Office agents are no longer competing on who can create the best PowerPoint presentations; they're now vying to prove "who knows the company best."

At the Yunqi Conference on September 22, Qianwen Office released six enterprise capabilities at once.

The most eye-catching feature is the "digital employee": it has a name, department, manager, job responsibilities, and scope of authorization; it can be deactivated, and its execution records are traceable. It seems that AI has finally transformed from a tool floating on a desktop into a colleague with an employee badge within the organizational structure.

However, if you only focus on this "employee badge," you might easily miss the four more important words mentioned at the press conference: corporate context.

According to Qianwen Office's description, it connects group chats, documents, knowledge bases, and business systems to identify people, projects, processes, and rules, and then retrieves relevant information based on specific tasks. In other words, whether digital employees can perform their tasks increasingly depends on whether they have a continuously updated company memory behind them.

This means that the competition among office agents has entered a new phase.

In the past, people competed on who could make PPTs, organize spreadsheets, and generate reports; now, the real difference lies in who knows what happened in the company, what rules it operates under now, and who to contact for a specific issue.

The first round of competition felt like testing a complete stranger for a temporary job.

Over the past year, the most common presentation method for office agents has been to give them a document and ask them to complete a task: summarize sales data, create a presentation slide, find publicly available information, and organize meeting minutes.

This type of test is intuitive and easy to share. The task has a clear start and end point, and the result can be seen in just a few minutes.

However, it primarily measures model, tool usage, and file operation capabilities. In a real company setting, the issues quickly become more complex.

Does a refund count as quarterly revenue? Should the unusual order from North China be included? How should we consolidate data if the same customer is listed under two names in both the CRM and financial systems? In the report, does "key customer" refer to contract amount, payment received, or strategic level?

These questions are difficult to answer using public knowledge and generic models, because the answers lie hidden in the company's history, its narrative, and its power and responsibility relationships.

A recent comparative review of office service agents exposed this problem. Using the same sales record containing missing departments, refunds, cross-quarter orders, and unusual amounts, the three products calculated quarterly totals that differed by up to 57.7%. A significant portion of this discrepancy stemmed from their different approaches to handling suspicious orders.

These figures don't necessarily mean that any particular product is worse. What they really illustrate is that while office tasks may appear to involve calculating spreadsheets, at a deeper level, they involve making judgments on behalf of the company.

An agent may be able to read cell data, but may not know the financial definitions; may be able to find anomalies, but may not know how the company handled them in the past; may be able to deliver a complete report, but may not know who has the authority to decide which order to include.

As the capabilities of the models become increasingly similar, the new dividing line begins to emerge when it comes to who can obtain and correctly use this internal information.

"Understanding the company" isn't about stuffing in more documents.

Enterprise context sounds like an upgraded version of a knowledge base, but it actually adds an extra layer of relationships.

A typical knowledge base addresses the question of "what information does the company have?" A business context also needs to answer: which project does this information belong to, who is responsible for this project, what rules apply to the current task, which information is outdated, who has the right to view it, and where should the results ultimately be written back to?

Qianwen Office proposes a path that is divided into connection, understanding, and reuse.

First, integrate group chats, documents, knowledge bases, and business systems; then organize the information into entities such as people, projects, process rules, etc.; finally, load the data as needed based on the task, instead of stuffing the entire company into the model.

The value of this design is easily understood in the case of Gu Ming Tea. Gu Ming organized Lark documents, knowledge bases, Q&A groups, and training schedules into a knowledge space for store operations, accessible to nearly 10,000 frontline employees. Content such as recipes, equipment inspections, and table opening and closing procedures were broken down into entities and relationships, with different positions retaining different permissions.

When a store clerk inquires about poster display guidelines, the system needs to do more than just "search for posters".

It needs to know who the questioner is, which store they belong to, what position they hold, which versions of the regulations apply, and which section in a bunch of training materials is relevant to the current question.

This is also the difference between enterprise context and long context window. The model can read a certain number of words at a time, which solves the capacity problem; enterprise context solves the problem of which information in the company is related to each other and which rules should be in effect at this moment.

The GuMing case study currently does not disclose efficiency, accuracy, or operational improvements, making it suitable for illustrating product form but insufficient to prove effectiveness. However, it demonstrates a direction: office agents are shifting from "processing documents after receiving them" to "staying within the organization long-term to understand the business."

The new moat grows within the company's daily collaboration.

Models can be changed, but the enterprise context is difficult to rebuild overnight.

An agent who truly understands the company needs to accumulate organizational relationships, project records, business terminology, historical decisions, and employee feedback over a long period. They also need to adapt to company changes: employee transfers, project renamings, adjustments to approval rules, and changes in client levels will all impact subsequent tasks.

This gives the office platform a natural place.

DingTalk, Lark, and WeChat Work originally served only as platforms for communication, documents, and processes. With the introduction of agents, these scattered records began to become raw materials for understanding the enterprise. Those closer to the organization's daily collaborations have easier access to fresh, continuous, and authorized information.

Therefore, the competitive focus of office agents is shifting.

Early competition focused on whether models could be written, tools could be debugged, and tasks could be completed; the next stage will focus on whether the platform can establish a continuously updated enterprise context and allow different agents to use it together.

This will also change business stickiness.

In the past, when companies switched office software, the biggest headaches were transferring files, changing processes, and training employees. In the future, they will also need to migrate a layer of more intangible assets: the relationship between people and projects, the company's own business vocabulary, the rules used by agents, preferences formed by historical feedback, and already validated task paths.

While files can be exported in batches, organizational relationships and decision-making contexts are difficult to package completely.

For manufacturers, this is a deeper moat than simple model capabilities; for enterprises, it may also become a new lock-in cost.

The more context there is, the greater the responsibility becomes.

"Knowing the company" does not automatically equate to "doing the right thing".

Qianwen Office states that the enterprise context inherits the organization, identity, and permission system of the enterprise within the instant messaging system. This design allows the agent to prune permissions before retrieving data, while also addressing existing problems.

A group that hasn't been maintained for years, a document with overly broad permissions, or an account that has left the company but is still in the project space—these were previously just potential problems with access control. However, within an enterprise context, these errors can become information that the agent can actively access to complete tasks.

Incorrectly established relationships, expired rules, and contradictory historical data will also be included in the answer. The agent, knowing more, may be more confident in executing an invalid rule.

The previous sales flow test also reminded us of this point: obtaining more data only solves the problem of "seeing"; how to identify abnormal orders still requires companies to provide clear definitions and ensure that someone is responsible for the final judgment.

Therefore, the quality of an enterprise context cannot be measured by the number of documents accessed. More valuable metrics are: whether the source of information can be traced, whether permissions are updated according to the person and position, whether business rules have a responsible person, and whether the agent continues to guess or returns the problem to the person when conflicts occur.

A truly mature corporate context is both a knowledge system and a responsibility system.

Companies don't choose the smartest assistants, but rather the most controllable corporate memories.

When companies select office agents, real-world testing remains important, but the testing methods need to be further developed.

Instead of providing all three products with the same Excel spreadsheet, give them a real task that requires cross-system collaboration: identify project changes in a group chat, confirm constraints in a contract, verify customer identities in the CRM, and then generate the next steps according to the latest approval rules. During testing, temporarily change the person in charge, revoke access to a document, or update a business rule to see if the system can reflect the changes promptly.

Such testing is more complicated, but it gets closer to what businesses actually buy.

Businesses are not simply buying an agent that can operate computers for employees; they are purchasing an infrastructure that transforms organizational information into action. This infrastructure needs to be business-savvy and also inform the business why it is doing what it is doing, what information it is using, and who is responsible for the results.

At the same time, enterprises need to maintain a portable set of contextual assets: business lexicon, permission models, process rules, decision logs, and acceptance samples. These should belong to the enterprise, not just exist within a single office agent product.

The first round of the office agent competition tested who could complete the task at hand quickly, like a skilled temporary worker.

The second round of the competition tested who could stay in the company long enough to understand the relationships and rules that were never fully written into any of the rules and regulations.

In this round, the strongest model is not necessarily the product that best understands the company. What companies really need to protect is not just documents and data, but the company's language that any agent can understand.

This article comes from the WeChat official account " AI Native Lab " , which continuously analyzes real-world AI implementation cases and shares enterprise AI practices and methodologies.

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.