Coverage, Right and Wrong, and Experience: Three High Grounds in the Enterprise AI Context War

Authors | Lin Ke, Zheng Hao

On September 2, Tencent's WorkBuddy open platform was launched in Shenzhen. The market is more concerned about which ecosystems have been included in WorkBuddy, but several developer agreements announced on the same day revealed Tencent's key intentions in its office AI strategy.

This agreement divides WorkBuddy's ecosystem data into three categories: developer-owned data belongs to the developer, user-owned data is determined by agreement between the developer and user, and Tencent only needs the third category:

Call frequency, resource consumption, response time, error logs, and feedback evaluation.

There are three main uses: statistical analysis, quality assessment, and labeling.

A company that was vying for market share abandoned user data, focusing only on the traces of failed tasks.

In the same week, Qianwen Office released its first-month data, showing over 30 million registered users within 30 days of its launch, with enterprise users accounting for more than half. Baidu KuKu AI's standalone client was released on August 14th, announcing over 25 million monthly active users for its AI office platform. ByteDance's Doubao Work was released on August 25th, deeply integrated with Lark. These four products have different forms, but they all share the same goal—targeting the context of enterprise users.

After years of competition, DingTalk, Lark, and WeChat Work have achieved a combined coverage rate of 92%. At this level, incremental growth has almost disappeared, and what remains is a battle for existing market share.

The entry point for the next decade is unlikely to be OA systems, but rather whose AI can help users do the most work. As long as work is done, context is required.

The context is not a homogeneous resource; it is forming three high grounds in a gradient, each more scarce than the last.

I. High Ground 1: Data Coverage

At the AI Productivity Conference on July 15, Kingsoft Office CEO Zhang Qingyuan put it very bluntly: "Every dialogue of the large model is a rebirth. It has no memory—all the memory and context are stored in the software." He gave the division of labor as follows: the large model provides intelligence, and the software provides context.

Being able to read information means enabling agents to see the information within the enterprise. Approval chains, customer data, meeting minutes—these are things scattered across different systems, and agents cannot do anything if they cannot see them.

This used to be the biggest advantage of collaborative platforms.

The more years a company uses DingTalk, the more context it accumulates over those years. DingTalk has approximately 200 million monthly active users and over 20 million enterprise organizations. Lark's ARR is expected to exceed 3 billion yuan in 2025, and its ARR is expected to grow by more than 100% year-on-year in the second quarter of 2026. WPS boasts document assets accumulated over many years on 676 million monthly active devices.

For these platforms, AI is born right in the middle of the data.

But this high ground is being leveled by the platform itself.

Back in March of this year, WeChat Work open-sourced wecom-cli, covering core categories such as messaging, documents, smart spreadsheets, to-do lists, calendars, and meetings; DingTalk simultaneously fully implemented CLI, rewriting thousands of core capabilities into standardized commands that AI can directly call; Lark also open-sourced its CLI plugin and announced support for all mainstream Agent tools.

The three companies did this with the intention of making their AI stronger, but once the universal protocol is released, anyone can use it.

On August 18th, Qianwen Office officially integrated with WeChat Work. After authorization, users can access WeChat Work's smart spreadsheet and document creation functions through Qianwen Office's chat interface.

These collaborative OA (Office Automation) walls weren't breached by competitors; they were proactive modifications made by the platform itself to adapt to the AI era.

When the context can be read by any authorized agent, the question of "how much can be read" is no longer a significant barrier.

Although you can't do anything without it, it's becoming more and more like air—essential, yet hard to price.

II. High Ground 2: True or False Questions That Can Be Judged

Companies are willing to pay agents only if they can clearly demonstrate whether the job was done correctly.

This hurdle is far higher than simply "reading more." An agent might write a report that seems reasonable, but are the data definitions accurate? An approval process might be pushed forward, but does the order of signatures conform to company policy?

Without a definitive criterion, an agent is just a machine that appears to be very busy.

Organizational frameworks based on collaborative OA systems become even scarcer here. Information such as who approved what and when, how the organizational structure is nested, how permissions are inherited, and who collaborated with whom on which project naturally carries a label of right and wrong.

Document versions from A to B to C are documented and traceable, and schedules and actual attendance can be cross-verified. This kind of context can tell the Agent whether it is doing things correctly.

DingTalk and Lark maintain their advantage in this high ground. The permission chains, approval chains, and collaboration relationships of more than 20 million organizations have been accumulated for a long time. Lark Aliy, which was upgraded to Doubao Work Partner, has a permission system that is close to the user's own operation and the entire process is traceable. Sensitive operations require manual confirmation.

In contrast, Kingsoft Office has hundreds of billions of cloud documents stored on 676 million monthly active devices. Although the volume is larger, the documents are essentially read-only corpora. Agents can understand them, but cannot act on them. There is no inherent criterion for judging whether a report is correct or not.

Jinshan is safest on the first high ground, but more precarious on the second. If the competition in the entire industry ultimately stops at the first high ground, Jinshan is unshakeable; if it continues to move forward, Jinshan's uncertainty will increase.

Yu Hongxiao, head of WorkBuddy's open platform, used the example of financial automation to further clarify the difference between the two levels of high ground.

For example, at the end of the month, when reconciling accounts, thousands of transactions are checked one by one, and even a single mistake can lead to a disaster. Bank statements are on the bank's network, invoices are in the tax system, and vouchers are in the ERP system—the agent cannot read this data, which is the first major problem.

The question of whether the discrepancy is due to omission, duplication, or the money still being in transit remains in the mind of the finance department, and this constitutes the second major challenge.

Without access to data and without know-how, even the smartest model is useless.

WorkBuddy's approach is to outsource the second highland to the ecosystem.

The open platform initially releases five modules: Buddy Applications, Experts, Skills, Connectors, and Hardware. FineReport integrates its data analytics know-how, Beisen integrates its HR scenario judgment capabilities, and Roger Technology integrates its tax compliance experience.

Each partner brings their own criteria for "what's right" in their field, filling the industry gaps that WorkBuddy itself lacks. The first batch of over 30 Buddy applications covers more than 20 fields, including finance, law, healthcare, education, and public welfare, with over 100 partners joining.

Shen Tao, Vice President of Strategy at FineReport, believes this opportunity is 10 to 100 times greater than the window of opportunity for ToB SaaS when he launched JianDaoYun ten years ago, and he proactively proposed that it be completely free; Xiao Feng, Vice President of Technology at Weimob Group, said that this entry point must be seized; Wang Yao, Vice President of Beisen, more calmly pointed out that there is currently no specific revenue sharing plan; Tencent itself also confirmed at the press conference that payment capabilities and distribution channels are still under accelerated construction.

This is a frank picture: everyone is betting on the future, no one is making money now.

III. High Ground 3: Turning Failure into Skill

If the agent makes a mistake, and the error can be categorized, rerun, or fed back into the model or scheduling engine, it is not just a failure, but a piece of training data.

This is the third layer of value in the context, and the only one with a compounding effect.

The third set of data that Tencent requested in the open platform agreement included call frequency, resource consumption, response time, error logs, and feedback evaluations. The purpose of these data was marked as statistical analysis, quality assessment, and labeling, which points to this very high ground.

Labeling implies that someone is categorizing failures, turning "why it didn't finish this time" into a criterion for the next attempt.

The supporting engineering facilities were started even earlier.

Tencent has jointly launched the WorkBuddy Bench benchmark test by four teams: Youtu Lab, Keen Security Lab, WorkBuddy team, and YunDing Security Lab.

The 260 questions cover four subsets: code, web, office, and security. The key design is dual Harness parallelism, where the same batch of tasks are run on two execution frameworks, CodeBuddy Code and Claude Code, to conduct experiments on the separation of framework variables and model variables.

The evaluation assets are only introduced after the Agent has finished running, and the entire process is conducted offline and isolated in a sandbox to prevent the memorization of answers.

The purpose of this test may be to establish a resettable, rerunnable, and attributable evaluation environment in non-coded enterprise work scenarios.

Whoever has such an environment first will be able to systematically turn the pitfalls users encounter every day into capabilities.

However, this high ground still faces systemic construction difficulties. More than half a year after its launch, WorkBuddy users have frequently reported problems such as repetitive execution of long-term tasks and confusion in skill memory.

The testing foundation is currently more focused on the intention, and the loop from failure trajectory to annotation to model improvement is currently mostly limited to the product level.

DingTalk and Lark, on the other hand, have a structural advantage in the third high ground.

Task data within the platform is preferentially fed back into its own model training and evaluation system. For example, if a user stumbles while running a task on DingTalk, Qianwen can learn from this failure, and the DingTalk experience will improve accordingly.

On July 30, ByteDance merged the Lark product team into Doubao, thus completing this closed loop at the organizational level. Moreover, their second stronghold is the strongest, which means that failure is inherently labeled. Whether the approval process is successful or not is a natural label, saving even the labeling step.

The third high ground requires continuous investment, and the key question is where the money for this investment will come from.

IV. Misalignment between Transformation and Training

In the summer of 2026, the three major participating companies each pushed forward with organizational consolidation.

Tencent transferred the business and team of QClaw Product Center to the department where WorkBuddy is located; Alibaba merged QoderWork, Wukong, and MuleRun into Qianwen Office, which was put under the management of Chen Yusen, who took over as CEO of DingTalk on June 11.

ByteDance merged the Lark product team into Doubao, with Zhao Qiren in the top position. Shortly after, a second consolidation was carried out, and Trae and Kouzi were also merged into the Doubao system.

The three adjustments differed in direction. Only ByteDance integrated its collaboration platform product team into the model side, Alibaba handed over the Agent to the head of the collaboration platform, and Tencent changed the relationship between the two Agents. The common thread was the same: resources were being concentrated towards a single Agent entry point.

After the merger, their respective assets and conflicts gradually came to light.

DingTalk and Lark currently rely primarily on "personality" rental income for their revenue.

A 500-person company uses Lark, paying per seat, with the price multiplied by the number of employees, and renewing the contract annually. This business is stable and predictable, which also supports the platform's ability to continuously maintain its organizational permission system, approval workflows, and collaboration records—the most valuable context.

However, the pricing logic for agents is shifting away from individual users. WorkBuddy uses a points system, deducting fees based on usage; Zhang Qingyuan stated in July that subscriptions and AI usage will coexist in the long term; Baidu KuKu AI Enterprise Edition is packaged according to customization. No new entrants are still charging per user.

The more robust the organizational structure, the higher the maintenance costs and the greater the reliance on revenue from memberships; however, the paradox is that the more widespread the agent network becomes, the more companies tend to pay for results rather than for the number of people.

The business that provides the most valuable context is precisely the business that was first rewritten in the Agent era.

WorkBuddy has gone in the opposite direction. It lacks an organizational framework and corresponding financial pressure. Tencent says it will not set commercial KPIs for the team and is still in the investment stage.

WorkBuddy's monthly active users, sourced from institutions, have exceeded 20 million. Although this is ahead of many large enterprise agents, it does not mean that the quality of the context and tasks is deep enough, since each context source has to be connected almost from scratch.

The open platform strategy seems to be trying to compensate for the lack of depth with breadth. In the white paper, the exclusivity of several cooperation models is listed as non-exclusive in three places.

Lin Zuolu, head of the open ecosystem, explained that Tencent did not set a KPI for the number of partners in the ecosystem to be considered a success. "Openness is what we do, and the ecosystem is an outcome."

Although the ecosystem is growing rapidly, it's not yet time for Tencent to consider building a competitive moat.

Therefore, the side with the best training materials bears the greatest pressure to transform; the side with the least pressure to transform has the least training materials.

Looking at Baidu's KuKu AI with 25 million monthly active users and Kingsoft's with 676 million monthly active devices, they are both in a precarious position, with different contexts, levels of influence, and monetization paths.

This means that currently, no single major manufacturer can simultaneously occupy all three of these key positions.

V. The layer that cannot be taken away

There are some scenarios that are difficult for major companies to reach using their current models, such as financial risk control, government decision-making, and state-owned asset compliance.

These scenarios may have higher context quality, the most rigid processes, the most severe consequences, and the most attributable causes of right and wrong, making them the places that agents most want to enter but also the most difficult to enter.

At the ecosystem launch event on September 2nd, Rogyr Technology, a provider of digital financial and tax services, showcased an AI-powered smart terminal: the Rogyr Supercomputing Box. This edge device can run models with 35B parameters and encapsulates a ready-to-use tax compliance AI. Whether it's used on the edge or in the cloud depends on the customer's requirements for data export.

WorkBuddy's data processing protocol draws a line, not covering data stored in the developer's own environment or a third-party environment.

In version 5.5.2, WorkBuddy quietly added a local model Beta entry at the bottom of the settings page. The first downloadable model is Qwen3.6 35B-A3B, and the data does not leave the local machine and does not consume cloud quota.

Edge-cloud separation allows agents to enter the scenario, but at the cost of the edge data not entering Tencent's runtime, and the most valuable failure trajectory will not be seen by the public cloud. The third high ground will be locked here.

The collaborative platform's response to this scenario is private deployment.

DingTalk can provide a dedicated environment for enterprises, while Lark's solution emphasizes data ownership and account consistency, and full traceability of operations. However, the essence of privatization is to move the platform into the customer's walled-off space, and the scale effect and network effect are greatly reduced in this scenario.

At that time, compliance will begin to become a competitive barrier.

On August 6, the China Academy of Information and Communications Technology (CAICT) released the first batch of assessment results for the capabilities of intelligent office agents, scoring them across four dimensions: intelligent interaction, full-scenario task execution, security protection, and skills management.

Qianwen Office became the first product in China to pass 16 tests. Doubao Work also emphasized its dual certification status when it launched on August 25th.

In August, Bay Engine WorkBuddy was launched on the Guangdong government system, and the first batch of provincial-level units, including the Guangdong Provincial Medical Insurance Bureau, have carried out dozens of scenario docking and pilot projects.

Whoever obtains the qualifications and delivery cases first will have an extra ticket to bid in high-sensitivity scenarios; and in high-sensitivity scenarios, what is delivered is more about trust.

The growth rate of trust and the expansion rate of the open ecosystem are two parallel tracks.

This means that the most valuable segment of the entire market for agents may not necessarily result in a winner-takes-all situation. It will be dispersed among a few companies with delivery capabilities and compliance qualifications in the long term, which naturally complements the broad logic of open platforms and also constitutes a natural boundary.

VI. A Two-Way Acceleration

Tencent's willingness to sacrifice user data in order to expand its ecosystem, even at the cost of evidence of failure, reveals not just a unique strategy of Tencent, but a common predicament of the entire industry.

The existing context is being leveled out by open protocols, and the same is true for the barriers in the experience layer. Even the things that major companies are vying for are becoming increasingly similar:

The ability to translate context into reliable execution, and the mechanisms for continuously improving this ability.

The three high grounds provide a ranking standard. The first high ground has been commoditized, and readability is no longer a moat; the second high ground forms a filter, and whether the B-end is willing to pay depends on whether the agent can clearly explain right and wrong; the third high ground is the only place with compound interest, but it requires the support of the first two high grounds—without sufficient context, there is insufficient execution, and therefore, it is impossible to accumulate enough failures, making it even more difficult to gain considerable compound interest experience.

This chain is no longer simply about comparing the quantity or quality of contexts, but rather about competing to be the first to connect the three high grounds and form a closed-loop capability.

However, building a closed-loop capability always requires a sustainable cash flow to create positive feedback; disregarding returns is only suitable for the initial stage of business exploration.

In the past, the seat system of collaborative platforms provided the deepest context, but the seat system is being eroded by the logic of agents.

Today, points-based and pay-as-you-go systems can match the usage patterns of the Agent era, but they cannot support the organizational service costs accumulated by collaborative platforms over the past decade in the short term.

Whoever successfully integrates the non-seat portion of their revenue structure first will have the resources to continuously invest in the third high ground, instead of being dragged back by the inertia of the old model.

The two routes are approaching each other faster than expected.

Collaboration platforms are expanding outwards. Doubao Work inherits Lark's permission system and knowledge base, and is available for download as an independent product to everyone. Qianwen Office runs simultaneously on DingTalk, Lark, and WeChat Work.

The open approach is moving inward, with WorkBuddy's identity system, cross-platform memory, and points system all creating usage inertia; Buddy applications and expert centers point to industry depth and the accumulation of organizational knowledge.

Every step these big companies take is an attempt to push the context further and reach higher levels.

As Zhang Qingyuan said at the July press conference, this applies to every participant in this competition: AI is the only and most important opportunity in the next few years.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.