In May 2026, Alipay announced that its AI payment transactions had exceeded 300 million. A month later, WeChat opened up Mini Program AI access to developers, with one controversial requirement: developers must authorize the platform to read their Mini Program source code.
The two events were less than 30 days apart, yet they represent two strategic paths that had diverged over a year earlier. According to a report by LatePost, Alipay is internally testing an AI version codenamed 'Project Bao' — not just adding an assistant, but allowing users to switch with one click to an entirely new, conversation-driven interface. Meanwhile, at an earnings conference call, WeChat's President, Liu Zhiping, set the direction: an AI agent will eventually be integrated, but it will be deeply connected with social relationships, Official Accounts, and Channels, with no separate timeline.
Two platforms, each holding hundreds of millions of users and millions of Mini Programs, have given opposite answers to the same question: When AI can operate services for users, should the entry point be completely rewritten or should it be hidden?
What Alipay Is Cutting Is More Than Just an Interface
To understand what Alipay is actually doing, we need to look at a specific user action.
In the past, ordering three lightly sugared lattes and then booking a ride to the airport within Alipay involved a standard process: find the Didi Mini Program entry, input the destination, confirm the ride; exit, find the Luckin Coffee Mini Program entry, select the product, adjust sugar level, add to cart, checkout; switching back and forth between the two Mini Programs to complete payments. Every step was a click, a page jump, and a wait.
What 'Project Bao' aims to do is compress this entire process into a single sentence. The user speaks into a dialog box: 'Help me book a ride to the airport and order three lightly sugared lattes nearby,' and the AI takes over all subsequent steps: understanding intent, breaking down tasks, calling the corresponding travel and food services, combining orders, and completing payment. The interaction interface is no longer a row of Mini Program icons but a chat window.
The radical nature of this change is evident from the internal product design process. According to LatePost, to determine the new interaction form, the project team produced over 100 different product design versions. The final choice of a conversation-centric solution is based on the judgment that natural language has become the mainstream method for AI interaction, and service distribution should rebuild the entry point along this direction, rather than patching AI onto the old framework.
This radical approach wasn't Alipay's initial choice. In the second half of 2023, when the Alipay business unit management initiated discussions on 'how to move towards intelligence,' the first question was: Should they modify the existing app or create a new one from scratch? The project team initially chose the latter. At the Bund Conference in September 2024, Alipay launched the standalone AI application 'Zhi Xiaobao,' positioned as an AI life assistant.
Zhi Xiaobao didn't succeed. According to informed sources, the daily active users of the standalone app were far lower than the in-app AI assistant. That assistant, remaining within Alipay and leveraging the homepage traffic, maintained a stable DAU in the millions, accumulating much more interaction data than the standalone app.
There was also a more practical constraint: At the time, Ant Group was focusing resources on the health application 'Ant Afu,' and the general-purpose AI 'Ling Guang' was also in development, leaving limited computing power and development resources. Creating another standalone app would not only compete for resources with these projects but also bear the huge cost of migrating users from zero.
In March 2025, the team changed course, abandoning the standalone app route. An internal judgment gradually formed: serving Alipay's existing 1 billion user base, allowing them to access AI services with zero migration cost, is more effective than building a new application from scratch outside the ecosystem. In December 2025, the Alipay AI Version project was formally established, with the initial team coming from the in-app intelligent assistant team, later joined by algorithm, C-end product, and Mini Program business teams.
The final product strategy is neither a standalone native app nor an embedded assistant within the existing app, but a one-click switch. After the new version goes live, the original Alipay opens by default, but users can set the AI version as their preferred interface. LatePost reported that this 'cautious' rollout method points to an internal concept called 'clearing the cage for new birds.'
WeChat Prevents AI from Getting Between People
WeChat's AI path has followed a different logic from the start.
Tencent President Liu Zhiping's statement during the Q3 2025 earnings call was almost unambiguous: The AI agent WeChat will launch will be deeply connected with social relationships, communication capabilities, Official Accounts, and Channels; it is a unique Agent. There's no aggressive timeline, and the company has twice publicly refuted rumors about an AI assistant.
Why can't WeChat just cut out a conversational interface like Alipay? The reason lies not in technical capability but in product nature. WeChat's core interface is the chat list, the most frequently opened page on billions of phones daily. Any attempt to superimpose an AI conversation entry on this interface risks being perceived by users as interference with social relationships. Alipay's homepage is a service entry; turning it into a chat window requires users to adapt to a new operating habit. WeChat's homepage is conversations between people; replacing or crowding out human conversations with AI conversations touches upon the user's most important psychological territory.
WeChat's approach is closer to a 'parasitic' logic. The AI assistant doesn't replace any interface; it hides within group chats and Official Accounts, waiting as an Agent to be invoked. Imagine this scenario: In a family WeChat group, someone shares a long Official Account article about parent-child camping sites. Other members don't need to open and read it; they simply ask the AI assistant within the group to summarize the key points and coordinate the group members' calendars to book the recommended site. The Agent digests the article's content, calls the booking service within a Mini Program, coordinates times based on multiple members' schedule information from the group chat, and finally pushes the booking result back to the group.
Throughout this process, the AI always operates within the context of the group chat; users still see this group, these people, these conversations. The Agent's 'task completion' is embedded within social relationships rather than popping up in a separate interface to showcase its own presence.
This restraint comes at a cost. In WeChat, services exist on the platform in the form of Mini Programs, numbering in the millions. For AI to complete these tasks for users, it needs to understand not only user intent but also the data structures, page logic, and interaction flows of these services themselves. Alipay faces the same problem, and their solutions have led to the most fundamental divergence on this track.
Screen Reading vs. Source Code Reading: Which Solution Is Harder?
In June 2026, the WeChat Open Community released the 'Mini Program AI Development Model (Beta) Access Guide,' outlining two modes.
The first is 'Automatic Mode.' Developers authorize the platform to read their Mini Program source code during the review process. The AI understands page structure and operation logic by analyzing the source code and directly controls the Mini Program. The second is 'Development Mode,' where developers encapsulate their services into Skills according to WeChat-defined protocols, containing atomic interfaces and atomic components. The AI completes tasks by calling these standardized interfaces.
Alipay's solution is a 'dual-track' approach. According to LatePost, on one hand, it encourages willing merchants to actively connect, turning their services into MCPs or Skills that the AI can directly call. On the other hand, with user authorization, the AI 'reads the screen' of existing Mini Program interfaces to operate services that haven't been adapted yet.
Comparing the two, the core difference is: when handling legacy Mini Programs not yet adapted, WeChat requires developers to hand over source code, while Alipay chooses to have the AI perform screen-reading operations for the user.
Judging from the WeChat Open Community documentation, the 'Automatic Mode' is a more thorough technical solution. After reading the source code, the AI's understanding of the page is structured, operation paths are clear and controllable, and it doesn't rely on visual recognition and interface simulation like screen reading, leading to lower error probability. However, this approach transfers pressure to developers. Source code is a Mini Program developer's core asset. Handing it over means fully exposing business logic, data structures, and interaction design to Tencent. For small and medium-sized merchants relying on Mini Programs for business, this isn't just a security concern but a commercial risk: once the service flow is completely mastered by the platform, how much room is left for traffic distribution and bargaining power?
If not choosing 'Automatic Mode,' the Development Mode is also not easy. Developers need to re-organize business processes, split them into atomic capabilities, encapsulate them into Skills according to WeChat-defined protocols, and go through a new review process. The workload of decomposing and encapsulating an entire process for a food ordering Mini Program — ordering, payment, coupon redemption, membership points — could be a significant fraction of the initial development cost. Who bears this cost? WeChat hasn't provided an incentive plan, at least not yet.
Alipay's screen-reading solution bypasses these issues. It doesn't require merchant cooperation, code changes, or even for merchants to know their Mini Program is being operated by AI. The user says to the dialog interface, 'Help me buy a train ticket to Shanghai,' and the AI opens the 12306 Mini Program interface, recognizing departure city, destination, train list, seat selection buttons, payment confirmation page, step-by-step simulating user finger taps. For merchants who have completed MCP or Skill integration, the AI can directly call standardized interfaces for a smoother experience; for the vast number of long-tail services not yet adapted, screen reading provides the lowest-barrier compatibility path.
The problem with screen reading is straightforward: its stability hasn't been validated at scale. Mini Program interfaces vary widely. Dynamic loading, pop-up ads, layout changes due to version updates can all increase the probability of AI recognition failure. If a payment confirmation button shifts a few pixels, can the AI accurately tap it? If misoperations occur during screen reading — misreading amounts, selecting wrong delivery addresses — who is responsible? Alipay hasn't publicly disclosed related disclaimer clauses or dispute resolution mechanisms.
The logic of this path is to get users to start using it first. When merchants see the order conversion brought by AI, they will naturally take the initiative to connect to standard interfaces to optimize the experience. The C-end pushes the B-end.
What Do 300 Million Transactions Verify?
Beyond products and ecosystems, Alipay has done something else related to how AI pays.
At the AI Payment Ecosystem Conference in May 2026, Alipay disclosed that AI payment transactions had exceeded 300 million, supporting 95% of general-purpose agent frameworks, and simultaneously released Token Pay and AI Wallet. These two products are key to understanding the infrastructure of the Agent economy.
Token Pay solves the problem of extremely small-amount, high-frequency payments. When AI is comparing prices between two food delivery platforms, it might need to initiate a 0.01 yuan verification transaction to confirm account validity; when AI is filtering for the best combination among multiple coupons, verifying each coupon is a payment action. These transaction amounts are tiny, but the frequency is far higher than human users. The previous payment system was designed for 'human confirmation, human payment.' Token Pay hands this action over to the Agent.
The AI Wallet is more like giving the Agent a budget card. Users set rules and limits, and the AI completes payments autonomously within those rules. At the conference, Ant Group CEO Han Xinyi offered a judgment: In the future, countless Agents may be active in economic activities, with interactions shifting from human-to-human to human-to-Agent and Agent-to-Agent.
The absolute number of 300 million isn't huge compared to Alipay's overall annual transaction volume, but its significance lies in verifying one thing: users are already allowing AI to complete real commercial fulfillment on their behalf, not just stopping at queries and price comparisons. From ordering a ride and coffee with a sentence to AI payment and deduction, this entire service loop's technical and user authorization chains have been connected.
WeChat Pay hasn't publicly disclosed specific plans regarding AI transformation on its side. WeChat Pay also covers a massive user base, but its scenarios are more tied to social transfers, red envelopes, and merchant payments. The form of the Agent economy might differ. Whether new differences will emerge on the payment infrastructure front depends on whether WeChat launches similar Agent payment capabilities alongside its formal AI assistant release.
The Ecosystem Is Being Torn Along Two Seams
Both Alipay and WeChat are pointing towards Agent service entry, but their differing intermediate paths will tear the Mini Program ecosystem along two diverging seams.
Alipay's screen-reading solution passively AI-fies a large number of long-tail Mini Programs. Merchants do nothing, yet users can already operate their services via AI. This will elicit two reactions: some merchants, seeing AI-driven order volumes rise, will actively connect MCPs or Skills to optimize experience and compete for more traffic distribution; others may resist because the order source becomes blurred. Previously, every user click within a Mini Program was trackable. Now, for the segment where AI performs screen reading, merchants can't access user behavior data.
The Alipay team clearly anticipated this. According to LatePost, following the launch of the AI version Alipay, an AI open platform for merchants and developers will be released soon. This platform will likely need to solve: How can merchants both enjoy the order growth brought by AI and retain visibility and control over service processes, user reach, and revenue distribution?
The pressure on WeChat's side is different. The threshold of source code authorization will screen developers into two groups. Top-tier developers with technical teams and commercial bargaining power may be willing to hand over source code or invest resources in packaging Skills in exchange for prioritized traffic distribution from WeChat's AI assistant. But a large number of small and medium-sized merchants may be unwilling to hand over source code and unable to bear the packaging costs. If traffic indeed skews towards authorized merchants after the WeChat AI assistant launches, unauthorized Mini Programs may be marginalized in the AI service distribution channel. Over time, WeChat's Mini Program ecosystem could further concentrate towards the top, creating tension with WeChat's long-standing emphasis on a 'decentralized' ecosystem narrative.
A more subtle issue lies in technical standards. Alipay promotes MCPs, while WeChat has defined its own set of Mini Program MCP protocols. Although the name is the same, the specific implementations are not entirely interoperable. A restaurant merchant wanting both Alipay AI and WeChat AI to call its ordering service might need to package it according to two separate specifications. This isn't an insurmountable technical challenge, but it is a cost. Whichever side achieves scale advantage first will have greater bargaining power to push industry de facto standards. At the point where Alipay's AI payments have exceeded 300 million, this advantage temporarily lies with Alipay.
Returning to the user side, the ultimate outcome of this transformation may redefine the relationship between people and their phones. If Alipay's conversational interface succeeds, the frequency and scenarios for users opening Alipay will change. It won't just be opened when paying, but whenever there's a need, asked with a casual sentence. If WeChat's Agent succeeds, the way users accomplish tasks within group chats will change. There's no need to jump out of the chat interface to find services; everything is done through the Agent within the group chat.
The 'Red Envelope War' between the two platforms on the eve of the 2014 Spring Festival changed where users kept their money. This time, the battle is over who users will entrust with the phrase 'help me do this.'
Twelve years ago, WeChat Red Packets were called a 'Pearl Harbor attack' by Jack Ma. Twelve years later, as rumors about WeChat AI swirled for months, Alipay has stepped onto the stage first. Which of the two paths is closer to the real needs of the Agent era? The answer won't be found in product launch events, but in how millions of Mini Programs are reactivated, and in the experience of billions of users after they first say 'help me' to their phones.
