Back to all posts

cat posts/chatgpt-wallet-skips-eu-how-agents-pay-here-today.md --category "Payment & E-Commerce" --views 15

ChatGPT Wallet skips the EU: how agents pay here today

OpenAI launched ChatGPT Wallet on 13 November 2026 for US Pro subscribers only, with no European date, and calls a security check on every purchase a new bar. Strong Customer Authentication has required exactly that in the EU since 2019, and agent-initiated card payments have been running at European merchants since July through Visa and Mastercard.

  • --author By Falak Mahmood
  • --date November 16, 2026
  • --read 9 min read
  • --views 15 views

OpenAI launched ChatGPT Wallet on 13 November 2026, describing it as the first wallet that lets agents pay without access to a user's real card details and that requires a security check on every purchase. It is available to ChatGPT Pro subscribers aged 18 and over in the United States, with eligible US-issued cards, through the dots agents. There is no European launch and no stated date for one. For a Swedish merchant the interesting part is not the absence. It is that agent-initiated card payments have been running in production in Europe since the summer, through Visa and Mastercard, and the security property OpenAI is announcing has been mandatory here since 2019.

What ChatGPT Wallet actually does

The mechanism is network tokenisation, which is well established. A user adds a card once. For each purchase the user sees the items, the total and the selected card, approves with a biometric check such as Face ID or a fingerprint, and the payment network issues a token. The charge reaches the card without the merchant or the agent ever holding the card number. OpenAI built it with American Express, Mastercard and Visa, and says businesses can accept these payments through their existing systems.

Two things are worth separating. The tokenisation is ordinary: network tokens have been in production for years and are what sits behind Apple Pay and Google Pay. The newer part is binding a per-transaction human approval to an agent's purchase, so the agent proposes and the human authorises. OpenAI has also said it intends to let users delegate purchases to agents entirely, with parameters and preapprovals, which is where the regulatory questions below stop being theoretical.

The new bar has been European law since 2019

OpenAI frames a security check on every purchase as setting a new standard. In the EU and EEA, Strong Customer Authentication under PSD2 has required exactly that for electronic payments since 14 September 2019. Two independent factors from knowledge, possession and inherence, on every transaction, with a defined set of exemptions. A biometric approval on a registered device is a textbook SCA flow.

So the announcement reads differently depending on where you sit. From the United States it is a meaningful raising of the floor, because card-not-present transactions there have never required per-transaction authentication. From Stockholm it is a vendor shipping, in 2026, a property that European regulation forced into every checkout seven years ago. That is not a reason to be smug about it. It is a reason to be precise about what is actually new, which is the agent's role rather than the authentication.

Agents already buy things in Europe

This is the part most coverage of the launch has missed. Agentic card payments went live at European merchants months ago, through the networks rather than through a model vendor.

Visa Intelligent Commerce. At Visa's Payments Forum in Paris on 2 July 2026, Visa reported AI agents completing live purchases at independent European merchants with more than 30 issuing banks participating, naming lastminute.com, Frasers, Cleverbridge and BrickDepot. Agents browse, select and initiate purchases inside consumer-defined limits. Each transaction is authenticated with Visa Payment Passkeys, which ties the payment to a verified user instruction and supports issuers' SCA compliance. Merchant-side recognition runs through Visa's Trusted Agent Protocol and an Agent Directory, so a shop can identify a verified agent and decide how much access it gets to listings and checkout.

Mastercard Agent Pay. Banco Santander and Mastercard announced in March 2026 what they described as Europe's first end-to-end payment executed by an AI agent on production systems. The design has two parts: Agentic Tokens, which are 16-digit tokens tied to a real card but scoped specifically for agent purchases, and Verifiable Intent, a tamper-resistant record of what the user actually authorised.

Both schemes converge on the same three ideas: a credential scoped to the agent rather than the human, a cryptographic record of the user's instruction, and a way for the merchant to tell an agent apart from a browser. ChatGPT Wallet solves the first two inside its own ecosystem. The networks are solving all three across everyone's.

Where the regulation does not fit yet

SCA was written for a human payer. The regulation speaks of factors the payer knows, possesses or is. An autonomous agent has none of these in any sense a supervisor would accept, which is why every live scheme in Europe anchors the authentication to a human act: a passkey, a biometric approval, a signed record of intent. The agent executes. It does not authenticate.

The prevailing legal reading, including a Taylor Wessing analysis from February 2026, is that PSD2 is technologically neutral and therefore already applies when an agent initiates a payment on a customer's behalf. That is a comfortable position for supervisors and an uncomfortable one for product teams, because the obligations exist without any guidance written for the case in front of them.

PSD3 and the Payment Services Regulation do not close the gap. The final texts were published in April 2026, the PSR applies roughly 21 months after entry into force, and most provisions therefore land around 2028. The agreed texts are silent on agents. They do bring spend-limit and transaction-blocking obligations and harmonised liability, which will matter, but nobody should plan on a rulebook arriving before the commercial pressure does.

In the meantime the industry leans on three mechanisms: tokenised credentials scoped to an agent, delegated-authority patterns such as merchant-initiated transactions where a prior agreement exists, and step-up authentication when a purchase looks unusual. Each is defensible. None was designed for this.

What this means if you sell online in Sweden

The practical question is not whether you adopt agentic payments. It is whether agents are already arriving at your checkout and what your systems do when they get there.

Decide your position deliberately. An agent that reaches your checkout today looks like a browser with unusual timing. If you have bot mitigation in front of payments, you may already be blocking a paying customer. If you do not, you may be accepting traffic you have not reasoned about. Both are decisions, and most merchants are currently making them by accident.

Find out what your acquirer and PSP support. Visa's Trusted Agent Protocol and Mastercard's Agentic Tokens only help if the parties in your payment chain implement them. Ask directly whether your PSP can surface an agent signal to you, and whether your acquirer supports agent-scoped tokens. The answer today is often no, which is useful to know before the question becomes urgent.

Treat the agent signal as data, not as trust. An agent identifier tells you a request claims to come from an agent. It is an input to a decision, in the same category as a device fingerprint. Build the plumbing so the signal is captured and logged now, and set policy when the volumes justify one.

// Capture the signal before you need a policy for it.
// Log it, attach it to the order, decide later.
const agentSignal = {
  declared: req.headers['signature-agent'] ?? null,  // agent identity claim
  ua: req.headers['user-agent'] ?? null,
  verified: false,            // true only after a directory lookup
  seenAt: new Date().toISOString(),
};

await orders.update({
  where: { id: orderId },
  data: { agentSignal },      // evidence, not authorisation
});

Keep the human authorisation auditable. Whatever scheme you end up in, the defensible position is the one the networks have adopted: a record linking each payment to a specific human instruction. If a chargeback arrives in eighteen months and the answer is that an agent decided, you want the signed intent rather than a log line.

Watch liability, not features. PSD2 already allocates liability for unauthorised transactions, and an agent acting outside its parameters is the obvious dispute of 2027. Your contract with a PSP, and any terms you accept from an agent platform, should say who carries that. Few currently do.

The Swedish and EU angle

Sweden is an awkward market for this specific launch and a good one for the alternatives. Card penetration is high, Swish is dominant for consumer-to-business payments, and neither ChatGPT Wallet nor the network agent schemes touch Swish at all. An agent that can buy from a Swedish merchant with a Visa or Mastercard credential still cannot pay the way a large share of Swedish consumers prefer to. If your checkout is Swish-first, agentic commerce arrives later for you regardless of what OpenAI ships, and that is worth saying plainly to anyone who read the headline and asked what you are doing about it. The integration mechanics for that rail are in our Swish Handel guide.

The same pattern keeps repeating with this vendor. The dots agents launched without the EEA, Switzerland and the UK for Pro subscribers, with no date, as covered in our analysis of the Dots EU gap. ChatGPT Wallet now ships inside dots, so the exclusion compounds: a product unavailable in Europe, delivered through a product unavailable in Europe. Planning around an announced US feature is not a strategy for a Swedish business. Planning around what Visa and Mastercard have already put into production here is.

One more point deserves saying about the AI Act, because it will come up. An agent that selects and buys on someone's behalf is not automatically high-risk. The transparency duty in Article 50 applies when a person interacts with an AI system, and it has been in force since 2 August 2026. If your storefront talks to agents, or your own assistant buys on a customer's behalf, that duty is live now, independent of anything in payments regulation.

Decision guide

  • Selling to consumers in the EU? Ask your PSP today whether it can pass you an agent signal and whether your acquirer supports agent-scoped tokens. Log the signal now, set policy when volume justifies it.
  • Swish-first checkout? Agentic commerce reaches you later than the headlines suggest. Say so internally rather than starting a project on a US launch.
  • Building an agent that spends money? Anchor every payment to a human authorisation you can produce eighteen months later, and keep the agent's credential scoped and separate from the user's card.
  • Waiting for PSD3 to settle it? The texts are silent on agents and most provisions land around 2028. Design for PSD2 as it stands, since supervisors already read it as applying.
  • Blocking bots at checkout? Check whether you are also blocking verified agents, and decide whether that is what you intend.

Working out what agent traffic means for your checkout? We design and build payment integrations for Swedish and EU merchants, including the audit trail a disputed agent transaction will need.

See our payment integration services →

subscribe # the AI news that matters, minus the noise

Book a Call

Sources

Tags

Related posts