"Please reorder the milk tea I bought from that shop yesterday." If one day you could actually say this to your phone, and it would find the merchant, confirm the product, place the order, pay, and arrange delivery, you probably wouldn't first think of a payment revolution. You'd likely just feel that your phone is finally starting to act as a capable assistant. This might be a more appropriate starting point for our discussion of AI payments today. Most people's understanding of AI payments isn't about giving AI a wallet, nor is it about letting a robot decide how to spend money. What people truly experience is whether a single sentence can directly become a transaction. You state your needs, and the system handles the rest. Therefore, what's truly worth discussing about AI payments isn't often the word "payment," as that word is too abstract and distant. For most internet users, the general feeling is: an app or AI assistant can handle their shopping needs directly when they express them in plain language. If we only look at the Chinese market, AI payment isn't actually that unfamiliar. The reason is simple. Chinese users already live in a very mature mobile internet ecosystem: food delivery, e-commerce, maps, ride-hailing, local services, mobile payment, and instant delivery are all interconnected. Often, what users want isn't a completely new payment method, but rather to avoid having to switch between apps and repeatedly click the same buttons. In China, AI payment is more like an upgrade to the e-commerce shopping experience than a complete reconstruction of the payment system from scratch. Alibaba's upgrade of the Qwen App in January this year is a typical example of this direction. According to the official introduction, Qwen has integrated with services within the Alibaba ecosystem, including Taobao, Taobao Flash Sale, Alipay, Fliggy, and Gaode Maps. Users can order food, complete payments in chat, and plan and book trips with just a sentence. It seems like AI has become smarter, but more accurately, it's the platform's existing capabilities that have been reorganized. This is why AI payment is most easily implemented in China not in grand visions, but in very everyday scenarios. For example, "Buy another box of the printer paper I bought last week and deliver it to my office." Or, "Order another delivery from that restaurant yesterday and have it delivered to my home." Or, "Book a restaurant close to the office, for four people, with a budget under 500 yuan." Users understand these commands immediately. Because the AI here isn't some mysterious new financial tool; it's simply a smarter shopping assistant. What the people need, the endlessly competitive Chinese internet giants will do. Today, there are roughly three paths to AI-powered payments. One is to let AI watch the screen and operate on your behalf, clicking buttons, filling in information, and submitting the process like a human. The advantage of this approach is that it can be tested anywhere, but the disadvantages are also obvious: problems easily arise when the page changes, and you often still need to confirm the key parts yourself. For example, the Doubao phone, which was immediately attacked by WeChat upon its release. Another approach is for the platform to simply hand over its service capabilities directly to AI. For Chinese giants, this path is actually the smoothest. Because payment, merchants, maps, orders, and delivery are already under their control. AI doesn't need to imitate the user clicking around; it can directly access the capabilities within the system. In terms of user experience, it's "I said one thing, and it actually did it for me." For example, Alibaba's Qianwen and ByteDance's Doubao. The third approach is even more difficult, involving cross-platform, cross-website, and cross-company collaborations. At this stage, the issue isn't just about product usability, but whether others are willing to let your AI represent users in this process. Currently, we haven't seen any Chinese internet companies attempting this hard-core approach. Today, China's greatest advantage clearly lies in the first two approaches, especially the second. This is why, for Chinese users, AI payment may not seem groundbreaking; their overall feeling is likely not "AI payment is here," but rather "mobile shopping is more convenient." However, once the perspective shifts overseas, things become somewhat different. While e-commerce, food delivery, maps, and payments also exist overseas, they lack the highly integrated platform structure found in China. Service entry points, merchant systems, payment tools, and fulfillment networks are often controlled by different companies. Therefore, when AI attempts to complete a transaction on behalf of a user, the challenge isn't just "can the order be placed?", but "will others accept you as an AI agent?" This explains why AI payments overseas have become so complex. A prime example is the conflict between Amazon and Perplexity. Perplexity's Comet browser touted AI shopping as a key selling point, stating in its official description that it can help users compare products, read reviews, and proceed all the way to checkout. This aligns with many people's most intuitive imagination of AI shopping: instead of browsing websites one by one, AI will pick and buy for you. However, this is precisely where the problem lies. According to court documents dated March 9, 2026, a core argument Amazon made in its lawsuit was that Perplexity's AI agent accessed user accounts and attempted to perform actions "with the user's consent, but without Amazon's consent." Amazon's reasons for disagreeing are largely the same as WeChat's reasons for disagreeing with Doubao Mobile: authorizing AI doesn't mean the website also authorizes it. Your willingness to let it buy things for you doesn't mean the platform is willing to allow a third-party agent into its system to initiate transactions on your behalf. While the stated reason is for customer privacy and security, the core concern is that AI shopping will directly divert billions of dollars of advertising revenue from its sellers. After all, AI shopping is straightforward; it goes straight for the product and doesn't care about ads. In China, these kinds of conflicts are often resolved by the platform's integrated capabilities; however, overseas, once AI needs to complete transactions across websites, merchants, and payment networks, these issues become unavoidable. Therefore, overseas, the primary question for AI payments isn't "how easy it is to use," but rather "whether this can be completed in a closed-loop end-to-end manner." This is why Google, Stripe, and Coinbase have recently been focusing their efforts on this area. They all see the same problem: in the past, internet transactions assumed that the order was placed by a human, who was sitting in front of the screen, browsing, confirming, and paying. But once the order is placed by an authorized AI, that assumption no longer holds true. How does the website know that this AI was truly sent by them? How can the merchant confirm that it hasn't made any unauthorized purchases? How should payment institutions determine whether or not to release the funds? Google aims to first resolve the fundamental trust issue. Launching AP2 in September 2025, it wasn't about teaching AI how to make payments, but rather about answering the fundamental question: when AI initiates a payment on behalf of a user, what makes the system believe it's genuine? Google's approach is to add a verifiable "authorization certificate" to these transactions. AP2 incorporates two crucial elements: a Cart Mandate, which can be understood as the user's signature confirming the specific purchase; and a Payment Mandate, shown to the payment network and card issuer, informing them that the transaction was initiated by an agent, whether the user was present, and the transaction's context. Google later integrated this system with PayPal, creating a more complete merchant solution: merchants can use their own conversational shopping assistant to serve users, and at the payment stage, PayPal Agent takes over, using AP2 to complete the authorization and payment process. Ultimately, Google's approach addresses the fundamental question of "why trust AI to place orders." Stripe's approach is closer to merchants and platforms. It actually involves two layers. The first layer is the ACP, or Agentic Commerce Protocol, launched in September 2025 with OpenAI. You can think of it as a "standard for AI-readable checkout." Stripe doesn't want AI to mimic human actions like clicking web pages and filling out forms; instead, it wants merchants to proactively open up their capabilities regarding products, inventory, and checkout, allowing AI to initiate purchases using a standardized interface. The advantage of this approach is that merchants remain primarily responsible for orders and fulfillment; product display, order processing, and risk control assessments are still under their control, but they don't need to rebuild their systems for each AI platform individually. Stripe later launched the Agentic Commerce Suite, making catalog access, checkout, and payment a one-stop tool to lower the barrier to entry for merchants. However, Stripe also saw a completely different scenario. It's not "AI helping people buy things," but rather "software paying for software directly." Therefore, in March 2026, it launched Machine Payments. In this scenario, AI doesn't buy a Coke, but rather APIs, data, computing power, content access rights, or even pay-per-use services. Stripe's solution is that merchants can directly convert their interfaces to pay-per-use rates, as low as 0.01 USDC, with the money going into Stripe's account, and final settlement and reconciliation still done in Stripe's familiar way. For agents, there's no need to register an account, apply for an API key, or go through a bunch of manual processes; they can make payments while making inquiries. In other words, Stripe is building both an AI shopping checkout system and a micro-payment track between machines. Coinbase takes a different approach. It's more like assuming that in the future, many AIs won't be "shopping assistants" but rather "software with wallets." Therefore, in February 2026, it launched Agentic Wallets. The core isn't to make AI better at browsing websites, but to give AI a real wallet that can spend, receive, and transact, and this wallet has security boundaries. Coinbase's system includes the wallet infrastructure itself, the x402 machine payment protocol, and pre-packaged capabilities such as deposits, transfers, transactions, and earnings management. Its main focus is on solving a different type of problem: if an AI wants to buy computing power, data, pay API fees, or complete some kind of automated transaction on-chain, can it avoid waiting for confirmation every time? Coinbase's official documentation calls this agent an "independent economic agent." This term is a bit 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 handle income, expenditure, settlement, and transactions. If you look at these three paths together, the differences become clear. Google wants to first solve "how do you prove that this transaction was truly authorized by me?" Stripe is more concerned with "how merchants can connect their products and checkout capabilities to AI." Coinbase, on the other hand, has taken a step further to solve "if the payer is software, how does it complete income and expenditure on its own?" This is precisely why, while they all appear to be working on AI payments, their focuses differ. Google is more like adding a layer of trust to cross-platform intermediary transactions; Stripe aims to integrate AI shopping and AI-driven service calls into existing business networks; Coinbase, on the other hand, is betting on a world where software directly pays for software. At this point, the difference becomes clear. In China, AI payments are more like a natural extension of a mature internet system. Platforms have already organized merchants, payments, delivery, maps, and local services; AI's role is to rearrange these capabilities, allowing users to go from "operating themselves" to "simply speaking." Overseas, AI payments are more like a belated catch-up. Because no platform naturally controls the entire transaction chain, once AI is to actually complete transactions for users, it immediately encounters real-world problems such as whether websites will approve it, whether merchants will accept it, whether payment institutions will recognize it, and who will be responsible if problems arise. In China, it's more like pushing the online shopping experience a step further. Overseas, it's more like renegotiating the rules for participation by multiple parties in internet software and online consumption scenarios. This is why Chinese users' first reaction to AI payments might be, "This function is quite convenient"; while overseas markets' perception of AI payments is more like, "The original online transaction methods may need to be changed." Returning to the initial scenario... "Please place another order for the milk tea I bought yesterday and deliver it to my door." The truly interesting part of this sentence isn't the word "payment," but rather how it delegates many tasks that you would normally have to do manually to the system. It needs to know which store you mentioned, which item, its specifications, delivery location, and payment method. Payment is important, of course, but it's just the least significant step in the whole process. Therefore, the focus of AI payment isn't the payment itself. What people really care about isn't which protocol the backend uses, whether it's on-chain settlement, whether it uses stablecoins, or whether it uses encryption algorithms. Ultimately, people will use a very simple standard to judge its usefulness: I told the AI something, can it actually get it done for me?
Miyuki
Weiliang
Alex
Catherine
Kikyo
Joy