At the Yunqi Conference on September 22, Qianwen Office unveiled six enterprise-grade capabilities in a single release.

The most attention-grabbing feature is the "digital employee," complete with a name, department, supervisor, job description, and authorization scope. It can be deactivated, and its execution records are fully traceable. This signals that AI is no longer just a floating desktop tool but has become a colleague with an employee badge integrated into the organizational structure.

However, focusing solely on this "badge" means missing the two more critical words from the press event: enterprise context.

According to Qianwen Office, its system connects group chats, documents, knowledge bases, and business systems to identify people, projects, processes, and rules, then retrieves relevant information based on specific tasks. In other words, a digital employee's effectiveness increasingly depends on whether it has access to a continuously updated corporate memory.

This development marks a new phase in the competition among office agents.

Previously, the benchmark was who could create better presentations, organize spreadsheets, or generate reports. Going forward, the real differentiator will be which system understands what has happened within a company, how it currently operates, and who should be approached for any given matter.

The First Round: Testing Like a Stranger Hired for a Day

Over the past year, the most common demonstration of office agents involved handing them a document and asking them to complete a task: summarizing sales data, building a presentation, searching public information, or taking meeting notes.

These demos were straightforward and easy to share. Tasks had clear starting and ending points, and results appeared within minutes.

But such tests primarily measure model capability, tool invocation, and file handling. In a real company, the challenges quickly become far more complex.

Should a refund count toward current-quarter revenue? Should an unusual order from the North China region be included? When the same customer appears under different names in the CRM and financial systems, how should they be consolidated? Does "key account" in a report refer to contract value, collection amount, or strategic tier?

These questions cannot be answered using public knowledge or general models because the answers lie within a company's history, accounting conventions, and lines of authority.

A recent head-to-head evaluation of office agents highlighted this problem. When given the same sales data containing missing departments, refunds, cross-quarter orders, and abnormal amounts, three products calculated quarterly totals that differed by as much as 57.7%. A large portion of the variance came from how each handled suspicious orders.

These numbers do not necessarily indicate that one product is worse than another. What they truly reveal is that office tasks are superficially about number-crunching but fundamentally about making judgment calls on behalf of the business.

An agent can read a spreadsheet cell but may not understand financial conventions. It can spot anomalies but may not know how the company historically handled them. It can deliver a complete report but may not know who has the authority to decide whether a specific order should be included.

As model capabilities converge, the ability to access and correctly use this internal information becomes the new dividing line.

Understanding the Company Is Not Just Adding More Documents

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

A standard knowledge base answers the question of what materials a company has. Enterprise context must also address: Which project do these materials belong to? Who is responsible for this project? What rules apply to the current task? Which information is outdated? Who is authorized to see it? And where should the final results be written back?

Qianwen Office's approach is structured in three stages: connection, comprehension, and reuse.

First, it integrates group chats, documents, knowledge bases, and business systems. Then, it organizes the information into entities such as people, projects, and process rules. Finally, it loads information based on the task at hand rather than feeding the entire company to the model.

The value of this design is easier to grasp through the case of Guming Tea. Guming organized Feishu documents, knowledge bases, Q&A groups, and training schedules into a store operations knowledge space serving nearly 10,000 frontline employees. Content such as recipes, equipment inspections, and opening/closing procedures was broken down into entities and relationships, with different roles retaining different permissions.

When a store employee asks about poster placement guidelines, the system needs to do far more than "search for a poster."

It must identify who is asking, which store they belong to, what role they hold, which versions of the guidelines apply, and which section of the training materials is relevant to the current question.

This also illustrates the difference between enterprise context and long-context windows. How many tokens a model can read at once solves a capacity problem. Enterprise context solves the problem of which pieces of company information are related to each other and which rules should take effect at a given moment.

The Guming case has not publicly disclosed efficiency, accuracy, or operational improvements. It is suitable for illustrating the product concept but not yet sufficient to prove effectiveness. Still, it points to a direction: office agents are moving from "processing files you hand them" to "staying within the organization long-term to understand the business."

The New Moat Grows in Everyday Corporate Collaboration

Models can be swapped out, but enterprise context cannot be rebuilt overnight.

A truly company-savvy agent requires long-term accumulation of organizational relationships, project records, business vocabulary, historical decisions, and employee feedback. It must also adapt to changes: employee transfers, project renames, approval rule adjustments, and changes in customer tiers all affect subsequent tasks.

This gives office platforms a natural advantage.

DingTalk, Feishu, and WeCom originally carried communication, documents, and workflows. Once agents enter the picture, these scattered records become raw material for understanding the enterprise. Whoever sits closer to daily organizational collaboration is more likely to obtain fresh, continuous, permission-aware information.

Therefore, the focus of office agent competition is shifting.

The early competition centered on whether a model could write, whether tools could be invoked, and whether tasks could run to completion. The next phase looks at whether a platform can build a continuously updated enterprise context and allow multiple agents to share it.

This will also change commercial stickiness.

In the past, switching office software was painful because of file migration, process changes, and employee training. In the future, there is an additional layer of invisible assets to migrate: relationships between people and projects, the company's own business vocabulary, rules that agents have used, preferences formed through historical feedback, and validated task paths.

Files can be batch-exported, but organizational relationships and decision-making context are difficult to package completely.

For vendors, this is a deeper moat than pure model capability. For enterprises, it may also become a new lock-in cost.

More Context Means Greater Responsibility

Understanding the company does not automatically mean doing things correctly.

Qianwen Office states that enterprise context inherits the organization, identity, and permission structure from the instant messaging system. This design allows agents to apply permission filtering before accessing data, but it also inherits existing real-world problems.

A group that has not been maintained for years, a document with overly broad permissions, or an account belonging to a former employee still present in a project space—these were previously just latent risks in permission management. Once incorporated into enterprise context, these errors can become information that agents actively retrieve when completing tasks.

Incorrect relationships, outdated rules, and contradictory historical materials will all be carried into answers. An agent that knows more may also execute a rule that is no longer valid with greater confidence.

The earlier sales data test also serves as a reminder: having more data only solves the visibility problem. How abnormal orders are classified still requires the enterprise to provide clear conventions and assign accountability for the final judgment.

Hence, the quality of enterprise context cannot be measured by how many documents are connected. More valuable indicators include: whether information can be traced to its source, whether permissions update with people and roles, whether business rules have owners, and whether the agent continues guessing or escalates to a human when conflicts arise.

A mature enterprise context is both a knowledge system and an accountability system.

Enterprises Are Choosing Controllable Corporate Memory, Not the Smartest Assistant

When enterprises evaluate office agents going forward, hands-on testing remains important, but the testing method needs to advance.

Instead of handing the same Excel file to three products, consider giving them a real task that requires cross-system work: finding project changes from group chats, confirming constraints from contracts, verifying customer identity in the CRM, and generating next actions according to the latest approval rules. During the test, change the responsible person mid-task, revoke a document permission, update a business rule, and see whether the system reflects the changes in a timely manner.

Such tests are more cumbersome but far closer to what enterprises are actually purchasing.

Enterprises are not just buying an agent that can operate a computer on behalf of employees. They are buying infrastructure that converts organizational information into action. It must understand the business deeply while also letting the enterprise know why it acted a certain way, what information it used, and who is accountable for the outcome.

At the same time, enterprises need to maintain a portable context asset: business vocabulary lists, permission models, process rules, decision records, and acceptance samples. These should belong to the enterprise, not exist solely within a single office agent's product.

The first round of office agent competition tested who could work like a skilled temporary worker, quickly finishing tasks at hand.

The second round tests who can stay within a company long enough to understand the relationships and rules that have never been fully written into any policy document.

At this stage, the most powerful model may not be the product that understands the company best. And what enterprises truly need to protect is not just files and data, but the corporate context that allows any agent to understand them.