, specifications, and prices the agent can read. Amex, in its April document, lists among its three priorities making its own benefits “discoverable and actionable across leading AI platforms,” starting with restaurant reservations.
What changes for a brand
When a person buys, the brand has a storefront, an ad, a salesperson. When an agent buys, the brand has only what the agent can read: the site’s pages, the product data, and what AI platforms have learned about it from the sources they consult. The choice happens there, before the person sees a result. A brand that is described poorly, or described by others, gets dropped without anyone noticing.
The work for a company starts with a few simple checks. What ChatGPT, Gemini, and Perplexity say today about the brand, and which sources they learned it from: that is the first measurement in an AI Brand Audit. Whether product data, prices, and terms are written in a way software can read without interpreting. Whether the site lets agents in or blocks them along with unwanted bots. The answers decide whether the brand exists at the moment an agent, acting for a customer, looks for who sells what it was asked to find.
Frequently asked questions
It is commerce in which an AI agent searches, chooses, and pays on a person’s behalf, within limits the person has set. It is the term payment networks use. Today the protocols and the first test transactions exist; widespread use has yet to begin.
In testing, yes. On March 2, 2026, Banco Santander and Mastercard completed Europe’s first payment executed end to end by an AI agent, on live infrastructure and within a regulated banking framework. The two companies specify that it was a controlled environment and that the service is not yet commercially available.
Visa measures it with Agent Score, a score that checks whether an agent can move through the pages, understand the content, and complete a purchase. Without that score, the checks are the same: product data is readable by software, the site lets agents in, and AI platforms describe the brand accurately.
