Feishu and DingTalk Exit; the Personal Agent Takes Over AI Office Work
Feishu is being folded into Doubao and DingTalk is ceding ground to Qwen Office (千问办公); enterprise IM has finished its historical mission. The center of AI office work shifts from organizational collaboration to individual productivity, with coding as the Agent's foundation.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Feishu wasn't killed; two sales teams simply merged
Eric (艾瑞克) takes it apart from the inside. The commercialization team is folding into Volcano Engine, which in substance is nothing more than two sales teams merging: they face the same wave of customers and the same budget, so where they used to fight each other over deals, merging just raises efficiency. The product team is not being swallowed by Doubao either — there was never a Doubao enterprise edition to begin with. It was a Doubao bot added inside Feishu (飞书), and the product people had always been seconded over from Feishu; all that changes now is that the reporting line goes back to Xie Xin (谢欣). The R&D line is unchanged. From Zhang Yiming (张一鸣)'s seat, this is using an org change to end internal friction, not Feishu ‘being killed’.
— EricOne crayfish compressed big-tech planning from years down to months
Minghao (明浩) lays out two timelines. First, the demonstration effect from the leading vendors: products like OpenAI and Claude Code brought an explosion in revenue. Second, over the Spring Festival period OpenCloud — the one everyone calls the crayfish (小龙虾) — blew up in China, and for the first time gave the management of the giants a clear picture of ‘what an Agent is supposed to look like’, a consensus nobody hesitated over. Planning that used to run on half-year and one-year cycles is now counted in months: the Spring Festival contingency plan drawn up in Q4 last year was already out of date by March this year. Across the whole AI-office battlefield, less than two quarters separated project approval from shipping the product.
— MinghaoTencent held no to-B cards, and produced a winner anyway
Minghao's reading is that the core fact is Tencent has no to-B cards, which makes it the most anxious — and that is exactly why a peripheral team, people out of Yingyongbao (应用宝), the browser and Tencent Meeting, built Workbody with an extremely small headcount. Workbody is in essence a personal productivity tool: it grabs local data, it has no enterprise-data moat and therefore no barrier to entry, but it happens to match the trend of individuals paying for AI. The credible DAU figure is a little over a million, from Yao Shunyu (姚舜宇)'s own account in LatePost (晚点). More important, on the coding battlefield Chinese vendors are being beaten down by overseas products, while office work — the non-programmer population — is incremental territory, and Workbody got to that blank space.
— MinghaoFeishu dismisses personal local data as garbage nobody else wanted
Eric argues that Feishu holds enterprise data, and that features like intelligent Q&A and ‘开天眼’ (the god's-eye view) already turn corporate documents into AI context — Zhang Yiming does not need an assistant; he can search inside Feishu and get the essence of everyone's documents. So Feishu looks down on personal-edition local-data tools, treating them as ‘picking up the garbage other people threw away’. Minghao pushes back: this is precisely big-company arrogance. Feishu being folded into Doubao shows Zhang Yiming has realized he cannot stake everything on Doubao, and the boundary between personal productivity and enterprise office work is blurring. Feishu bet on the wrong direction.
— EricNobody beat anybody: Feishu and DingTalk both left the stage
Minghao's point is that Feishu and DingTalk fought for five or six years, and in the end it was not that one defeated the other — both of them walked off the stage of history, one after the other. Feishu goes into the Doubao system; on top of DingTalk, Qwen Office (千问办公) has grown. Enterprise IM in the narrow sense has completed its historical mission, because human-machine collaboration in the AI era does not need IM as its base layer. But the full digital-and-AI stack for enterprises is a new question, and each company stands in a different place on it. Neither giant's to-B storefront ever produced a standalone commercial result; both ended up as pieces in the group's big AI story.
— MinghaoAll three Doubao monetization paths fail; the user profiles are the asset
Yanan (亚楠) goes through Doubao's consumer monetization attempts: product links generating GMV run to only tens of thousands of orders; brand advertising (GEO) has a ceiling of a few tens of millions and damages the user experience; consumer paid penetration is low to a degree they call ‘令人发指’, appalling. None of the three roads works, so the only turn left is to B. But Doubao has one hidden value: its account system connects with Douyin (抖音), and the quality of its user profiles far exceeds Douyin's, which can raise the precision of ad recommendations. ByteDance's advertising revenue last year, domestic and overseas combined, was close to US$90 billion — even a 1% lift is US$1 billion. That is the part of Doubao-as-an-asset that sits below the waterline.
— YananWhen nobody knows the answer, A/B testing only wastes time
The discussion notes that the A/B testing and central-platform mechanisms ByteDance has been proudest of fall apart in front of AI: multi-team coordination and A/B tests waste time, and AI changes so fast that by the time coordination is finished the market window has already passed. Workbody was built quickly by a one-person group and only stacked up by adding people later; ByteDance, by contrast, patches a product together out of many teams. When nobody knows the right answer, running more experiments is worse than building one or two things with conviction. Even the product naming carries path dependency — a single-character brand like Cui (翠) is hard for Chinese users to pronounce correctly.
Office Agents just hide coding behind the interface
Minghao's summary: the general-purpose office Agent merely hides coding behind the interface. Generating websites, handling spreadsheets, editing files — behind all of it is tool-calling ability, and coding is the core constituent of an Agent. Over the past six months and the next six, coding is an absolute consensus: whether it is model training and inference or post-training, the self-evolution loop is built on code ability. That also explains why the leading vendors all treat code ability as the most central point to break through. The other camp argues for the world-model route, but in the short term coding is still the infrastructure.
— MinghaoIn their own words · checked verbatim
It's selling shovels to the people digging for gold — meaning it's still a tool, and a tool is still there to make you more efficient.
他是一个卖挖金子的铲子，等于他还是工具，工具还是用来提效的。
Eric7:44
On the AI battlefield today, that time span — the cycle really may have to be counted in months.
今天在AI这个战场里，这个时间跨度的。周期可能真的得按月算
Minghao16:23
Look at Feishu and DingTalk: they've been fighting for five or six years now. And in the end, you see, it wasn't that one of them beat the other — it was that both of them walked off the stage of history, one after the other.
你看飞叔跟钉钉就是也打了这56年了。然后其实到最后你看不是谁击败了谁，而是两人先后退出历史舞台
Minghao49:29
Any product the boss fixes his eye on is going to die. Because the moment he pays attention, every move gets bent out of shape.
凡是老板盯上了产品，一定会死。就是因为他一旦关注动作就会变形。
Figures
| Doubao DAU | 200 million | 1:02 |
| Doubao daily token consumption | on the order of 100 trillion; annual cost well into the tens of billions of RMB | 5:07 |
| Feishu ARR | about US$1 billion | 1:01:03 |
| ByteDance ad revenue last year | close to US$90 billion, domestic and overseas combined | 1:15:38 |
| ByteDance capital expenditure | tens of billions of US dollars a year | 1:21:54 |
Glossary
- Agent
- An AI program that can perceive, decide and execute tasks on its own — an office Agent can draft documents or work through spreadsheets for you.
- ARR (Annual Recurring Revenue)
- The annualized recurring subscription revenue of a SaaS business, the gauge of how healthy a to-B model is.
- CAPEX (Capital Expenditure)
- What a company spends on long-lived assets such as compute and servers — the main cash burn for large-model companies.
- GEO (Generative Engine Optimization)
- Using payment or technical means to influence how a large model recommends products or describes brands in its answers.
How to listen
Product managers, investors and industry analysts watching AI office work and BAT reorganizations, plus founders trying to understand how Chinese large models get commercialized.