WeChat Pay announced that its AI-only card now supports two more apps, DeepSeek Harness and OpenClaw. The card was already linked to WorkBuddy and QClaw. Once a user authorizes and binds the card, they can simply state what they need in a chat, and the AI handles everything from recommending options to placing the order and paying.
Right now, the card can be used to pay for more than 700 Pay Skills listed on Skillhub.
A card kept separate from the main account
The core of the design is account structure. The AI-only card is fully isolated from a user's main WeChat Pay account. Users set their own balance on the card and can transfer money in or out at any time — the AI can only spend what's sitting in that card. Every deduction still requires the user to confirm a second time before it goes through.
The official flow is four steps: copy the install command, authorize and bind the card, let the AI call a Pay Skill, and confirm the payment. Public reports have quoted a comparison for it: “it's like when you were a kid and an adult handed you cash to go buy them a pack of cigarettes.” The adult hands over a specific amount, and the wallet stays in their own pocket.
Translated into risk-control language: a sandboxed account with a spending cap, plus a manual approval step. Whether AI agents should be allowed to pay for things has been debated for the past year, and the sticking point has always been trust. An isolated account is currently the cheapest fix — it doesn't require rebuilding the underlying payment system, and it doesn't hand the model access to the main account.
700 Pay Skills are the supply side
Getting payments working is only half the job; the other half is having something worth buying. The 700-plus Pay Skills on Skillhub are that supply: these skills charge for use, and when the AI calls one, the money comes out of the dedicated card.
That's a different logic from traditional e-commerce. Payments used to be attached to physical goods; this time they're attached to capability — a single API call, a chunk of video rendering, a data query. The price per call is usually tiny, but the frequency can be high. And that's exactly the kind of scenario where a second confirmation hurts the most: if a popup appears every time you spend a few cents, the experience falls apart.
That's why setting the spending limit matters more than it looks. When a user loads a fixed amount onto the card in advance, they're effectively front-loading the authorization step; every confirmation after that is really there to stop the model from going off the rails, not to prevent fraud.
Who's responsible for a bad purchase
When a model places the order, responsibility isn't clearly defined yet. Legally, the user pressed the pay button, so it counts as the user's own decision — but the user is really just confirming a plan the model came up with, and it's hard for them to tell whether that plan contains a hallucination or was somehow steered.
Overseas, Visa has already connected its payment network to ChatGPT, following a similar path: turning spending limits and authorization into a middleware layer. The fact that this wave of moves in China and abroad landed so close together suggests everyone has converged on the same first answer to “should AI be allowed to touch money” — start with a small, instantly resettable card.
The real test will come once this scales up. Once hundreds of thousands of small transactions initiated by models are running every day, the entire process for complaints, refunds, and dispute resolution will need to be rewritten. For now, this isolated-account design keeps the risk capped at the limit of a single card — and that's the smartest part of it.
Sources: WeChat Pay official announcement, CocoLoop, IT Home; the card's balance-isolation mechanism, the double-confirmation flow, and the number of callable Pay Skills on Skillhub have been verified.