OpenAI introduced ChatGPT Wallet this week, a payments feature that lets agents complete purchases on a user's behalf. The phrase that carries the design is the one in the announcement: explicit user approval, every time.

What it is. The wallet is built into ChatGPT Work, Codex and OpenAI's dots platform. An agent that decides a purchase is needed can initiate a transaction, and the user approves it before anything is charged. OpenAI says existing card protections apply, which is the reassurance that lets a bank or a card network treat this as a normal card-not-present transaction rather than a new category of risk.

Why a wallet is the missing piece. For two years the demo of agentic commerce has been easy and the last mile has been impossible. An agent can compare prices, fill a cart and find a discount code. It cannot type a card number, and every party involved has a reason to stop it: the merchant does not want an unattributable actor at checkout, the bank does not want to adjudicate a dispute with a bot, and the user does not want a model hallucinating a purchase.

This week made that friction visible. United Airlines moved to block unauthorized agents from booking and managing trips, citing credential exposure and rules that require a human at check-in. Amazon has blocked Meta's AI agent from shopping on its platform. Chase's updated terms bar external AI from using customer credentials. The pattern is not technophobia. It is institutions refusing to accept an actor they cannot identify, authorize or hold responsible.

A wallet with a mandatory human approval step is OpenAI's attempt to supply exactly that: a named principal, a recorded consent, and a rail that already has chargeback rules.

What is still unknown. OpenAI has not laid out merchant coverage, per-transaction or daily limits, how disputes between a user and a merchant are handled when the purchase was agent-initiated, or whether merchants can refuse wallet transactions. Each of those is a detail that decides whether this is a convenience feature or a payments network.

The strategic stakes are high because the wallet is a natural home for the rest of the money in the AI stack. Agent subscriptions, API usage, marketplace purchases by Business Agents, and eventually spending on behalf of a company rather than a person all want the same primitive. Whoever owns the approval screen owns the relationship, in the way that the card network owns the checkout page today. Google has its own payments stack and an assistant that can already transact; Apple has the device and the card; Meta is stitching commerce into its agents.

There is a regulatory angle too. A wallet that holds a balance and settles with merchants starts to look like a money transmitter in some jurisdictions, and the compliance burden is the reason most AI companies have stayed away from this for as long as they have.

What to watch. Whether merchants opt in or block the wallet the way United blocked agents, what chargeback rates look like in the first quarter, and whether approval stays mandatory once users get tired of tapping confirm. The interesting question about agentic commerce was never whether a model can buy something. It is whether anyone will accept its money.