Aktiver — semantic AI platform Deterministic · traceable · behind your firewall
Insights / Agentic commerce

OpenAI gives ChatGPT agents a wallet, with a biometric check on every purchase

On Friday, November 13, OpenAI launched ChatGPT Wallet, a payment layer that lets its agents, called dots, complete purchases on a user’s behalf. TechCrunch reports that the agent helps choose the items and then checks out, while the user reviews the items, the total and the card to be charged and approves each purchase with a security check such as Face ID or a fingerprint scan.

Access is narrow at first: a limited group of U.S. ChatGPT Pro subscribers aged 18 or older, using U.S.-issued Visa, Mastercard or American Express cards. According to ID Tech, the wallet was built with those three card networks, and OpenAI says it will widen availability as it works on compatibility and reliability with businesses, banks and payment networks.

/ 01How the wallet works

A user adds a card once. When a purchase is approved, the wallet pays with a payment token issued by the card network rather than the card number, security code or expiration date, so neither the dot nor the merchant sees the real card details. The agent passes the merchant the shipping address and contact information needed to fulfil the order. Each approval creates a record of the purchase the user authorized the agent to make.

The ordinary card machinery stays in place. Android Authority reports that OpenAI says the wallet preserves existing card rewards and does not affect card protections for eligible disputes, and ID Tech notes that banks and networks continue to evaluate each transaction.

What the launch does not include is just as telling. TechCrunch reports there are no spending limits at launch. OpenAI plans to add spending caps and rules based on purchase category or time period, so that some purchases can be pre-approved and handled with more autonomy. In OpenAI’s words, as quoted by Android Authority: “Over time, we expect users to be able to delegate purchases entirely to their agent by setting clear parameters.” Those pre-approvals, it says, would come with “enhanced safeguards” that build on biometric checks and authorization records.

/ 02Why it matters

Agent shopping has so far been held back less by models than by the plumbing around them. Retailers have been wary: TechCrunch notes that Amazon is the largest retailer to block AI agents, steering users toward its own assistant, and that DoorDash and Uber have reported low volumes of agent-placed orders. Android Authority adds that United Airlines has announced a policy restricting agent access to its digital platforms.

OpenAI is addressing the merchant side directly. Alongside the wallet it published a Commerce Partner Playbook, which ID Tech says asks websites to configure their CDNs and firewalls to recognise signed requests from ChatGPT agents using Web Bot Auth, and to work with their payment and risk providers, because fraud systems may flag genuine agent purchases. OpenAI says it is working with Forter, Riskified, Sardine, Signifyd, Adyen, Checkout.com and Stripe. For American Express, ID Tech reports, this is the first integration under what it calls Agentic Commerce Experiences.

The payment networks frame the issue as one of trust. “An agent making a purchase should strengthen, not bypass, the trust built into payments,” Sherri Haymond, Mastercard’s executive vice president and global head of digital commercialization, told ID Tech. With ChatGPT reporting 1.2 billion weekly users, the scale at which that trust gets tested could be large once access widens.

/ 03The traceability angle

Today’s design keeps a human in every transaction. The user sees the basket, the total and the card, and a biometric check ties the approval to a person on a device. The resulting record is simple: this user approved this purchase at this time.

The planned next step changes that. Once a user can say “buy groceries under a set amount each week” and let the agent proceed without a tap, the authorization record has to carry more. It has to show which standing rule the agent relied on, how the agent classified the product against the rule’s category, what price it saw, and which user request started the chain. A dispute about an agent purchase will turn on those links, not on whether a card token was valid.

The same applies to the merchant side. An agent matching a request to a product depends on the catalog: categories, variants, prices and stock. If the agent misreads what a product is, a category rule can be satisfied on paper and violated in substance. Structured, explicit product data and an auditable link from each order back to the rule and the request are what make that kind of error visible after the fact.

When an agent spends money, the receipt has to say which rule let it.

/ 04What to watch

  • The delegation controls. How OpenAI defines spending caps, purchase categories and time windows, and what the authorization record captures once purchases no longer need a per-order biometric check.
  • Merchant uptake. Whether major retailers that restrict agents change course, and how widely the Web Bot Auth approach in the Commerce Partner Playbook is adopted.
  • Disputes. How card issuers handle chargebacks on agent purchases, given that card protections for eligible disputes are said to carry over.
  • Rollout. When access extends beyond U.S. Pro subscribers, and to which card types and markets.
Aktiver’s take — Telecom, media & e-commerce

Per-purchase biometric approval is a sound starting point, but the authorization record is the part that will matter. Once spending caps and category rules arrive, the question changes from “did the user tap approve” to “which rule allowed this order, and did the agent read the catalog correctly”.

That is a graph problem. Model the user’s delegation as explicit rules, the merchant catalog as products with categories and prices, and each agent purchase as a node linked to the request that started it, the rule that permitted it and the product it matched. A support or risk agent can then answer, with evidence, why an order was placed and whether it stayed inside its limits, and give the same answer every time it is asked. Ingestion Hub turns catalog feeds in CSV, JSON or SQL into an ontology that a merchant’s own experts refine; Knowledge Builder agents are held to that ontology rather than to a guess about what a product is.

How we’d model it

Agent purchaseDelegation rulePurchase categoryCatalog productAuthorization recordauthorizedBylimitsmatchesrecordedInAgentpurchaseDelegationrulePurchasecategoryCatalogproductAuthorizationrecordauthorizedBylimitsmatchesrecordedIn
  • Agent purchase —authorizedBy→ Delegation rule
  • Delegation rule —limits→ Purchase category
  • Agent purchase —matches→ Catalog product
  • Agent purchase —recordedIn→ Authorization record

See where this applies →

Go deeper

See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.

← All insights Subscribe via RSS → Book a demo →
Keep reading