In the AI era, the real opponent has never been competing products, but old paradigms. Over the past decade, the evolution path of enterprise office software has been almost linear: from PC to mobile, from documents to collaboration, from workflows to platforms. They solved one problem—how people can collaborate more efficiently. But in the AI era, that very problem is losing validity. When AI begins to have the ability to understand, plan, and execute, what enterprises truly need is no longer a “collaboration tool,” but a work system that can host AI, dispatch AI, and be accountable for outcomes. This is precisely the starting point of DingTalk’s “self-revolution.”
Why DingTalk Must Cut Into Itself
In 2025, things were not easy for DingTalk. The dividend of mobile internet has faded, and collaborative office software has entered a phase of stock (incremental-to-zero-sum) competition; rivals such as Feishu, WeCom (WeChat Work), and Microsoft Copilot have continued to increase their bets on AI-ification; and users’ excitement about “one more AI feature” is fading quickly. More crucially, there is a structural problem: in the AI era, DingTalk and its competitors have not formed a genuine “generation gap.”
This is also the reality Chen Hang (nickname: Wuzhao) saw first after his return—if DingTalk merely treats AI as a functional plug-in, DingTalk will ultimately be replaced by “a DingTalk that is more like AI.” Thus, a more aggressive question was put on the table: wait for others to use AI to revolutionize DingTalk, or have DingTalk “kill” itself with its own hands first? The answer is the latter.
The essence of this transformation is not a product upgrade, but a role reversal. In the old DingTalk: humans were the executors and thinkers; software was a tool and an assistant. In the AI era, Wuzhao’s new judgment is: AI becomes the subject of work—perceiving, judging, planning, executing; humans retreat to the decision-making layer—setting goals, making trade-offs, and backstopping results.
This means that messages, spreadsheets, search, workflows, approvals—operations that previously required manual completion by people—should, in essence, all be taken over by the system; it means that “tapping an app” and “running a process” are no longer the core of efficiency, and “issuing a command—getting a result” is. Under this premise, DingTalk can no longer be just an app. It must evolve into an “operating system that serves AI.”
Agent OS: Not a New Product, but a New Foundation
This is also the context for DingTalk’s official launch of Agent OS on December 23, 2025. Rather than calling it a new product, it is more like a complete architectural rewrite. Agent OS does not care which app you clicked; it cares about how a task is broken down into steps, which Agents execute them, how permissions are controlled, how results are audited, and how failures are backstopped.
Within this system, DingTalk ONE is no longer just a UI entry point, but a unified desktop for human–AI collaboration; AI Search & Ask is not merely search and Q&A, but the starting point of tasks; “Wukong” plays the role of a dispatcher, translating natural-language commands into executable chains; and DEAP enables enterprises to build, manage, and govern their own Agents.
With these capabilities assembled together, DingTalk for the first time no longer emphasizes “collaboration,” but emphasizes delivery. But a more realistic question quickly emerged: if AI is to truly execute tasks, how does it understand the real world?
Enterprises’ most important data has, for a long time, not been in systems. Face-to-face communication in sales, inspection records on factory floors, Excel spreadsheets on administrative staff’s local computers, decisions spoken casually in meeting rooms—massive amounts of information exist offline, or exist in extremely unstructured ways. Once this data is missing, no matter how smart the AI is, it can only remain at the level of “armchair strategizing.” This is precisely why DingTalk has begun making hardware.
DingTalk A1 is not a piece of hardware made for hardware’s sake. Its role is more like the “sensory organ” of DingTalk’s AI system. Through an almost imperceptible recording card, spoken language in meetings, interviews, inspections, and sales is, for the first time, systematically collected, structured, accumulated, and fed into DingTalk’s AI system, becoming data assets that can be analyzed, called upon, and reused.
In this sense, A1 is not selling recording; it is selling the enterprise’s data entrance to the real world.
Even more intriguing is that seemingly ordinary physical button. Right now, it is a voice memo; but in DingTalk’s planning, it is regarded as the future physical entrance for “one-click calling an AI Agent.” When AI can truly take over tasks, pressing this button means directly triggering a complete workflow, rather than opening an app.
This Is Not the Endgame, but the Direction Is Already Clear
If A1 is the “ears,” then AI Spreadsheets are the “central nervous system” of DingTalk’s entire system.
In countless enterprises, what truly carries business operations is not ERP or CRM, but Excel. They are non-standard, scattered, and maintained by people, yet they truly exist. DingTalk did not try to replace these processes with a standardized system; instead, it chose to let AI enter the spreadsheet itself, making the spreadsheet a convergence point for data, workflows, and AI applications.
Voice-based form filling, photo-to-structured data, and cross-system data linkage allow many processes that previously ran only by relying on human labor to barely keep them going to, for the first time, gain the possibility of being AI-ified—while almost not requiring any change to frontline employees’ work habits. This is not disruption, but an extremely pragmatic reconstruction.
Of course, what is truly difficult is not only the product.
At the organizational level, DingTalk has also gone through a “self-negation.” After Wuzhao’s return, he required the team to be accountable for results and accountable to customers; he laid out all the accumulated demands built up over many years, forcing trade-offs; he split the organization into smaller combat units and pushed change with an extremely high-frequency iteration rhythm; he used products and progress, rather than slogans, to filter for the people who truly believe in AI.
In just a few short months—from AI DingTalk 1.0 to Agent OS, from software to hardware, from tools to systems—DingTalk has almost been rebooting a mature big-tech product in the way a startup would.
This action of “killing DingTalk,” of course, is still far from being called a success. The endgame of Agent OS remains blurry; whether AI can stably take on the role of the work subject, and whether enterprises are willing to truly hand over execution authority, both need time to be validated.
But at least one thing can already be confirmed: in the AI era, the greatest risk is no longer doing something wrong, but not daring to overthrow what used to be right.
DingTalk chose to cut into itself first; it may not be a sure win, but if the way work is done truly gets redefined in the future, at least it is sitting at the table.
[Disclaimer]: The above content reflects analysis of publicly available information, expert insights, and BCC research. It does not constitute investment advice. BCC is not responsible for any losses resulting from reliance on the views expressed herein. Investors should exercise caution.