"Order another milk tea from that shop I bought from yesterday and have it delivered to me."
If one day you could really say this to your phone and it would find the merchant, confirm the product, place the order, make the payment and arrange delivery for you, you probably would not first think of some payment revolution. You would most likely just feel that your phone has finally begun to act like an assistant that can actually get things done.
This may be the more appropriate starting point for our discussion of AI payments today.
Because what most people understand by AI payments is not giving AI a wallet, nor letting a robot decide on its own how to spend money. What people can genuinely feel is whether a single sentence can be turned directly into a transaction. You state your need, and the system completes everything that follows on your behalf.
So what is truly worth discussing about AI payments is often not the word "payment" itself, which is too empty and too remote. For the vast majority of internet users, the felt experience is likely to be more down to earth: when I express a shopping need to an app or an AI assistant in plain language, it can get it done directly.
The intuition of Chinese users
If we look only at the Chinese market, AI payments are in fact not that unfamiliar.
The reason is simple. Chinese users have long lived within a highly mature mobile internet: food delivery, e-commerce, maps, ride-hailing, local services, mobile payments and instant delivery are inherently connected. In many cases, what users want is not an entirely new payment method, but simply not having to switch back and forth between apps and not having to repeatedly tap the same buttons.
ThereforeIn China, AI payments are more like an upgrade to the e-commerce shopping experience than a rebuilding of the payment system from scratch.
When Alibaba upgraded the Qwen App in January this year, the direction it set out was typical. According to the official introduction, Qwen has been connected to services within the Alibaba ecosystem such as Taobao, Taobao Flash Sale, Alipay, Fliggy and Amap, and users can, through a single sentence, order food, complete payment within a chat, and plan and book trips. It looks as though AI has become smarter, but more accurately, capabilities that the platforms had already connected are now being reorganized.
This is also why what makes AI payments most viable in China is not some grand vision, but a number of highly everyday scenarios.
For example, “Help me buy another box of the printing paper I bought last week and have it delivered to the office.”
For example, “Order the same takeout from that restaurant again from yesterday and have it delivered to my home.”
Another example: “Book a restaurant near the office for four people, with a budget of 500 or less.”
Users understand these statements immediately. That is because the AI here does not resemble a mysterious new financial instrument; it is simply a smarter shopping assistant.
Whatever the public needs, China's internet giants, locked in relentless competition, will build. Today, with respect to AI payments, we can roughly identify three implementation paths:
One is to have the AI watch the screen and operate on your behalf, clicking buttons, filling in information and submitting processes like a human. The advantage of this approach is that it can be tried anywhere; the drawback is also obvious: it is prone to problems whenever the page changes, and at critical points you often still have to confirm things yourself in the end. Take, for example, the Doubao phone that was besieged by WeChat as soon as it left the factory.
Another is for platforms simply to hand their own service capabilities directly to AI to invoke. For China's big tech companies, this path is actually the smoothest. That is because payments, merchants, maps, orders and delivery were all already in their own hands. The AI does not need to imitate users clicking around; it can directly invoke capabilities within the system. As users experience it, this means “I say one sentence, and it actually gets it done for me.” Examples include Alibaba's Qwen and ByteDance's Doubao.
The third, and more difficult, approach is the cross-platform, cross-website and cross-company one. At this stage, the question is no longer merely whether the product is easy to use, but whether others are willing to let your AI come in on behalf of users to do this. At present, we do not appear to have seen any Chinese internet company attempt to take on this hard mode.
Today, China's greatest advantages clearly still lie in the first two approaches, especially the second.
This is also why, for Chinese users, AI payments may not seem all that earth-shattering; the overall impression is likely not “AI payments have arrived” but rather “shopping on a phone is more convenient.”
The situation overseas is different
But once we shift our perspective overseas, things become somewhat different.
Overseas markets of course also have e-commerce, food delivery, maps and payments, but they do not have China's highly integrated platform structure. Service entry points, merchant systems, payment tools and fulfillment networks are often held by different companies. As a result, once AI wants to complete a transaction on a user's behalf, the question it faces is not merely "can an order be placed," but "will others recognize you as an AI agent."
This is why overseas AI payments suddenly become complicated.
A very typical example is the conflict between Amazon and Perplexity. Perplexity's Comet browser treats AI shopping as an important selling point, and its official description directly states that it can help users compare products and read reviews all the way through to checkout. This in fact closely matches many people's most intuitive idea of AI shopping: you do not have to browse website by website yourself; instead, you let AI help you choose and help you buy.
But the problem arises precisely here.
According to a court document dated March 9, 2026, one of Amazon's core arguments in the litigation is that Perplexity's AI agent accessed user accounts and attempted to perform operations under circumstances where "the user was willing, but Amazon had not consented."
The reason Amazon did not consent is broadly the same as the reason WeChat did not consent to the Doubao phone: your authorization to AI does not mean the website has also authorized AI. Your willingness to let it buy things for you does not mean the platform is willing to let a third-party agent enter its system and initiate transactions on your behalf.
Of course, the external explanation is based on considerations of customer privacy and security, but the core concern is still that AI shopping will directly cut off the advertising fees, amounting to tens of billions, that its own sellers spend on placements.
After all, AI shopping has a straightforward style: whatever needs to be bought, it goes directly for the product, and it will not look at advertisements.
In China, such conflicts are often absorbed by the integrated capabilities within platforms; but overseas, once AI needs to complete transactions across websites, across merchants and across payment networks, these problems immediately cannot be avoided.
Therefore, overseas, AI payments are first of all not a question of "whether it is easy to use," but a question of "whether this matter can actually be completed end-to-end as a closed loop."
Solutions for overseas AI payments
This is also why Google, Stripe and Coinbase have recently all been pushing hard on this matter.
What they see is the same problem: in the internet transactions of the past, the default assumption was that the person placing the order was a human, and that this person was sitting in front of the screen, browsing, confirming and paying themselves. But once the person placing the order becomes an authorized AI, that original default premise no longer holds. How does a website know that this AI was really sent by you? How does a merchant confirm that it has not made random purchases? And how should a payment institution determine whether this payment should be approved?
Google wants to solve the trust problem at the very front first. It launched AP2 in September 2025, not to teach AI how to pay, but to answer one question first: when an AI initiates a payment on behalf of a user, what does the system rely on to believe that this is real? Google's approach is to add a set of verifiable "authorization proofs" to this type of transaction. In AP2 it designed two very critical things: one is called Cart Mandate, which can be understood as the user's signed confirmation of this specific purchase; the other is called Payment Mandate, which is shown to payment networks and card issuers, telling them that this is a transaction initiated by an agent, whether the user was present at the time, and what scenario the transaction belongs to. Google later connected this system with PayPal to create a more complete merchant solution: merchants can use their own conversational shopping assistants to receive users, and at the payment step, PayPal Agent takes over and completes authorization and payment in the AP2 manner. Google's path, in the final analysis, is to first fix the matter of "on what basis should AI orders be trusted."
Stripe's approach is closer to merchants and platforms. It has in fact built two layers. The first layer is ACP, namely the Agentic Commerce Protocol, launched together with OpenAI in September 2025. You can understand it as a set of "AI-readable checkout standards." Stripe does not want AI to imitate humans clicking through web pages and filling out forms; instead, it wants merchants to proactively open up capabilities such as products, inventory and checkout, so that AI can initiate purchases through a set of standard interfaces. The advantage of this is that merchants remain the primary party responsible for orders and fulfillment, and product display, order processing and risk-control judgments remain in their own hands, but they do not need to rebuild their systems separately for every AI platform. Stripe later launched the Agentic Commerce Suite, making catalog integration, checkout and payment into one-stop tools as much as possible, lowering the threshold for merchants to integrate.
But Stripe also saw another completely different scenario. It is not "AI helps people buy things," but "software pays software directly." Therefore, in March 2026 it launched Machine Payments. In this scenario, AI is not buying a bottle of cola, but buying APIs, data, computing power, content access rights, or even invoking a service on a per-call basis. Stripe's solution is that merchants can turn their interfaces directly into per-call charges, with payments accepted even as low as 0.01 USDC; the money enters a Stripe account and is ultimately settled and reconciled in the way Stripe is familiar with. For an agent, it does not need to first register an account, apply for an API key and then go through a pile of manual processes; instead, it can pay while invoking. In other words, Stripe is building the checkout for AI shopping on one hand, and the micropayment rails between machines on the other.
Coinbase's angle is different again. It is more like assuming that in the future many AIs will not be "shopping assistants" at all, but "software with wallets." Therefore, in February 2026 it launched Agentic Wallets. The core is not to make AI better at browsing websites, but to let AI truly possess a wallet that can spend money, receive money and trade, and this wallet also has security boundaries. In Coinbase's system, one part is the wallet infrastructure itself, another part is a machine payment protocol such as x402, and another part is capabilities it has already packaged, such as top-ups, transfers, trading and yield management. Its key aim is to solve another type of problem: if an AI wants to buy computing power, buy data, pay API fees, or complete some kind of automated transaction on-chain on its own, can it avoid waiting for a human to click confirm every time? Base's official documentation calls this kind of agent an "independent economic actor." This phrasing is somewhat strong, but the meaning is clear: what Coinbase wants to do is not to make AI more like a shopping assistant, but to make AI itself a software entity that can receive and spend, settle and trade.
If these three paths are viewed together, the differences become very clear.
Google wants to first solve "how do you prove that this transaction was really authorized by me."
Stripe cares more about "how merchants connect their products and checkout capabilities to AI."
Coinbase goes one step further and addresses "if the payer is itself software, how does it complete its own receipts and payments."
Precisely because of this, they all appear to be working on AI payments, but their focal points are not the same. Google is more like adding a layer of trust mechanisms for cross-platform agent transactions; Stripe wants to connect both AI shopping and AI-invoked services into the existing commercial network as much as possible; Coinbase is betting more on a world where software pays software directly.
Two stories
At this point, the difference in this matter is very clear.
In China, AI payments feel more like a natural extension of a mature internet ecosystem. Platforms have already organized merchants, payments, delivery, maps and local services; what AI needs to do is rearrange these capabilities so that users shift from 'operating things themselves' to 'just saying what they want.'
Overseas, AI payments feel more like belated catch-up. Because no platform naturally holds the entire chain in its own hands, once AI genuinely seeks to complete transactions on users' behalf, it immediately runs into practical questions: whether websites will allow it, whether merchants will accept it, whether payment institutions will recognize it, and who bears responsibility when something goes wrong.
In China, it feels more like pushing the experience one step forward.
Overseas, it feels more like renegotiating the rules.
Precisely for this reason, the first reaction of Chinese users to AI payments may be 'this feature is quite convenient,' whereas the overseas market's perception of AI payments will feel more like 'the internet's original way of transacting may be about to change.'
Finally
Let us return to the scenario at the very beginning.
'Order another delivery of the milk tea I bought from that shop yesterday.'
What is truly interesting about this sentence is not the word 'pay,' but that it hands over to the system many things you would originally have had to do yourself. It needs to know which shop you mean, which product, what specifications, where to deliver, and which payment method to use. Payment is of course important, but it is merely the least conspicuous step in the whole matter.
So the point of AI payments is not payment itself.
What everyone truly cares about will not be which protocol is used in the backend, whether settlement is on-chain, whether stablecoins are used, or whether cryptographic algorithms are used. In the end, everyone will judge whether it is useful by a very simple standard:
I said one sentence to the AI — can it actually get the thing done for me?
