In 2016, Allen Zhang (Zhang Xiaolong) drew the line: WeChat would not build a mini-program store, no categorization, no rankings, no recommendations (21st Century Business Herald). For the next decade, this decentralization logic became the cornerstone of WeChat’s governance. But AI instructions are inherently about distribution. When a user tells Xiaowei “order me a milk tea,” it has to choose between Meituan and JD, between different price points. That choice can come from the user’s historical preferences or from the platform’s ranking weights—technically, the two are nearly indistinguishable. Tencent did not solve this contradiction. They hid it inside five layers of constraint.
On June 20, WeChat rolled out its native AI assistant, Xiaowei, in a gray release. Top-left corner, a green eye; tap it and you enter a chat interface labeled “Beta.” Users can ask it in natural language to send messages, set reminders, summarize PDFs, or trigger Meituan to search for milk tea. The capability drawing the most attention is one-line mini-program generation. Say “build me an exercise check-in tool,” and a few seconds later you get a lightweight page with training logs, history review, and duration stats (21st Century Business Herald hands-on). This generative capability shipped with five layers of constraint. Each layer can technically be loosened. WeChat loosened none of them.
The first layer: the generated mini-program is for personal use only and cannot be shared. In hands-on testing, reporters from 21st Century Business Herald found that after generation there is no share button, and Xiaowei explicitly states that external sharing is not supported.
The second layer: the output is not a full mini-program, only a pure front-end page running locally inside the WeChat client. No network access, no backend, no data persistence. When the reporter asked to add a calorie-tracking feature to an existing exercise check-in tool, Xiaowei responded that it does not yet support additions to existing tools—you can only regenerate an integrated version from scratch.
The third layer: no payments. Xiaowei can call up Meituan or JD to search for products, but it only returns product listings; purchase and payment still require the user to manually jump into the mini-program to confirm. On June 17, WeChat Pay launched an AI-exclusive card with three lines of defense: isolated from the primary account, balance at the user’s discretion, and confirmation on every transaction. But this card currently connects only to WorkBuddy, not Xiaowei (Tencent News). Alipay’s Abao took the opposite route: no isolated account; AI operates the primary account directly, relying on manual confirmation plus a “you dare to pay, I dare to compensate” safety net (Sohu). Among WeChat’s 1.4 billion monthly active users, a large share are financially sensitive. Batch abnormal transactions, once they occur, can trigger systemic public-opinion fallout. What the isolated account does is contain the blast radius.
The fourth layer: the main model is WeChat’s in-house WeLM; only complex reasoning falls back to DeepSeek; the group-level Hunyuan is not used. WeLM handles everyday conversation and semantic understanding for most service calls.
The fifth layer: the entry point is placed in the top-left corner, avoiding the two most prominent positions—the top of the chat list and the search box. It is there, but it does not seize the social interface.
These five layers together point to the same judgment: Tencent is deliberately locking AI on the personal-agent side, preventing it from sliding toward platform distribution.
Ant Group’s Lingguang took the opposite road. Launched in November 2025, it generates a “flash app” in 30 seconds. In April 2026, Ant announced a 100-million-yuan incentive fund to motivate creators, encouraging users to publish their generated flash apps to Lingguang Quan (Lingguang’s sharing circle), where works can collect likes and be remixed. As of April, the platform hosted over 30 million flash apps (People’s Daily Online).
Lingguang’s logic is a UGC app marketplace: creator incentives drive volume; social distribution filters quality. WeChat refused this path. It needs to protect the existing interests of 8.4 million mini-program developers, and it worries that shareable AI-generated apps would turn into a new content garbage dump and a moderation black hole. WeChat mini-program operating rules require any mini-program containing user-generated content to integrate content-safety moderation (WeChat Open Docs), and AI-generated tools are in essence a new form of UGC: high volume, low barrier, manual pre-review nearly impossible. The moment sharing is allowed, vast amounts of unmoderated content would flood the social graph. The developer guidelines for AI ecosystem integration, released on June 8 this year, are still in closed beta, with integration left to each developer’s discretion; the governance framework is far from formed (Sina Finance).
After hands-on testing, TMTPost categorized Xiaowei as a single-user Agent, with a service radius that essentially closes at fulfilling one person’s needs (TMTPost). It found that Xiaowei can help you draft a Moments post, but when you tell it to publish directly to Moments, it refuses. The subject of social expression is always the human; AI does not act on their behalf.
Lingguang’s route carries its own risks. Low-quality templates, engagement-farming creators, homogeneous junk apps—these problems are already surfacing against a base of 30 million flash apps. WeChat’s refusal to go the store route is, in part, dodging the same trap.
The difference is that Lingguang chose to open first and govern later, betting that the ecosystem would find its own order in the chaos. WeChat chose to govern first and open later, removing uncertainty from the public communication domain up front. Each strategy carries its own cost, and there is no clear winner yet.
AI routing itself does not equal centralized distribution. WeChat could design the candidate set so that everything comes from the user’s history and authorized services, with AI responsible only for matching intent to existing relationships. This path is technically sound, and it is the direction the five layers are converging toward. But there is one contradiction it cannot sidestep: the moment AI begins to handle real transactions, it has to rank among multiple candidates. Once ranking involves prioritizing some services and demoting others, distribution power re-enters the room. That power wears the disguise of natural language understanding, making it harder for users to notice than a leaderboard.
Tencent’s current five-layer constraint is a locally optimal gray-release strategy at the governance level. It blocks AI garbage from eroding the social graph, and it confines batch abnormal-transaction risk to a local blast radius. Before moderation mechanisms and payment risk controls are battle-tested, this constraint is rational.
But Xiaowei’s generative capability stops at the pure front-end local-tool layer—it can build check-in logs, mood diaries, to-do lists. The things that genuinely need AI to handle—placing orders, booking appointments, processing refunds, these multi-step transactions—live in the transaction layer, which Xiaowei cannot reach.
The transaction layer is currently handed to WorkBuddy. This is Tencent’s desktop-side office-productivity agent, launched in March, integrated with the AI-exclusive card in June. It runs inside enterprise scenarios, with organizational authorization, audit chains, and clear accountable parties. Meituan’s life assistant uses WorkBuddy’s expert mechanism to order takeout and buy group-buy vouchers for users; payment goes through the exclusive card, with confirmation on every transaction.
This division of labor is a transitional solution. It relocates the sharpest question—how to rank when AI handles real transactions—into the B-side scenario, and bypasses it on the C side.
The discussion about opening up sharing can wait: AI content moderation and payment risk controls are not yet ready, and opening up now means throwing a system whose governance framework is unfinished into the social network of 1.4 billion users. The more urgent question is transparency in intent routing. When Xiaowei calls up Meituan instead of JD to search for milk tea, when it returns results for one brand instead of another, users do not know how that choice was made, and developers do not know under what conditions their service will be prioritized. WeChat can continue to refuse leaderboards, but it cannot leave developers guessing inside a black box, nor leave users unaware of what decisions AI made on their behalf.
If this layer is not done well, Xiaowei will fail on both sides. Users will find it cannot get things done—every step requires manual confirmation, so they might as well open the mini-program directly. Developers will feel that WeChat is using AI to redistribute traffic through a new black box, with the decade-old promise of no rankings and no recommendations quietly circumvented at a technological bend. The trust on both sides is two faces of the same thing. Tencent got the five-layer constraint right. What comes next is the explanation layer.