On September 7, the "China Office Agent User Behavior (Incomplete) Report" was released, based on real usage data from LobsterAI, the office agent product from NetEase Youdao, since its launch in February 2026. The product has accumulated more than 1 million users. This is the first publicly released usage-distribution dataset for office agents in China.
A caveat needs to be stated upfront: all the figures come from a single product's own tracking data. NetEase Youdao is both the data source and the publisher, there has been no third-party verification, and the user base can't be said to represent the whole industry. It reads more like a product's self-portrait than a picture of the national office-agent user base.
The Consumption Curve Is Steeper Than Subscription Models Assume
The hardest number in the report is the compute distribution: the top 20% of users consumed 87.4% of total compute, and the top 5% alone accounted for 53.5% of token consumption. The stickiness gap between paying and free users is just as stark — paying users came in at 6.2 times the token consumption, 5.2 times the task volume, and 3.0 times the monthly active days of free users.
Put those three multiples side by side: active days differ by only 3x, but token consumption differs by 6.2x. Paying users don't show up dramatically more often — each time they do, they do far heavier work. That's exactly where the report lands its conclusion: heavy usage, not visit frequency, is the main driver of paid conversion.
Doing the Math
That curve puts direct pressure on pricing. Assume an office agent charges a flat unlimited monthly fee — a rough calculation based on the 87.4% compute concentration means the heaviest 5% of users have to be subsidized by the remaining 95%, a group that still includes a large share of free users. Several Chinese office-agent products currently run a hybrid of a monthly fee plus credits, and this data roughly explains why a pure flat-rate model is hard to sustain.
One mitigation the report points to is caching: 96.8% of content enters the cache after being read once, and reusing that cache can cut costs by roughly 80%. That ratio isn't surprising for office scenarios — the same contract or project document gets referenced repeatedly within a single task, so the repeated tokens don't need to be billed as fresh input every time. Model choice also keeps unit costs down: the report shows the DeepSeek family accounts for 85.8% of total calls.
Tasks Are Getting Longer
The size of a single task grew 3.1 times over five months, up 53% month-over-month in August; the report also states that average task size roughly doubles every three months. By task type, programming and coding account for 38.5%, while document writing is only 10.1%; by output, web pages make up 69.1% of what's produced versus 8% for documents.
That ratio suggests the name "office agent" is a bit of a misnomer. Most people using it are writing code and building web pages — paperwork is a small slice. The industry breakdown backs this up: software development and IT account for 34%, design and film/video production 13%, internet product operations 9.7%, and finance 9.1%.
Who's Using It, and When
Geographically, the four cities of Beijing, Shanghai, Guangzhou, and Shenzhen together account for 27%, while non-tier-1 cities make up more than 60%. By time of day, the weekday peak falls at 3 p.m., but 60.2% of compute consumption happens outside office hours. Close to 50% of users try out a new model within three days of its release.
The fact that 60% of usage falls outside office hours says more about how the product gets used than the afternoon peak does. Tasks run long, users submit and walk away, then come back later to collect the results — a pattern that's a step removed from instant back-and-forth Q&A.
Sources: "China Office Agent User Behavior (Incomplete) Report", CocoLoop, 36Kr, Ifeng Tech; compute-concentration and multiple figures verified against the report's published numbers, sourced from NetEase Youdao's own product telemetry.