Next Token词元之外
Menu
← All blog articles

Blog / WEEKLY #006

Why Do China’s Personal Agents Keep Turning to Merchant Workflows?

From life assistance to store operations, the deployment of Agents is shaped by data, permissions to act, and measurable outcomes. Merchant scenarios are more concrete, but that does not mean consumer services have disappeared.

Helping someone organize their life sounds like a natural use for an Agent. But when a task involves purchases, appointments, accounts, and after-sales service, a single natural-language request quickly connects to multiple systems. Knowing what to do is only the beginning for a model; it also needs to be able to read information, initiate actions, and confirm the results.

Store operations offer another starting point. Orders, reviews, and marketing campaigns happen every day. Operators know what takes time and can more easily point out whether a report has missed a problem. The discussion of personal assistants in episode 006 eventually turned to merchant tools on platforms such as Meituan. That shift is worth explaining, but it does not justify concluding that only merchant Agents can succeed in China.

Merchant tasks come with some of their own context

When an unfamiliar user says, “Help me plan tomorrow,” the system may not know their work schedule, family responsibilities, budget, or preferences. By comparison, store operations can start with a narrower question: summarize yesterday’s reviews, find recurring complaints, or list order anomalies that the store manager should look at.

A narrower scope does not necessarily make a task technically simple. It does make the business objects easier to identify. Which store, what time period, which data, and who should see it can often be determined from existing workflows. Users can also draw on their experience to judge whether the result is relevant.

Meituan’s official website describes CatPaw as handling data summaries, anomaly detection, and operational recommendations for merchants, while emphasizing that key decisions such as pricing remain with the operator. This is the platform’s description of the product’s intended use, not independently measured data on its benefits. It illustrates a pragmatic boundary: first organize scattered information into a form that operators can assess.

Being close to the business does not mean automatically having permission

A platform that serves many merchants can build up an understanding of business workflows. Whether a particular Agent can use a particular store’s data still depends on actual authorization and integration. Being part of the same ecosystem is no reason to assume that every account, order, and operation can be accessed freely.

You can assess a deployment plan with three questions. Can it see the data needed to complete the task? Does it have an entry point for carrying out the relevant action? Can the result be checked afterward? These questions correspond to information, action, and feedback. Missing any one of them may leave the system at the level of offering suggestions.

Suppose a system can analyze negative reviews but cannot edit product descriptions. It can still serve as an operational assistant, but its deliverable should be a list of items to address; it should not claim to have improved the product page. Describing the scope of its capabilities accurately makes it easier for a person to take over the remaining work.

The same applies to personal assistants. Getting travel recommendations and completing a booking whose status can be checked are different levels of service. Two products with similar model capabilities may have entirely different practical uses because they have access to different interfaces.

Consumer services offer a counterexample

Alibaba’s official introduction to the Qwen App explicitly lists consumer scenarios such as placing an order through Taobao’s instant retail service and completing payment in the conversation. Whether the experience is consistently reliable or not, this shows at least that personal shopping tasks are within the product’s deployment scope. They cannot be summed up as having no useful role.

It is worth noting that these capabilities connect to existing product, transaction, and payment systems. What they share with merchant tools is the ability to hand an intention off to actual business processes. Whether the user is a consumer or an operator does not change this basic requirement.

So the more useful question is how far a service can go, rather than a simple division between China and other countries or between individuals and merchants. A platform can first cover tasks for which it can provide a complete action path. An independent assistant needs to connect to more services through authorized interfaces, partnerships, or interaction with user interfaces. The costs and reliability of these approaches need to be compared for specific tasks.

Business value is clearer, and easier to question

Another appeal of merchant scenarios is that users can often describe the losses caused by a problem: how much time repeated organization takes, which anomalies are often missed, and who will act on a report once it is ready. This gives the product a relatively clear direction for evaluation.

However, a faster report does not automatically mean better business performance. Reducing time spent organizing information, improving decisions, and increasing revenue are outcomes at different levels. Without follow-up records, at most you can confirm the first; you cannot attribute all changes in revenue to the Agent.

Personal assistants face a similar issue, though their value is not necessarily measured in revenue. Convenience, peace of mind, or completing a task that had been put off can all count. Merchant tasks have the advantage that their boundaries and feedback are often clearer; that does not mean people’s personal lives are not worth serving.

When an Agent first enters a store or an e-commerce merchant console, this can be understood as making use of existing business conditions. What ultimately determines how far it can go is still the information it can access, the actions it can perform, and whether the results can get back to the user.