
Agent payments, one year in
OpenAI, Google, Visa, Mastercard and PayPal each built a way for agents to buy. What has launched, what the numbers show, and what a shop must prepare.
A year ago, an assistant that buys for you was a demo. Since September 2025, the companies behind the assistants and the companies behind the payments have each published a way for agents to shop, and several have moved real money with it.
The picture in September 2026 is clearer than the noise around it. Agents increasingly do the looking. People still approve the buying. And shops do most of the preparing.
Who built what
OpenAI and Stripe went first. In September 2025 they released the Agentic Commerce Protocol, open source under the Apache 2.0 licence, and launched Instant Checkout in ChatGPT for Etsy sellers in the US. Stripe added a payment token limited to one merchant and one basket value, so the assistant could start a payment without seeing the card. The shop stays the merchant of record.
Google answered the same month with its Agent Payments Protocol, AP2, backed by more than 60 organisations. It works on proof. A signed intent mandate records what the person asked for, with price limits and conditions. A cart mandate records the exact items and price the person approved. In April Google handed AP2 to the FIDO Alliance and released a version for purchases when the person is not present, within instructions given in advance, such as tickets bought the moment they go on sale.
In January Google added the Universal Commerce Protocol, developed with Shopify, Etsy, Wayfair, Target and Walmart and endorsed by more than 20 others, Visa, Mastercard, American Express, Stripe and Adyen among them. It covers the journey from discovery to after sales, and Google said it would power checkout from eligible US retailers in AI Mode and the Gemini app, with the retailer as seller of record.
The card networks built for trust. Visa Intelligent Commerce includes the Trusted Agent Protocol, which gives a merchant a consistent signal of an agent’s identity, so verified agents can be told apart from other automated traffic. Mastercard Agent Pay gives developers a way to identify their agents, and with the FIDO Alliance Mastercard is developing a credential that confirms amount, merchant and product, so everyone involved can see that the shopper approved the payment. PayPal introduced Store Sync in October 2025, which sends a merchant’s catalogue to ChatGPT, Microsoft Copilot and Perplexity, and announced Agent Ready, payments from AI surfaces with buyer protection, for early 2026.
The March correction
The most useful lesson came in March, when OpenAI pulled back from Instant Checkout. It said the first version had not offered the flexibility it wanted, let merchants use their own checkout, and put its effort into product discovery. Analysts told CNBC that OpenAI had struggled to onboard merchants, to show accurate product data, and to support baskets with several items or loyalty memberships.
What replaced it is telling. Merchants now send product feeds and promotions to ChatGPT, and Target, Sephora and Nordstrom already do. Walmart launched its own app inside ChatGPT, with account linking, loyalty and Walmart payments. Shopify said its merchants would complete purchases in a browser inside the app. The protocol survived. The idea that the assistant owns the till did not, at least for now.
What has actually launched
The clearest evidence comes from the card networks, and it is early. On 2 July Visa said more than 30 European issuers had completed live transactions executed by agents, at merchants including lastminute.com, Frasers, Cleverbridge and BrickDepot. Swisscard, Viseca and Cornèrcard are among the banks named. Every purchase was authorised by the consumer and secured with a Visa payment passkey. The chief executive of lastminute.com called it early days.
In March, Santander and Mastercard completed what they called Europe’s first live payment executed end to end by an AI agent, inside the bank’s regulated framework and expressly not a commercial rollout. In the US, Mastercard had said all its cardholders would be enabled for Agent Pay by the 2025 holiday season.
Demand is easier to see in referrals than in purchases. Adobe measured traffic from AI services to US retail sites up 693 percent over the 2025 holiday season and 393 percent in the first quarter of 2026. In July it was 62 percent above a year earlier, and those visitors converted 60 percent better than other traffic, the eleventh month in a row they did better.
None of these figures counts purchases completed by an agent, and the announcements we have read give no volumes. People research in the assistant and buy on the shop’s own site. The protocols are ready before the habit is.
What a shop must prepare
Product data comes first. Agents buy from what they can read, and Adobe found that about 34 percent of product pages cannot be properly accessed by AI. OpenAI’s retreat was partly a data problem. Price, stock, variants, delivery times, returns and tax belong in a feed and on the page, current and identical in both. Google has added dozens of new Merchant Center attributes for this, from answers to common product questions to compatible accessories and substitutes.
Then the checkout. Do not build for a single protocol. PayPal calls itself protocol agnostic, and Visa’s Intelligent Commerce Connect handles ACP, UCP, the Trusted Agent Protocol and Stripe’s Machine Payments Protocol through one integration. For most shops the route runs through the payment provider and shop system they already use, plus a checkout a guest can finish without obstacles.
Then trust. Decide which agents you let in. A site that blocks every bot will start to block paying customers, and one that admits everything invites scraping and fraud. Signed agent identities are how the networks propose to tell the two apart.
Then fraud and disputes. Google wrote AP2 around three questions: did the user authorise this purchase, does the request reflect what they really wanted, and who is accountable if it goes wrong. Ask your payment provider the same three, and how a dispute over an agent’s purchase is handled.
This is the work Cited and Studio are for. Cited makes a company readable for the assistants, with structured data, clear facts and pages they can quote, measured every week. Studio keeps the website and its content in the company’s own repository, where an assistant drafts and a person decides.
What stays a person’s decision
Every launch we have read keeps a person at the deciding point. Visa’s European purchases were authorised by the consumer. Google’s auto browse in Chrome is designed to pause and ask before it buys anything. Even AP2’s unattended purchases run only on a mandate the person signed in advance, with limits.
That is the right default. An agent can compare offers faster than any of us. Knowing what a household or a company actually needs, and answering for the money spent, stays with people. The same holds on the shop’s side: prices, discounts, return rules and which agents may buy are the owner’s decisions, written down clearly enough for a machine to follow.
Start with the feed. Make the product data complete and current, check that assistants can read it, keep the checkout simple, and ask your payment provider which agent programmes it already supports. All of it helps human buyers too.




