ChatGPT Wallet Can Pay for You, but It Never Sees Your Card Number

Published on: 2026-11-15

ChatGPT Wallet Can Pay for You, but It Never Sees Your Card Number

Two announcements landed in the same week, and they answer the same question in opposite ways.

On November 13, OpenAI introduced ChatGPT Wallet. Agents inside ChatGPT Work, Codex and dots can complete a purchase once you approve it — you can ask a dot to help decide what to buy, then let it check out on your behalf. On November 14, Google added Google Wallet to the set of apps that Gemini's personal intelligence can connect to. Once authorised, typing @Google Wallet or @wallet into a conversation lets Gemini answer questions about cash flow, asset allocation and budgets, surface subscriptions you forgot about, flag recurring charges, and mark price increases.

One can pay. The other can only look. That reads like a difference in product strategy, but the more durable change is somewhere else: for the first time, permission itself has been broken into visible layers. Consent used to be a single button. Now it is a set of decisions, and each one has a different consequence.

One: is the agent reading, or is it paying?

📖 Glossary

AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.

KAIHE AIBOX concept device on a studio steel desk beside an EACH PURCHASE holographic sphere panel

On the ChatGPT Wallet side, money actually moves. OpenAI developed the feature with American Express, Mastercard and Visa, charges no commission on transactions at this stage, and says it builds on the protections already attached to your existing card. The official ChatGPT account made two narrower claims as well: the agent does not see your actual card number, and every purchase goes through a security check.

Google drew a harder line. Connecting Google Wallet lets Gemini read only the account information you authorise. It cannot transfer money, manage funds, or invest. Credit cards and bank accounts are linked through Plaid. What the assistant can do is reorganise spending history into charts and reminders — forgotten subscriptions, recurring charges, price increases.

This layer sets the boundary around money, and it has nothing to do with how capable the model is. One can change your balance; the other can change your mind. The difference is not technical difficulty but the kind of mistake you have to live with: a bad summary can be re-read, while a bad payment has to be recovered.

Two: is approval given once, or requested every time?

The rollout tells you more about the design than the feature list does. ChatGPT Wallet starts as a limited release: the United States, adults 18 and over, ChatGPT Pro subscribers, with a US billing address and an eligible card issued in the US. That is not a standing mandate handed to an agent. Each purchase is pulled out and checked on its own.

Google does not talk about individual transactions. It talks about the entrance. The connection has to be switched on manually under personal intelligence settings, in connected apps; it is opt-in, and it can be revoked at any time. Josh Woodward, vice president of Google Labs, said the rollout would first reach all Gemini app users in the US.

Put the two together and a shared choice appears: neither made "authorise once, valid forever" the default. The one that can pay pushes the check down to each transaction. The one that can only read leaves the switch in your hands. Granularity matters more than the yes-or-no decision. Where the permission stops determines where you can still intervene when something goes wrong.

Three: can you take it back?

KAIHE AIBOX concept device on an attic desk beside a READ ONLY holographic windows panel

This is the question people ask least often, and it belongs first.

Was the card number handed over? OpenAI says the agent never touches it. Can the permission be withdrawn? Google says yes, at any time — and in any case it cannot move money. Both sides appear to have kept an escape route.

One sentence still needs adding. Revoking access takes a second; un-reading data does not. Read-only is not the same as taking nothing. What gets read here is spending behaviour, which is not a balance and yet says more about a life than a balance does: when you buy, what you buy regularly, which trial you forgot to cancel, which price rise you missed. Once that pattern has left where it used to live, "revoke" no longer covers it.

Permission governs whether an AI may touch something. It does not govern where that something was in the first place. Those are separate problems, and they are routinely treated as one.

Where the edges are

Three lines are worth holding.

Both announcements are US-only for now. Do not rebuild a workflow around "I can use this today", and do not read a limited release as a general launch.

Several widely circulated capabilities have not shipped. Per-purchase biometric approval and tokenised payments that replace card numbers come from non-primary sources. Spending caps, category and time-window limits, and fully delegated purchasing are described as plans, not current behaviour. They belong in a sentence about what may come later, not in a list of what works now.

Neither is obviously safer than the other, and that comparison is not useful. Different permission scopes produce different shapes of risk: on one side money can be spent, on the other habits can be read. There is no single scale for those two things. What you can do is decide which consequence you are willing to carry.

Using the three questions elsewhere

The same sequence works well beyond wallets. A signed-in browser session, a cloud drive connection, calendar access, permission to send mail on your behalf — all of them are versions of the same decision. Ask in this order:

Read or execute. Is it observing, or is it acting?

Once or each time. Did you grant a standing mandate, or does every action come back to you?

Reversible or not. Where is the switch, how fast does it work, and what has already left?

Three questions before one click is cheaper than an afternoon spent hunting through settings afterwards. That, more than either product, is what this week leaves behind.

Both companies made permission finer this week: per-transaction checks, read-only access, revocation at any time. That is the right direction. But permission answers "may the AI touch this", not "where was this kept". Access can be cancelled in a second; data that has left the machine cannot be pulled back. KAIHE AIBOX keeps files and memory on the box you own, so an agent can work without moving them out first. Whether to let AI near a wallet is the next decision. Knowing which things never left is the one before it.

Recommended Products

A1 Home Entry A1 Pro Enhanced A2 Professional A2 Pro Advanced X1 Enterprise G1 Flagship
© KAIHE AI - Agent Computer Specialist