Alipay's Agents Are Buying While US In-Chat Checkout Retires

Most writing about agent commerce is forecasting. Here is a number that is not: Alipay says its AI Pay service processed more than 120 million transactions in the week of 5 to 11 February 2026, which it describes as the first agentic payment service to reach that mark.

That figure is the company's own, in the company's own release, so treat it as a company claim rather than an audited one. It is also the largest concrete number anybody has published about agent-mediated purchasing, and it is a weekly figure rather than a cumulative one.

What is actually running

• Alipay AI Pay, launched in 2025, lets an agent complete a payment. Per the same release it now runs inside apps and mini programs for physical retailers including Luckin Coffee, on smart glasses including Rokid, and in the Qwen app.
• The Agentic Commerce Trust Protocol, launched by Alipay with partners in January 2026. Alibaba's Qwen app was the first platform to adopt it, connecting to Taobao Instant Commerce and AI Pay.
• Qwen itself, which reached 300 million monthly active users across Alibaba's consumer platforms by early 2026, with roughly 140 million first-time AI shopping experiences logged during the Chinese New Year campaign.

The shape worth noticing is not the volume. It is that identity, trust and payment were solved together, by parties who already had all three, and the agent layer was added on top of rails that already worked.

The US went the other way in the same period

While that was happening, in-chat checkout in the US contracted rather than expanded. We wrote about the reasoning at the time in why OpenAI killed Instant Checkout, and the short version is that owning the transaction turned out to be a harder business than owning the recommendation.

Why the rails matter more than the models

Alipay did not need to invent agent identity from nothing. It already knew who the user was, already held the payment instrument, and already had a merchant network that trusted it to settle. Adding an agent meant extending an existing trust relationship by one participant.

In the US, no single party holds all three. The assistant knows the intent, the card network holds the instrument, the merchant holds the inventory, and none of them knows whether the agent in front of them is authorised. That is the gap the emerging identity protocols are built for, and it is why they arrived before the volume rather than after it.

China solved the agent problem by extending a trust relationship that already existed. The US is building the trust relationship first. Those are different timelines, not different levels of ambition.

What this does and does not tell a DTC operator

• It does not tell you agent checkout is imminent in your market. Nothing here is a US or EU forecast, and the figures are from a market with a payment infrastructure yours does not share.
• It does tell you the demand side is not hypothetical. 140 million first-time AI shopping experiences in a single campaign is a large number of people finding the interaction acceptable.
• The recommendation layer arrives before the checkout layer either way. Whether or not an agent can pay on your behalf, it can already decide which brands to name. That is the part you can measure today.

That last point is the practical one. Being nameable does not depend on any protocol shipping. We measure it in the Brand Index, and the current finding is that a meaningful share of well-known DTC brands are not named at all when assistants are asked what to buy in their own category.

The honest summary: one market has agent-mediated purchasing at scale on rails it already owned, another is still assembling the trust layer, and nobody can responsibly tell you which sequence turns out to be the right one. What is not in dispute is that in both markets the assistant is already answering the question that used to start with a search box.