AIANALYSISPODCAST
Card networks, model labs, and a new crop of startups are racing to build the payments layer for AI shopping. The unresolved question is not the tech. It is the bill.
On June 3, 2026, Clink opened public access to a piece of payment infrastructure designed for one specific kind of buyer: an AI agent that nobody physically hands a card to. The launch sat at the end of a sequence of events that started earlier in 2026, when Clink executed what it calls the first Asia-Pacific agentic transaction through Visa's Visa Intelligent Commerce (VIC) integration. Visa itself marked the milestone as the first time its partners completed "secure agentic transactions" end to end. The agent bought something. The rail processed it. The receipt had a name on it. The question the launch was really answering was who would be on the hook if the agent had bought the wrong thing.
The race to wire AI agents into the existing payment stack is being sold as a wallet story. It is not. The wallet is the easy part. A recent Clink press release reads as "and now anyone can use ours." The product a startup can actually build a moat around is the bill, the dispute, and the merchant of record.
Three protocols, one checkout
Google's Universal Commerce Protocol (UCP) and OpenAI's Agentic Commerce Protocol (ACP) each try to own the full checkout grammar: catalog, promotions, loyalty, shipping, and payment. Merchants asked to integrate both end up writing two parsers for the same product page, and a third if they want a card network in the loop. Visa's APAC agentic-ready program launched with 50-plus partners across the region to push VIC as the rail beneath either protocol, and the U.S. press release framed the same strategy as opening a global door for AI-driven shopping. Clink's pitch to builders, aired in detail on the Chinese-language panel 可以给你的 Agent 发一点零花钱了, is to sit above all of it: a "ramp layer" that connects UCP, ACP, Shopify, Magento, and WooCommerce so the merchant writes the integration once. Stripe, by contrast, is both the acquirer and a protocol author, which is why a neutral ramp has room to exist.
Why fiat won the first round
The choice of rail is not aesthetic. Clink and most of the early agent-commerce builders have stayed on card and bank rails rather than stablecoin, and the reason is not yield. Card networks come with a century-old dispute regime: a chargeback, a chargeback code, a liability path. Stablecoin is irreversible once paid, and consumer confidence in an agent buying a $400 plane ticket on a non-refundable rail is, by the builders' own admission, the hardest part to engineer around. Stablecoin escrow, a smart contract that releases on delivery confirmation and is sometimes called "on-chain Alipay," is being explored by Stripe and Coinbase to soften the irreversibility. Visa and Mastercard have bought stablecoin firms, mostly for B2B cross-border SWIFT replacement, not for consumer agent flows. The escrow pattern shifts liability; it does not eliminate it. Someone still has to sign the receipt.
The China counterfactual
In China, the same problem is being answered with a different trust regime. Alipay's Agent Marketing Platform (AMP) and Tencent's Agent Card both route agent spend through mobile wallets that already own the consumer relationship, so the merchant-capture problem Western panelists keep circling does not exist. What does not exist there, either, is the dispute and refund regime that makes U.S. and EU card rails feel safe for delegated spend. China runs a higher-trust experiment: the wallet is the customer, and the consumer has already opted in. The U.S. and EU run a different one: the card is the customer, and a chargeback is the safety net. Which trust regime a merchant ends up in depends on geography, not on which protocol the agent is using.
The watch item
The protocol war is a UX integration problem a ramp layer can paper over, so the winner there matters less than the trust regime underneath. The thing to watch is whether the chargeback regime, the merchant-of-record convention, and the card-credential custody model that card networks spent a century building stay intact when the buyer is non-human. The agent-payments race is being framed as a wallet race. The product that ships will be decided by who owns the bill when the bot is wrong.