Tencent WorkBuddy, ByteDance Doubao, and Alibaba Qwen's Digital Employee Battle: The Victory Lies in the "Foundation," Not the "Floors"

11/20 2026 582

Produced by | He Xi

Layout | Ye Yuan

On August 14, Qwen Office added Zhipu GLM-5.3 and DeepSeek V4 Pro to its model selector. Three days later, on the evening of August 17, those two models vanished from the product. Alibaba offered no explanation.

Three months later, on November 16, Qwen Office brought back ten models at once—Qwen4-Max, Kimi, GLM, DeepSeek, MiniMax, and Step all made the list.

That same week, Doubao Work labeled each model with a number: a quota consumption multiplier. The lower the multiplier, the fewer AI points are deducted for running similar tasks. It opened to a broader user base on November 12, along with multiplier and time-slot discounts.

Most people read this as "all three are racing to build a model ecosystem." Frankly, though, connecting models was already worthless by 2026.

What these three events really reveal is this: the upstairs floors are already fully renovated. The number of models, the chat interface, the basic features—anyone can spend a few days laying them out to match everyone else. Once that's done, everyone looks the same.

Floors can be rented, bought, or renovated. Foundations cannot.

Below, the author starts from this "model freedom" to discuss why the outcome of this round of the digital employee battle depends not on whose floors are higher, but on whose foundation can be modified.

01 Floors Are Folding, and "Model Freedom" Is a False Equalization

First, let's be clear about one thing: multi-model aggregation is an action that is "hard to decide, easy to deploy."

Technically, connecting a model is an engineering effort measured in "days." It took Qwen a few days to install two external models and only one evening to remove them; when it launched ten models at once in November, no one complained about the timeline. WorkBuddy was even earlier—when it launched on March 9, 2026, it already supported seamless switching among Hunyuan, DeepSeek, GLM, Kimi, and MiniMax, and later added an Auto mode for automatic routing.

So the hard part isn't connecting models; it's figuring out what you have left after connecting them.

Qwen's August "three-day tour" best illustrates this hesitation. If users come here only to pick a convenient model, where does Qwen's commercial return come from? If the platform's value lies in "matching tasks to the most suitable model," then whose model it is shouldn't matter. By November, the answer had leaned toward the latter—ten models launched together, paired with the claim of "autonomous selection based on task complexity, response speed, and token cost."

This amounts to admitting that the moat of office agents isn't the models themselves.

Doubao took another path but arrived at the same point. On October 29, Kimi K3.1 Pro, GLM 5.3, and DeepSeek V4.1 Flash were first opened to Team and Enterprise editions, with usage deducted directly from the enterprise account's AI points, so enterprises no longer had to procure separately from model vendors. On November 12, it opened to a broader user base, along with multiplier and time-slot discounts.

This step matters more than it looks. It turns "choosing a model" from a technical issue into a cost-governance issue. Users don't need to understand models; they only need to understand how much this task is worth and how much quota it should consume.

After model capabilities converge, this is the only place left to differentiate—turning incomparable things into things that can be accounted for.

Tencent is even more thorough: from day one, it doesn't let users worry about choosing models. Auto mode schedules automatically. The benefit is peace of mind; the cost is that the routing logic isn't detailed in public materials, so users can hardly know which company's model ultimately handled their sentence.

The three companies' action paths differ, but the result is the same: front-end products are rapidly converging. This is "floor folding."

When models can be priced, replaced, and routed, a moat built on models has already failed. Competition can only retreat—back to the layer that can't be bought with money or moved away.

02 The Foundation's Underlying Color: Three Existing Ecosystems That Cannot Be Quickly Replicated

The foundation of digital employees isn't compute power or parameters; it's the work ecosystem each company has built over ten years.

More critically, there's another layer: the "effective radius" of these three foundations is completely different, and it directly determines what an agent can and cannot do.

Tencent's foundation is relationships.

QuestMobile data shows that in September 2026, WeCom's PC monthly active users were about 20.3 million, versus DingTalk's about 12.57 million and Feishu's about 1.94 million. This is PC-only, not all platforms, but it's enough to show one thing: the densest organizational relationship chains in Chinese workplaces are in Tencent's hands.

WorkBuddy spreads along this chain with almost no need to educate the market. On October 14, WeCom and WorkBuddy released a joint version; after an enterprise activates and assigns seats, employees can invoke it directly from the sidebar of the contacts list. By the Tencent Global Digital Ecosystem Summit at the end of October, Tencent released "Buddy One," using WorkBuddy as a workbench to connect Tencent Docs, Meetings, Drive, and ima.

Product-side data holds up too. According to a Citigroup research report citing Tencent management in July 2026, WorkBuddy had 20 million cross-platform monthly active users and 13 million daily active users, with a DAU/MAU ratio between 65% and 75%. Separately, according to Tencent's disclosure at the Digital Ecosystem Summit at the end of October, the enterprise edition already served 50% of China's Fortune 500.

But the relationship domain's shortcomings are equally obvious. Agents are strong at finding people, pulling meetings, synchronizing information, and pushing processes, but weak at being far from core business decisions. It can help you pull twenty people into a meeting room, but it can't replace you in judging whether the meeting should be held.

ByteDance's foundation is knowledge.

The documents, OKRs, and meeting minutes Feishu has accumulated over ten years are among the most structured knowledge assets in Chinese enterprises. Doubao Work is deeply integrated with Feishu, and agents can summarize minutes based on enterprise context, search historical documents, and reuse accumulated methodologies.

The problem lies in two places. The first is reach: Feishu's PC monthly active users are about 1.94 million, not in the same league as WeCom's roughly 20.3 million. No matter how high the knowledge density, if the desktop entrance is narrow, the daily scenarios it can reach are limited. ByteDance's response wasn't to build products but to restructure the organization—on July 30, the Feishu product team and Doubao product team were merged into a new team, with Xie Xin reporting to Doubao head Zhao Qi while retaining the title of Feishu CEO, and the GTM team merged into Volcano Engine; on August 24, the TRAE and Coze teams were merged in; on August 25, the "Doubao Work" brand was released. Three adjustments in forty days essentially used organizational means to compensate for the entrance.

The second is more troublesome: shelf life. A channel policy from two years ago or a pricing model from last year, fed to an agent, will make it confidently give wrong judgments—and it won't tell you that the judgment is based on expired facts.

Knowledge density is an asset and a liability. The thicker the knowledge base, the more it needs a mechanism for "when to deprecate." None of the three companies has publicly discussed this mechanism, including ByteDance, which treats knowledge as a core asset. Nor can this be solved casually through technology—who has the right to declare a piece of knowledge obsolete is organizational power.

Alibaba's foundation is transactions.

DingTalk's process skeleton plus the e-commerce toolchain gives Qwen Office a depth that others find hard to replicate. The e-commerce edition released on October 14, after switching into "e-commerce operations" mode, can directly invoke more than a dozen merchant tools including Business Advisor, Qianniu, Alimama Wanxiangtai, Youzan, FastMoss, and LinkFox, completing product selection, creative materials, advertising, and operations analysis in the same workbench.

The data is solid too. Qwen Office CEO Chen Yusen (who has also served as DingTalk CEO since June 2026) said at the launch event that about 20% of existing paying users come from e-commerce; e-commerce users complete an average of 38 tasks per week, with usage depth and payment rates clearly above average.

But this foundation has the narrowest radius. Outside e-commerce, Qwen's depth in general office scenarios is discounted. The transaction domain can grow a very sharp knife, but that knife doesn't work well when moved to other industries.

So these three foundations aren't superior or inferior—only a matter of fit. Tencent's and ByteDance's foundations are broad but far from money; Alibaba's foundation is close to money but narrow.

03 No Matter How Thick the Foundation, If You Can't Feed It In, It's Zero

Having seen the three foundations, the next question is the one that truly determines the outcome: no matter how thick the foundation, has it been fed to the agent?

Having a foundation doesn't mean using it well. Two barriers stand in between: organizational friction and execution architecture.

The first barrier is permissions.

For an agent to work, it must first see the data. But data is scattered across different departments, different systems, and different business groups. This isn't a technical problem; it's an organizational one—no one wants to hand over their data permissions, even to another team in their own company.

ByteDance passed this barrier most smoothly. The Feishu product team was merged into Doubao, reporting lines were straightened, the AI transformation path for existing customers was shortest, and there was no need for cross-business-group negotiation internally. The organization moved first, the product second—the order wasn't reversed.

Qwen's base is DingTalk, which in theory should be smoothest, since organizational relationships are supplied by default. But e-commerce and general business belong to different systems, and toolchain integration and permission opening still require internal run-in. The outside world can't see the progress here, only the results: the e-commerce edition's completeness is clearly higher than general scenarios, and that itself is a signal.

Tencent is the most typical. According to public information, WorkBuddy is advanced by the CSIG system, while WeCom belongs to WXG—two business groups. The product launched in March and reached tens of millions of users, but only with the joint version on October 14 did the two lines truly complete the handshake on data permissions and entrance.

The product ran ahead while the organization chased from behind. Those seven months are the price tag of friction cost.

The second barrier is execution architecture.

There is no official definition here; the author can only infer from public product forms. For the same task—say, "pull sales data from five platforms and generate a review report"—the three companies' approaches are clearly different.

WorkBuddy leans toward "teams." In public materials, it emphasizes parallel collaboration among multiple expert agents: a data expert, an analysis expert, and a design expert each handle a segment, the main agent schedules, local files are authorized for reading and writing, and results are saved directly as local documents. The benefit is fault tolerance—if one path goes wrong, it can be rerun separately without affecting the whole.

Doubao leans toward "direct output." It emphasizes directly operating the browser and computer to complete tasks across software; long tasks can be run on a cloud computer without interruption even if the device is shut down. The benefit is speed and broad coverage; the downside is that in extremely complex scenarios it tends to go all the way down one path, with high rework costs.

Qwen leans toward "self-checking." In its November statement about connecting ten models, one line is key: allocate by task complexity—complex reasoning to Qwen4-Max, solution checking to GLM-5.3 and DeepSeek V4 Pro, daily copywriting to cheaper tiers. This is effectively one model producing a solution and another cross-reviewing it, leaning toward an engineer's mindset: verify the logic before delivery. At the same time, it strings more than a dozen merchant backends into the same dialog box, which is closer to toolchain orchestration.

Strictly speaking, Qwen overlays two features—"self-checking plus toolchain"—with the former managing quality and the latter managing the closed loop.

The three routes have no absolute superiority or inferiority, but they determine the cost structure of the same task. The team route spends scheduling cost, the self-checking route spends reasoning cost and time, and the direct-output route spends rework cost.

So the outcome of digital employees will ultimately land on a very unsexy metric: unit task economics—how much money and time it takes to complete a job correctly, and how many reworks.

This metric is more honest than DAU, and more honest than the number of models. Regarding entrance heat, a set of numbers is often mixed together: WorkBuddy's 13 million DAU and Doubao Work's 2.049 million PC MAU. One is DAU, the other MAU; one spans three platforms, the other is single-end. Direct comparison is unfair, but the confusion in this set of numbers is itself a reminder: entrance heat does not equal productivity.

For enterprises that need to pay, the author suggests asking only three questions: At which layer do your core decisions happen—relationships, knowledge, or transactions? To adopt this agent, how many systems' permissions must be opened, how many approvals must be passed, and how long until launch? How much quota does a typical task consume, and who bears the cost of failed retries?

The first question determines whom to choose, the second determines whether it can be implemented, and the third determines whether it will spiral out of control after scaling.

Conclusion

Back to the title: why is the outcome of this battle determined by the "foundation" rather than the "floors"?

Because floors can be rented. Models can be integrated, interfaces can be replicated, features can be added—as long as you're willing to spend the money, you can have them up and running to match the competition within a few days. This November's round of model arms racing is essentially three companies painting their floors the same color on the same day.

What truly sets them apart is three immovable foundations—Tencent's collaborative relationships, ByteDance's knowledge density, and Alibaba's transaction linkages. And whether those foundations can be put to use comes down to three hard challenges: whether the organization is willing to relinquish permissions, whether knowledge has a deprecation mechanism, and whether task orchestration can support general complex scenarios.

None of these three hard challenges can be solved by buying APIs. They require overhauling organizational structures, permission allocation, and interest chains. What needs to be knocked down is the load-bearing wall—not given a fresh coat of paint.

So the endgame of this battle isn't about how high each company's floors reach. It was already foreshadowed the moment each company decided to renovate its own foundation.

Solemnly declare: the copyright of this article belongs to the original author. The reprinted article is only for the purpose of spreading more information. If the author's information is marked incorrectly, please contact us immediately to modify or delete it. Thank you.