On June 8th, in the afternoon, WeChat suddenly pushed a message to me. At first, I thought it was just another routine notification from the WeChat Open Platform. Our own New Distribution mini program (FMCG Insider) has been running for three years. Over the years, there have been too many feature upgrades, but to be honest, few have made me feel that 'the business logic has changed.' But this time, I stared at that notification for a long time. Because I quickly realized that it is directly relevant to everyone in FMCG. WeChat released guidelines for developers to integrate WeChat AI. In plain terms: you can connect the service capabilities in your mini program into WeChat AI. Once connected, the mini program is no longer just a page waiting for people to click; it becomes a service that WeChat AI can understand, invoke, and execute. Many people will treat this as a piece of tech news. My judgment is exactly the opposite. It doesn't change a feature; it changes something more fundamental: in the future, users won't need to enter a mini program themselves to find a service. You just say a sentence—'Is there a nearby drink that can be delivered today?' or 'Order my usual coffee'—and the AI understands, matches, invokes, and delivers the result to you. This is no longer people finding services. This is AI invoking services. FMCG companies, whether AI recognizes and can invoke your brand, your channels, and your local services, may become the next hurdle for your business. Next, I'll explain my understanding step by step. To understand the significance of this, we first need to talk about the concept of 'entry points.' Essentially, all products and services compete for entry points. In the early days, the competition was about 'being searchable.' So companies focused on official websites, keywords, and search rankings. If a brand or service provider disappeared from search results, it was basically nonexistent on the internet. Later, entry points became platform recommendations. Taobao, JD.com, Meituan, Douyin, Xiaohongshu, and Video Accounts each have their own distribution logic. You not only need to be found by search but also recommended by algorithms. Then came mini programs. They solved a very practical problem: you don't need to download an app to access a wide range of services within WeChat. Brands build membership, retailers build malls, distributors take orders, and stores do community conversion. But mini programs have had a persistent flaw. You have to make users know your name, remember your entry point, be willing to enter your page, and have the patience to operate step by step. For high-frequency, essential needs, it's fine, but for low-frequency, complex, or impulsive decisions, this path is too long. So in recent years, many companies' mini programs have become another kind of 'page engineering': more and more pages, more frequent activities, but actual usage is not high. Now, this logic is starting to loosen. Because AI is not just answering questions; it's starting to do things. This is not just WeChat's move. On January 15th this year, Alibaba's Qianwen connected to Taobao, Alipay, Taobao Flash Purchase, Fliggy, and Amap, launching over 400 'task-execution' functions at once, linking demand expression, task execution, and payment fulfillment into one line. Within two months, its C-end monthly active users exceeded 100 million. ByteDance took a more direct approach: partnering with ZTE Nubia to create a 'Doubao phone,' giving AI system-level permissions, using screen reading and simulated clicks to operate the screen for you. All the big players have realized the same thing: AI that only chats is worth little; only when it can invoke services and execute tasks can it drive transactions. WeChat's uniqueness lies in the fact that it is not a single-point application. It has mini programs, WeChat Pay, Official Accounts, Video Accounts, WeCom, communities, merchants, user relationships, and a large number of local services and transaction scenarios that have been running for years. As of the end of March this year, the combined monthly active users of WeChat and WeChat (international version) were 1.432 billion. If an AI that can do things starts appearing in the WeChat of over a billion people, it won't just change a feature button; it will change the organization of this service capability. In the past, it was the page era: users entered pages and then found functions. In the future, it may be the invocation era: users state their needs, and AI matches capabilities. The difference is not just a technical term; it's a whole set of business logic. In the past, companies needed to 'make users see me.' In the future, they must also answer a new question: Can AI understand me? Can it invoke me? After invocation, can I deliver reliably? What exactly did WeChat do this time? If you only look at the technical documentation, most people in the FMCG industry won't understand it. Terms like automatic mode, development mode, atomic interfaces, SKILL, and mini program MCP are all new. Translated into business language, it's actually simple: WeChat is turning each mini program from a 'collection of pages' into a 'collection of capabilities.' In the past, an e-commerce mini program had a homepage, category pages, product details, shopping cart, orders, and membership. Users clicked in themselves, searched, filtered, and placed orders. But in the logic of AI invocation, the mini program must tell WeChat AI: I can search products, check inventory, generate orders, track logistics, issue coupons, and display membership benefits. This is no longer pages. This is capabilities. WeChat offers two ways to connect.
- One is automatic mode: after the developer authorizes, the platform reads your source code during review, automatically analyzes the pages, and lets AI directly operate your mini program—this is practical for small and medium developers who can't rebuild a set of interfaces;
- The other is development mode: developers proactively break their business into smaller units (so-called atomic interfaces and atomic components), package them as SKILLs, and hand them over to AI for invocation. The two are not mutually exclusive and can be enabled together. The term SKILL is a concept in technical documentation. In business, I prefer to call it a 'service manual.' What exactly can you do? What conditions are required? What results are returned? Which actions require user confirmation? Which scenarios require login, authorization, payment, or manual intervention? In the past, many companies hid these things in pages, in processes, in customer service scripts, or even in the minds of some veteran salespeople. If AI wants to invoke you, you can no longer be vague. You must clearly articulate your capabilities. Digging one layer deeper, you'll see something more interesting. Currently, big companies are taking several technical paths to let AI take over services. One path is 'screen reading plus simulated clicks': AI directly looks at the screen, recognizes elements, and simulates finger taps, without requiring cooperation or authorization from the other party. ByteDance's Doubao follows this path. It's like climbing in through a window—you can get in and do things, but it's fragile. Another path is 'interface invocation': the application proactively declares which capabilities can be invoked, and AI follows the interface, with authorization, boundaries, and records. Google's AppFunctions, released in February this year, is a typical example; it calls this 'MCP on mobile.' It's like entering through the front door, presenting credentials, and doing things in the open. WeChat's two modes this time happen to bet on both paths: automatic mode lets the platform read your source code and parse pages for AI, leaning toward the 'screen reading' end; development mode lets you package capabilities as SKILLs and be invoked via interfaces, leaning toward the 'interface' end. But what it really wants is the latter. It hopes you proactively package your capabilities clearly, so AI can come through the front door to invoke you, rather than relying on the platform to guess and read your pages. And whether you can 'enter through the front door' depends on one thing: whether your service has been standardized. So on the surface, this is a technical integration, but at the bottom, it's the standardization of enterprise services. WeChat AI won't do your business for you; it just connects demand and capability more directly. What truly determines whether you can be invoked is whether your service is clear, stable, and executable. This is somewhat similar to Video Accounts back then. WeChat didn't produce content itself; it opened up the infrastructure and let creators grow on their own. Today, it's the same: it didn't announce an all-powerful agent; it first gave developers a framework: are you willing to open up your capabilities for AI to invoke? This is very WeChat. It doesn't define the answer first; it paves the road. Once the road is paved, whoever can run fast depends on their own abilities. Why is this relevant to FMCG? AI is still a tool in the office, writing weekly reports, making proposals, and organizing meeting minutes. These are valuable, but if AI stays only in the office, its transformation of FMCG is limited. The truly complex parts of FMCG are never in the office; they are in channels, terminals, inventory, fulfillment, and reach. This time, mini programs connecting to AI opens another possibility: AI not only helps improve efficiency, but it also moves to the front line of business—transactions. Under this new logic, the first to reassess themselves are distributors. In the past, the core capability of distributors was to get goods to terminals, take orders, and deliver goods—in short, they were the local market fulfillment organizations. In the invocation era, they may need to become another role: local service interfaces. For example. In the past, when terminals served by distributors needed restocking, they relied on salespeople visiting, making phone calls, or clicking item by item on an ordering mini program. In the future, these actions may all be condensed into a single sentence in WeChat—a small shop owner says to WeChat, 'Restock the items that are about to run out this week, and include any promotions,' and the demand goes directly to the distributor. But whether you can truly be invoked this way doesn't depend on whether you've connected to AI, but on whether your inventory, prices, payment terms, and promotions have become data that systems can read. Today, most distributors' data is still scattered in WeChat groups, Excel spreadsheets, salespeople's memories, and bosses' minds. This is also the judgment I've always made: for FMCG companies to use AI, the starting point is not buying tools, but rebuilding the data foundation, process mechanisms, and assessment mechanisms.
Without a data foundation, AI doesn't know what products you have;
Without process mechanisms, there's no one to receive the demand AI brings;
Without assessment mechanisms, salespeople won't maintain data seriously. So WeChat AI is not an easy traffic dividend for distributors; it's more like a mirror, reflecting whether your local service capabilities have been systematized. Beyond distributors, brands and stores can't escape either. The difference is that distributors receive orders from small shops; brands and stores receive the words spoken by consumers themselves. In the past, brands competed for consumer mindshare. A slogan, a package, a promotion—they fought for a place in the consumer's mind; then through shelf placement and salesperson recommendations, they turned that place into a final purchase. But in the invocation era, users may not first think of your brand; they might first ask AI: 'I want to buy a box of room-temperature yogurt for my child. Is there anything suitable nearby?' At this point, AI needs to judge for them: whether it's available nearby, whether you have promotions, whether membership coupons are available, which size is suitable for family stocking, and how soon it can be delivered. Behind this is not a slogan or a single campaign; it requires turning SKUs, prices, inventory, promotions, membership benefits, store coverage, and fulfillment methods into structured data that AI can read. In the past, brands competed to be remembered by users; in the future, they must also compete to be understood by AI. The store side is the same. In the past, a store's value was location, shelves, salespeople, customer relationships, and immediate availability—customers walked in, saw the goods, and bought. But if users become more accustomed to asking AI first and then letting AI recommend nearby services, the 'walking in' action may be skipped. Customers may not enter the store first; they might ask, 'Which nearby store has ice-cold beer and can deliver within half an hour?' At this point, whether a store gets recommended depends not only on location but also on inventory accuracy, price clarity, fulfillment speed, and review stability. Stores are not just there; they must be dispatchable. But I don't want to overstate this. Currently, it's still in beta: the developer-side access point has just been launched, and there's no official timeline for full rollout or how it will perform. After all, this is just the beginning. How far WeChat AI can go, how many people will use it, and whether it can truly work—there are no answers today. It doesn't offer a wave of dividends; it's more like a direction. But the direction is clear enough: if 'invocation' truly takes hold, it won't distribute opportunities evenly. Those invoked will likely not be the loudest shouters or the flashiest pages, but those companies whose services are clearly described, data is connected, and fulfillment can withstand the load. This may be especially difficult for FMCG. In the past, FMCG excelled at distribution, display, promotion, and customer relationships; but AI only recognizes services it can read—whether you have product data, real-time inventory, executable fulfillment nodes, and whether there are people willing to maintain that data continuously. These things that used to be backstage may become front-end entry points in the future. If you can't explain it clearly, you can't be invoked. If you can't be invoked, you have no opportunity. So what's truly worth doing today is not rushing to connect a new feature, but taking the time to sort out your services: which ones can be triggered by a single sentence, which can be broken into interfaces, and which can truly handle demand. The page era is not over yet. But the invocation era is already knocking on the door.