OpenAI launches ChatGPT Wallet for Dots agents to buy with your card

The launch

OpenAI began rolling out ChatGPT Wallet on November 13, 2026, according to a report by The Verge. The wallet allows Dots agents, a platform launched earlier this year, to make purchases on behalf of the user using a card registered to the account.

The feature is currently available only to ChatGPT Pro subscribers in the United States who are over 18, have a US billing address, and eligible cards from American Express, Chase, Mastercard, and Visa.

Every purchase goes through a confirmation step before it's finalized: the agent notifies the user, who must approve with biometric verification (Face ID or fingerprint scan) before the payment is completed. There is currently no automatic purchase without this step.

How the payment works behind the scenes

According to OpenAI, ChatGPT Wallet does not share the actual card number with either the Dots agent or the purchase website. Instead, the company uses payment tokenization, the same model used by tap to pay, Apple Pay, and Google Wallet: the physical card becomes a token that represents the transaction, and it's this token that circulates between the agent, the wallet, and the store.

This architecture matters for developers: from the perspective of an e-commerce backend, a purchase made by a Dots agent arrives like any other tokenized transaction. In principle, there is no exposure of the actual card number at any stage of the flow.

Full delegation doesn't exist yet (but it's on the roadmap)

Today, every purchase requires explicit human approval for each transaction. OpenAI, however, already signals the next step on its support page:

users to be able to delegate purchases entirely to their agent by setting clear parameters—such as spending limits, purchase categories, or time periods—and pre-approve those purchases.

OpenAI, ChatGPT Wallet support page

This is the point that matters to anyone building on top of agents. The human-approval-on-every-action model is the simplest safety net to implement, but also the one that scales the least. Rule-based authorization (spending limit, category, time window) is the same type of control that any financial delegation system already has to solve, except now applied to an LLM agent, which can be induced by prompt injection to misinterpret an instruction.

Approval fatigue: a risk the design already knows

The technical reading here is that this confirmation flow repeats an old security UX problem: the more times the system asks for biometric approval, the greater the chance the user approves on autopilot, without reading what they're confirming. It's the same effect observed on mobile app permission screens and OAuth consent prompts clicked without reading.

If OpenAI moves toward rule-delegated purchases, this risk shifts: instead of click fatigue, the problem becomes the quality of the rule the user sets at the moment they define spending limit and category, a configuration error there has a direct effect on the real credit card.

The obstacle that isn't technical: stores block agents

Not every store will accept a purchase made by a Dots agent. OpenAI itself acknowledges this:

websites need to ensure agents are not inadvertently blocked, and payment and risk systems need to recognize legitimate purchases made by agents.

OpenAI, ChatGPT Wallet support page

This conflict has already surfaced before: earlier this year, Amazon blocked Meta's Muse agents on its shopping platform. It's a snapshot of a problem without a standardized solution yet: the same antifraud and anti-bot systems that protect a store against scraping and malicious automated purchasing also block, by default, any traffic that looks like an agent, legitimate or not.

What changes for builders

For engineering teams that maintain e-commerce, marketplace, or any payment system, ChatGPT Wallet is a signal: AI agent traffic doing checkout stops being a hypothesis and becomes a real production case.

• Bot detection needs a third category. Today most antifraud systems classify traffic as either human or malicious bot. A Dots agent making a legitimate purchase, authorized by the cardholder's biometrics, is neither, and risk systems that can't tell the difference will generate false positives.
• Tokenization is already a prerequisite, not a differentiator. Anyone who still processes the actual card number at any point in checkout will feel more pain as agent purchases grow, because OpenAI's model, like Apple Pay's and Google Wallet's, assumes end-to-end tokenization.
• Rule-based authorization is the next battleground. If the full delegation promised by OpenAI gets off the ground, the logic of spending limit, category, and time window will need to be validated on both the agent's side and the side of the card issuer and the merchant, a distributed architecture problem, not just an interface one.

What remains open

OpenAI has not detailed a timeline for full purchase delegation, nor which card issuers beyond the four mentioned will join. There is also no industry standard yet that tells a merchant how to reliably distinguish an authorized AI agent from an ordinary bot.

Until that standard exists, each store decides on its own whether to block or allow agent traffic, as Amazon did with Meta's Muse. For those building payment infrastructure today, the practical question has already arrived: does your system's checkout recognize a legitimate agent, or does it treat everything that isn't human as fraud?