China Office Agent User Behavior Report: After a Task Is Assigned, Who Confirms It Is Done?

Youdao’s September report (October 19, 2026) records 77% WeChat use among its IM users; channels overlap. These are single-platform observations.

That entry point is worth watching. But the adoption of an entry point and the completion of work are two things that need to be verified separately. What follows is independent analysis by Open Market Notes: when a single chat message can start a task, the real change in the office may not be that people make a few fewer mouse clicks. It may be that assigning responsibility becomes less costly, while the cost of confirming that the responsibility has been fulfilled does not necessarily fall at the same rate.

How many conditions does “do this for me” leave out?

Imagine an e-commerce operations employee sending an instruction during their commute: prepare a product listing plan based on the materials for a new product. The agent gathers information, organizes a spreadsheet, writes a description, and even generates accompanying images. This is a hypothetical work scenario, not a user case from the report.

Judged by the number of files, the task has made progress; judged as a business activity, it may not yet be complete. Whether the product model matches, whether the images are licensed for commercial use, whether the price is still valid, and whether the copy claims more than the product can actually do are all part of the deliverable. If the system can also publish directly, the target store and publication time need to be clear as well. The more briefly a person states a request, the more the system needs to decide what it can infer and what it must ask about.

The value of a chat entry point therefore cannot be measured only by how easy it is to give an instruction. It also depends on whether misunderstandings can be caught promptly, whether users can see where a task is stuck, and whether there is an inspectable result at the end. A system that keeps displaying “processing” may feel more reassuring than one that clearly says “product specifications are missing, so I cannot continue yet,” but it is not necessarily more useful.

Youdao's product code and documentation show the ability to connect files, browsers, tools, and chat channels. For organizations, these connections mean that new handoff points need to be designed: who provides the inputs, who approves external actions, and who is responsible for the final version. These are questions of how work is organized; model performance alone cannot resolve them.

Activity is a signal; acceptance is an outcome

When reading any user behavior report, start with a simple question: who is in the denominator? The users of a particular platform, the subset who enable a particular feature, and all the employees of a company are different groups. One person using multiple devices and models may also generate multiple records. A metric expressed as a percentage is not automatically suitable for comparison with others.

Behavioral data is useful for identifying phenomena that deserve further investigation. It can help product teams find entry points, obstacles, and needs, but it cannot, on its own, tell us whether a task saved an organization time. More activity may come from new users or from repeated attempts by existing users; more calls may reflect more complex work or more retries. Making a judgment requires separating these possibilities rather than choosing the most optimistic one.

One workable approach to evaluation in an enterprise is to define the deliverable first. For example, give a manual process and an assisted process the same set of product materials and require each to produce a draft listing that meets the same standards. Compare not just when the first file appears, but also the time spent checking, revising, waiting for additional information, and reworking a draft after it is returned. Wherever possible, quality should be assessed against the same criteria by people who do not know how the work was produced. This does not guarantee a perfect experiment, but it can reduce the risk of mistaking “looks busy” for “actually useful.”

Beyond the compute bill, there is a rework bill

Token usage is a measure of model usage, not a measure of results. Reading long documents, retaining context, and trying multiple approaches may reasonably make a task use more computation; the same growth may also come from unproductive loops. What managers should really care about is the total cost of each acceptable result, rather than whether any one measure of usage rises or falls on its own.

DeepSeek’s caching documentation explains that matching input prefixes can be reused; outputs still require computation. Caching does not verify answers.

Suppose a research assignment requires frequent access to the same set of background materials. Reusing inputs could then be valuable. But if the wrong subject is being researched, the final report may still need to be redone, however cheap each reading is. For tasks with external consequences, a cheap error is not necessarily a low-cost error. An unchecked file sent to a client can also consume time in communication and explanation. Those costs generally do not appear in full on the model bill.

This is why “each call is cheaper” should be discussed alongside “each acceptance check is more reliable.” The former can encourage experimentation; the latter determines whether those attempts can become a dependable part of formal work. Conflating the two makes technical progress carry a business conclusion it has not yet demonstrated.

Finishing the file does not end the need for handoffs

METR's methodology update of February 24, 2026 notes that participant selection and the simultaneous use of multiple agents affect efficiency measurement. This article does not use that update to judge the net benefits of office agents.

Another organizational question worth considering separately is how a file becomes an agreed version. Continuing the earlier hypothetical scenario: the operations employee receives the draft listing, a designer changes the images, and sales revises the product description. Which version is ready to publish? If the system then regenerates the whole set of files from old materials, will the previous human work be overwritten? These questions are related to the quality of a single generated output, but improving that quality alone cannot fully resolve them.

A clearer approach is to distinguish drafts from approved versions and make the basis for changes traceable. For fields that can easily change, such as prices, specifications, and deadlines, record when they were checked and who is responsible for confirming them. An agent can help organize the differences, but “identifying a difference” and “having the authority to choose the final version” are still different permissions. If every participant assumes someone else has already confirmed the content, greater speed may simply allow unconfirmed content to spread outward faster.

These handoff requirements need not turn into complicated forms. A short note may be enough: what changed this time, which materials were used, and what remains unconfirmed. Its value is that the next person can continue the work without having to reconstruct the whole process by guesswork. When a personal assistant becomes part of team collaboration, whether it can leave this context behind becomes part of delivery quality.

What really needs to be established is an orderly delivery process

The appeal of office agents is that they take operational burdens off people's hands. Preserving that value means avoiding a situation in which the operations saved simply turn back into an opaque supervision burden.

A more dependable workflow should make the goal, key supporting information, current status, and final output visible to the user. Actions involving external sending, payments, or overwriting original files need to be handled differently from ordinary organizing and drafting. A retry after a failure should also make clear which steps have already happened, to avoid duplicate sends or changes. These are design principles, not a guarantee of any particular product's existing capabilities.

This report offers an opportunity to observe changes in how people begin office work. For users and companies, the more worthwhile next question is whether assigning a task produces a deliverable that can be checked. The entry point determines whether people are willing to start; clear responsibility, evidence, and acceptance checks determine whether they are willing to rely on it over the long term. Ultimately, judging the maturity of these products means returning to the work itself.

Source note: This article links to NetEase Youdao's original report and draws on product documentation, DeepSeek technical documentation, and METR's methodology update. The scenarios and evaluation recommendations are original analysis by Open Market Notes. Usage on a single platform is not generalized to China as a whole, and this article does not constitute investment advice.