ChatGPT Wallet: OpenAI's Second Try at Agentic AI Shopping

OpenAI scrapped Instant Checkout, its first attempt at in-chat buying, because it wasn’t flexible enough, and has now launched ChatGPT Wallet to let agents pay on a user’s behalf. For merchants and payment teams, the more important news is what OpenAI shipped alongside it: a playbook for letting agent traffic through the bot defenses that currently block it. OpenAI is also arriving after Anthropic, Meta and xAI, whose agents can already shop.

What Happened

The Decoder reports that ChatGPT Wallet launches only in the US, through OpenAI’s Dots agent. Users must be adults with a ChatGPT Pro subscription and an eligible US bank card.

The flow is simple on the surface. A user adds a card once. Before every purchase, the system shows the items, the total and the selected card, and the user approves each payment with a check such as Face ID or a fingerprint. Neither the agent nor the store sees the real card number. American Express, Mastercard and Visa generate a payment token instead: a stand-in card number with its own security code. According to OpenAI, existing card rewards and the usual bank protections for eligible disputes still apply.

OpenAI also describes a fixed purchase path. The customer says what they want, and ChatGPT searches the open web, APIs or merchant product feeds for options. The user compares them and picks one. ChatGPT then completes the purchase either by driving a browser on the merchant’s existing site or through a plugin that supports structured actions such as managing the cart.

Why It Matters

The design separates two jobs that agent shopping has blurred: deciding what to buy and authorizing payment. The agent can search, compare and fill a cart, but each payment still needs a biometric approval from the person, and the merchant receives a token rather than a reusable card number. That limits what a hijacked or confused agent can spend today. OpenAI says it plans to let users eventually hand purchases to agents entirely, within set spending limits, product categories and timeframes. That later step is where the controls will matter most, since nobody will be approving each payment.

Distribution is the other constraint. Many websites and payment systems still block AI agents as bots or reject their purchases for security reasons. OpenAI’s answer is a Commerce Partner Playbook that explains how merchants can adjust firewalls, provide product data and work with providers such as Stripe, Adyen and Riskified so agent-driven purchases go through.

What This Means for Builders

OpenAI is catching up, not leading. The Decoder notes that Anthropic’s Claude Duet, Meta’s Muse and xAI’s Grok Bot already have built-in payments, partly through single-use cards via Stripe Link, while Google and Perplexity run their own checkout tools. Stack Archive’s earlier look at Claude Duet’s proactive agent preview and its coverage of AI agent payment rails show the same pattern: payments are becoming a standard agent capability, and the open question is whether merchants will let the traffic in.

For e-commerce teams, that turns into concrete work. Agents will reach a store either by driving its existing web pages or through a structured plugin. A browser-driven agent depends on bot rules not rejecting it. A plugin depends on product data and cart actions being exposed cleanly. Fraud teams also have a new decision to make: a tokenized, biometrically approved purchase from an agent is a different risk profile from an anonymous headless browser, and rules that treat both the same will turn away legitimate orders.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.