Feishu and DingTalk Exit; the Star of AI Office Is the Personal Agent
Feishu folded into Doubao, DingTalk making way for Qwen Office — the historical mission of enterprise IM is over; the core of AI office shifts from organizational collaboration to personal productivity, and coding capability becomes 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 just merged
Eric breaks it down from the inside: the commercialization team merging into Volcano Engine is essentially just two sales teams merging, because they face the same wave of customers and the same budget, and used to fight each other for orders — merging raises efficiency. The product team wasn't swallowed by Doubao; a Doubao enterprise edition never existed, it's just a Doubao bot added inside Feishu, and the product team was always seconded from Feishu, so now it's only the reporting line returning to Xie Xin. The R&D line hasn't changed at all. From Zhang Yiming's angle, this is using org adjustment to solve internal friction, not Feishu being ‘eliminated’.
— EricOne crayfish compressed big-company planning from years to months
Minghao gives two timelines: first, the demonstration effect of leading vendors — products like OpenAI and Claude Code brought revenue explosions; second, OpenCloud (crayfish) going viral in China during Spring Festival, which made giant-company management clear for the first time what an ‘Agent should look like’, forming a consensus without hesitation. Plans originally made on a half-year or one-year cycle are now compressed to monthly — the Spring Festival plan made in last year's Q4 was obsolete by this March. Across the whole AI office battlefield, from project kickoff to product launch, there was less than two quarters.
— MinghaoTencent had no to-B card, and made a winner anyway
Minghao argues the core is that Tencent had no to-B card, so it was the most anxious, which let an edge team (backgrounds in Yingyongbao, the browser, Tencent Meeting) build Workbody with a tiny team. It is essentially a personal productivity tool, grabbing local data, with no enterprise data moat, hence no barrier — but it happens to match the ‘individuals pay for AI’ trend. The credible DAU figure is a bit over a million, from Yao Shunyu's own statement in a LatePost report. More importantly, in the coding battlefield domestic vendors are being beaten down by overseas products, while the office scenario (non-programmer population) is incremental, and Workbody seized that gap.
— MinghaoFeishu dismisses personal local data as trash others don't want
Eric argues Feishu holds enterprise data, and its intelligent Q&A and ‘open the sky eye’ features can already turn enterprise documents into AI context — Zhang Yiming doesn't need an assistant, he can just search inside Feishu and get the essence of everyone's documents. So Feishu looks down on personal-version local-data tools, seeing them as ‘picking up trash others don't want’. But Minghao counters that this is exactly big-company arrogance: Feishu being folded into Doubao shows Zhang Yiming realized he couldn't bet everything on Doubao, and the boundary between personal productivity and enterprise office is blurring — Feishu bet on the wrong direction.
— EricIt's not who beat whom — Feishu and DingTalk exit together
Minghao proposes that Feishu and DingTalk fought for five or six years, and in the end it wasn't that one beat the other, but that both exited the stage of history one after the other — Feishu entered the Doubao system, and above DingTalk grew Qwen Office. Enterprise IM in the narrow sense has completed its historical mission, because human-machine collaboration in the AI era doesn't need IM as its foundation; but the full enterprise digital AI stack is a new topic, and each company's state differs. Previously neither big company's to-B facade produced independent commercial results, and both ended up as pieces of the group's big AI story.
— MinghaoDoubao's three monetization paths all fail; the user profile is what's valuable
Yanan takes stock of Doubao's C-end monetization attempts: product-jump GMV is only tens of thousands of orders, brand advertising (GEO) caps out at tens of millions and hurts user experience, and C-end paid penetration is low to the point of being ‘outrageous’. All three paths are dead ends, so it can only turn to to-B. But Doubao has a hidden value: it shares an account system with Douyin, and its user profiles are far higher quality than Douyin's, which can improve ad recommendation precision — ByteDance's domestic and overseas ad revenue last year was nearly 90 billion USD, and even a 1% lift is 1 billion USD. That is the below-the-waterline part of Doubao as an asset.
— YananWhen nobody knows the answer, AB testing just wastes time
The discussion notes that ByteDance's proud AB testing and middle-platform mechanisms crumble before AI: multi-team coordination and AB testing waste time, and AI changes too fast — by the time coordination finishes, the market window has passed. Workbody was built fast by a single-person group, then stacked up by continuously adding people later; ByteDance, by contrast, has many teams cobbling together one product. When nobody knows the right answer, many attempts are worse than doing one or two with conviction. Even product naming carries path dependency — single-character brands like ‘Cui’ are hard for Chinese users to read accurately.
Office Agents just hide coding behind the interface
Minghao concludes: general office Agents just hide coding behind the interface — generating websites, handling spreadsheets, editing files are all tool-calling capabilities underneath, and coding is the core constituent of an Agent. Over the past half year and the next half year, coding is the absolute consensus; whether it's model training, inference or post-training, the self-evolution process is built on code capability. This also explains why leading vendors all treat code capability as the most core breakthrough point; another camp advocates the world-model route, but in the short term coding is still the infrastructure.
— MinghaoIn their own words · checked verbatim
He's selling shovels for gold mining — he's still a tool, and tools are still for efficiency.
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
In today's AI battlefield, the time span — the cycle may really 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 fought for five or six years. And in the end, it's not that one beat the other, but that both exited 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 eyes on is bound to die. Because once he pays attention, the moves get distorted.
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 a million billion, annual cost several hundred billion RMB | 5:07 |
| Feishu ARR | about 1 billion USD | 1:01:03 |
| ByteDance ad revenue last year | nearly 90 billion USD combined domestic and overseas | 1:15:38 |
| ByteDance capital expenditure | several tens of billions of USD a year | 1:21:54 |
Glossary
- Agent
- An AI program that can autonomously perceive, decide and execute tasks; an office Agent, for example, can draft documents and handle spreadsheets for you.
- ARR
- Annual recurring revenue — a SaaS company's subscription revenue calculated on an annualized basis, measuring the health of a to-B business model.
- CAPEX
- Capital expenditure — spending on long-term assets such as compute and servers, the main money-burning item for large-model companies.
- GEO
- Generative engine optimization — influencing product recommendations or brand descriptions in large-model answers through paid or technical means.
How to listen
Product managers, investors and industry analysts watching AI office and BAT org restructuring, plus founders who want to understand domestic large-model commercialization.