Why has the competition for context become the new battlefield of office AI?

Recently, Google released the Gemini Agent for enterprise office scenarios.

This agent is powered not only by Google's in-house Gemini, but also by Anthropic's Claude, a competing large language model.

Google's explanation on this point is straightforward: leading models may be replaced every few months, and the context, skills and data accumulated by enterprises over the long term are the real core assets.

Clearly, models are no longer a long-term moat, and the competition for office AI has shifted to the battle for contextual resources.

On October 14, WeCom and WorkBuddy also announced an in-depth strategic partnership.

WorkBuddy has officially joined the WeCom contact list as a conversational bot. The two sides have integrated their account systems, and opened up underlying capabilities including messaging, documents, meetings, cloud disk and more within the scope of enterprise authorization.

One is the collaborative office platform with the largest user scale in China, and the other is the office agent with the largest user scale in China.

The combination of the two gives office AI the opportunity to access both internal enterprise collaboration and external business operation scenarios at the same time, further expanding the boundary of enterprise "context".

Over the past month, companies including Microsoft, OpenAI and Anthropic have also made intensive layouts around "context", successively launching capabilities such as Agent identity, collaboration space, external connectors and organizational memory.

Office software vendors are trying to embed Agents deeper into workflows, while model companies have begun to build dedicated office environments for AI.

Why has the battle for context become the new battlefield for office AI?

What exactly is the "context" that office AI is competing for?

The context in office scenarios goes far beyond several uploaded files.

Chat records, emails, project documents, meeting minutes and business data are all part of context. Organizational division of labor, version status, approval processes and access permissions also count.

At a deeper level, there are implicit rules formed by enterprises over a long period of time: what caliber to use for a quotation, which links a product release needs to go through, and why a certain decision was rejected in the past.

This information is scattered across documents, chats, calendars and business systems. Employees rely on long-term work experience to connect these pieces of information and understand the context of events. To complete full tasks independently, Agents must also establish the same logical connections.

The series of new features released by Google this time are all built around this goal.

Gemini Agent is deeply integrated into Google Workspace, accessing emails, files, meetings and collaboration relationships; it also connects to Microsoft 365, Slack, Salesforce, ServiceNow and business databases, bringing information outside the platform into the task environment.

Enterprises can also assign independent identities, mailboxes, calendars and storage spaces to the "Coworker Agent", allowing it to participate in long-term work as a fixed role.

Microsoft has taken a similar path. On September 25, Satya Nadella announced the largest update of Copilot to date, calling it the "operating system for work".

The new version of Copilot covers conversation, content creation, tool building and autonomous execution, with Word, Excel and PowerPoint directly integrated into the task chain.

The agent has independent identity, memory, computing resources and workspace, and can access team collaboration and document processes. It reads enterprise context on the one hand, and generates new context through discussion, execution and correction on the other.

Google and Microsoft have taken the lead in this competition. Years of operation of office software have allowed them to accumulate a large amount of context assets, and their next focus is to deepen internal capabilities and expand external connections.

The situation faced by OpenAI and Anthropic is different.

The two companies have powerful large models, but lack long-operated office software and naturally precipitated enterprise context. Therefore, they are building a working environment that can carry context.

At the end of September, OpenAI launched ChatGPT Space, Pages and team collaboration features.

Space stores project materials and shared background information; Pages carries work results that are continuously modified; the resident agent Dots is equipped with a cloud computer and browser, and can handle tasks for a long time.

Discussions, writing, modification and delivery are completed as much as possible inside ChatGPT, and the whole working process is precipitated accordingly.

Anthropic integrates Chat and Cowork into a unified Claude, and adds document and presentation tools.

Users do not need to select chat or execution mode in advance. Claude can read files, call connectors, generate documents and continue to advance tasks in the same conversation.

The two types of vendors have thus entered each other's core areas: office software vendors add intelligent capabilities to their original products, while model companies build identities, collaboration spaces and organizational memory for AI.

Office AI is becoming more and more like an office suite, and office suites are also increasingly developing into Agent platforms.

From partial context to full-domain context, what has changed?

Vendors compete for context, but what enterprises ultimately care about is operational results. The core demand of office AI is shifting from efficiency improvement to revenue growth.

According to Accenture's *2026 China Enterprise Digital Transformation Index*, 82% of Chinese enterprises hope to use AI to open up new business growth space, and only 18% of enterprises take cost optimization as their main expectation.

The same report shows that 88% of the surveyed enterprises have passed the AI pilot stage, but only 14% of enterprises have achieved significant improvements in production efficiency, revenue or profit. AI tools are spreading rapidly, but the business returns are still limited.

For consumer-facing enterprises, growth ultimately falls on a series of specific actions:

Identify high-intent customers, reduce follow-up omissions, recommend products based on real needs, handle complaints in a timely manner, and find opportunities for repurchase and additional purchases from continuous communication.

ByteDance and Alibaba are both pushing office AI into business operation links.

Doubao Work leverages Feishu's group chat, meeting and project data to try to connect customer insights, marketing content and team execution.

Qianwen Office connects DingTalk and business systems to participate in lead sorting, sales follow-up and customer service analysis.

The in-depth integration between WeCom and WorkBuddy also cuts into this value chain.

Tencent has guided this product collaboration to a new idea — "full-domain context".

According to the official definition, full-domain context includes not only internal collaboration data such as documents and meetings, but also external customer demands, service records and business feedback.

Internal information records how enterprises work, while customer information reflects how the market responds.

Only when the two types of information are connected can Agents judge who should handle a consultation, whether a demand has been included in the product plan, and which departments are needed to cooperate for a business opportunity.

WeCom is exactly at the intersection of the two types of information — it connects internal organizational structure, group chats, meetings, documents and business systems, and connects customers, upstream and downstream partners through WeChat externally.

As of the release of WeCom 5.0, the platform has connected more than 14 million real enterprises and organizations, and more than 750 million WeChat users receive services through WeCom every day.

This connection has already generated quantifiable business value before.

Cases disclosed by WeCom show that Rainbow Department Store once organized 50,000 shopping guides to serve 10 million customers, established 57,000 communities, and achieved an annual online GMV of 5.1 billion yuan.

Shopping guides can still respond to consultations and complete transactions after the store is closed, and the relevant turnover has increased by 16%. After accessing the CRM system, shopping guides can also view member information and consumption records to improve the pertinence of product recommendations.

AI is beginning to transform this service capability from individual experience into organizational capability.

WeCom 5.0.9 launched in June this year has provided the "service summary" function, which can extract demands, intentions and transaction bottlenecks based on the communication records between employees and customers, and screen out customers that need key follow-up every day.

In the past, customer intention judgment mainly relied on sales experience and morning meeting reviews.

Now, enterprises can solidify the judgment logic into unified rules. Even new employees can quickly understand where the last conversation stopped, what the customer cares about, and at which step the transaction is stuck when they open the chat history.

WorkBuddy further integrates customer communication into the Agent's task chain.

WeCom provides customer relationship, organizational identity and permission boundaries, while WorkBuddy is responsible for analyzing information, calling tools and delivering results, thus connecting the customer front line with the internal enterprise execution process.

For example, when a customer asks in WeChat whether a product is suitable for him, and also cares about in-store inventory and discounts, the shopping guide can let WorkBuddy combine historical communication, product materials, promotion plans and internal documents to sort out product comparisons, in-store visit suggestions and communication scripts, and send them to the customer after confirmation. Appointments, transactions and subsequent follow-up remain in WeCom.

At the same time, high-conversion communication methods, demand judgment and follow-up rhythm can also be precipitated into reusable methods for the whole team, instead of being mastered only by a few top sales performers.

"Full-domain context" thus expands the boundary of office AI.

After internal and external information converges, office AI begins to move from office workflows to business operation processes.

How does the expansion of context boundary reshape the business model of office AI?

The continuous expansion of context boundaries is also changing the business model of office AI, providing a new solution to the long-standing pain point of "large user scale but low payment rate" in the collaborative office industry.

According to DingTalk's disclosure, as of August 2025, the platform serves more than 26 million enterprise organizations, with more than 190,000 paying organizations, and the conversion rate is still less than 1%.

The global market has a similar situation. As of June 2026, the number of paid seats of Microsoft 365 Copilot has exceeded 30 million. Compared with more than 450 million commercial paid seats of Microsoft 365, its penetration rate is close to 7%.

Traditional office software is usually charged by seat: one employee corresponds to one account, and the price has little to do with usage frequency, task difficulty and business results.

This model is built on a low marginal cost. Office AI brings a new cost structure, which can be roughly divided into three parts.

First, identity and permissions. Who can use the Agent, and which data the Agent can access. The seat model is still suitable for pricing this part.

Second, tasks and consumption. How many models, tools and business systems the Agent calls, how many queries, updates and process executions it has completed. Usage-based billing can cover the continuously generated costs.

Third, business results brought by full-domain context. Whether the customer service problem is solved, whether the lead is filtered, whether the appointment is successful, and whether the order is closed. When the results can be clearly defined, vendors can further charge by task or output.

The billing method of the industry has begun to change along this path.

In this update, the WeCom backend adopts the design of "fixed quota + additional quota package".

Enterprises first get a predictable basic quota, and then add additional usage according to actual needs. Administrators can view consumption by cycle, department and member, and adjust the quota limit individually or in batches.

Other vendors are also adopting hybrid billing models.

Microsoft launched Microsoft 365 E7, priced at $99 per user per month, packaging Microsoft 365 E5, Copilot, Agent 365 and Entra Suite into one seat.

Identity, permissions, security and basic AI capabilities are covered by the subscription, while part of the Agent operation and complex tasks continue to consume Copilot Credits. Enterprises can thus manage both users and actual workload at the same time.

In its 2026 semi-annual report, Kingsoft Office proposed that WPS 365 will explore the "seat + AI usage" model. The basic subscription corresponds to office entry and organizational management, while AI usage is calculated according to the usage depth and resource consumption of complex tasks.

In the next step, the new model of charging by task or result may be implemented faster in scenarios with high standardization and easy-to-confirm business results.

Martin Lau, President of Tencent, said at this year's Q2 earnings call that WorkBuddy will continue to search for high-value-added application scenarios, bring additional value and returns to users, and even help users earn more money in some cases.

Following this path, context first becomes the capability foundation of office AI, then expands its business boundary, and now begins to undertake the functions of measurement and attribution.

This change is not a single product function upgrade, but an overall transformation covering models, applications and computing power systems.

As Ma Huateng said in Tencent's Q2 financial report: "We are building a brand new, AI-empowered Tencent at three levels: intelligence, applications and infrastructure."

The integration of WeCom and WorkBuddy is a preliminary exploration in this direction, and also allows Tencent's ecological advantages to further realize value.

Similar changes are also taking place in companies such as Microsoft, Salesforce, Alibaba and ByteDance. Different companies have mastered different context resources, and their commercialization paths will also diverge accordingly.

The future of office AI ultimately depends on who can help enterprise customers achieve major results and bring substantial improvement to enterprise revenue.

From this perspective, among Tencent, Alibaba and ByteDance, which one is most likely to achieve this goal?

This article is from the WeChat official account "Deep Flow Institute", written by Wu Jiangfeng, and authorized for release by 36Kr.