Author: Ren Tongyao, Editor: Li Motian
Dario spreads anxiety, Altman is left dizzy and slumped in his seat, and large models take the spotlight for a few days at a time. Amidst a wave of FOMO, every business owner who has a love-hate relationship with AI has realized: it is time to take down the plaque that says "Heaven rewards the diligent" and sign up for a crash course on enterprise Agents.
Enterprise Agents are the next big market after coding, a consensus highly unified on both the supply and demand sides. Therefore, after fighting the AI office war with elevator ads on the eastern front, the battle has swiftly shifted to the core western front: getting bosses to foot the bill.
However, when bosses open their budgets and install Agents on every employee's desktop, they find that improving company efficiency seems to be stuck at the "last mile." The first batch of companies to catch the trend have all realized a problem:
An Agent that helps employees write weekly reports is not necessarily an Agent that can do the actual work well.
At the Apsara Conference on September 22, Chen Yusen, CEO of Qwen Office, mentioned this issue: "For Agents to land in real enterprise scenarios, they must not only understand the business but also collaborate within the enterprise and organization like humans, and every task execution must have boundaries and be traceable."
As one of the participants in the office war, Qwen Office did not take the route of heavy venture capital investment, but instead focused its main energy on the rear. At the Apsara Conference, Qwen Office released a series of solutions in one go, including Agent hosting capabilities, digital employees, collaboration, applications, and a security center, all pointing to one question:
How exactly should a real enterprise Agent work?
What Is Your Context?
Among the numerous solutions released by Qwen Office, the focus is probably not the Agent itself, but the "context issue."
What is context? It refers to the specific context in the workplace.
For example, when a boss says, "Let Xiao Chen handle this project," the context of this decision includes who Xiao Chen is, what the project is, what other projects Xiao Chen has done before, and why Xiao Chen is suitable for this project, and so on.
Chen Yusen gave an example at the press conference: asking what the gross profit margin of the East China region was last quarter? No Agent can answer this, because the Agent doesn't even know what company it is working for, making it impossible to start anything. It's like hiring a researcher without giving them an access card, so they can only receive packages at the front desk.
This is actually the last mile for Agents from personal use to enterprise use:
Chatbot products can solve "general" problems, but once they enter the "specific" context of an enterprise, they fall short without context. For example, if you ask how many days of annual leave the company offers, and the Agent cannot see the company's management regulations, it can only thoroughly review the labor law and give a standard answer.
Currently, the most widely used applications of Agents, such as writing weekly and daily reports and making PPTs, are characterized by very thin contexts; uploading a few documents and writing a few prompts is enough for the Agent to handle it. But the context of a scaled enterprise has an unimaginable level of complexity.
For instance, the data caliber issue in catering enterprises can stump a whole bunch of Agents: is the sales revenue for the store or same-store? Does it include dine-in or takeout? Which one to choose when different departments have different data time points? If the context is not completed, the Agent will run out of memory on the spot.
When individuals use AI, the model can automatically record the user's habits into long-term memory. But the "enterprise memory" in an enterprise organization is very complex.
Deloitte explicitly pointed out the issue of "fragmented context" in a previous report. Enterprise context is accumulated in ERP systems, and business logic is embedded in ABAP code and employees' daily communications. For an Agent to understand a company's knowledge is a massive project in itself.
This leads to a new problem: having context is not enough; this context must also be callable by the Agent.
Writing a weekly report for an individual is nothing more than uploading a few documents. Asking an enterprise organization to turn fragmented context into structured documents and data—even if employees are willing to do it, by the time it's actually done, ASI will have been realized three times over.
Model capabilities can be tailored to the business, ranging from lavish to frugal; Harness capabilities will gradually converge among various Agents. All participants in the office war have realized that to break into the enterprise, context is the key. As shown on Chen Yusen's PPT: Context is all you need.
Therefore, at the Apsara Conference, Chen Yusen immediately highlighted the focus of Qwen Office: a product named Enterprise Context.
Let's take a look!
Software used by humans and tools used by AI are literally two different species.
The buttons, menus, and views of SaaS software perfectly match human interaction habits, but Agents need structured, callable, and verifiable interfaces. However, SaaS software in large companies has been running for decades and cannot afford to make drastic overhauls. Whether Agents can call them smoothly is the key to improving efficiency.
In other words, the relationship between Agents and SaaS is not a zero-sum game, but one of mutual integration. The connection between the two is a key step in integrating the context.
A signal was conveyed at the Apsara Conference: Qwen Office is advancing cooperation with the two global software giants, SAP and Salesforce, becoming their first officially partnered enterprise Agent partner in China. A large amount of enterprise context can be activated by Agents.
After solving the source of the context, there is still the problem of how to bridge the gap between the "context accumulated by humans" and the "context that Agents can call."
For most legally operating enterprises, the context basically has two characteristics:
First, it is fragmented. Even in enterprises with a strong document culture, the context is scattered in various corners and lacks a systematic structure, including but not limited to structured data in SaaS software, impromptu discussions in group chats, ad-hoc meetings initiated by clients, and review summaries in chat records, etc.
Just like in history, besides official historical records, many daily texts such as administrative documents, personal diaries, and local gazetteers are also of great value for historical research. How to collect these fragmented historical materials is the pain point of enterprise-level context.
Second, the scale is massive. Chen Yusen gave an example: based on the data accumulated by enterprises on DingTalk, the internal data scale of medium to large enterprises is around 150PB, which even the most powerful context window cannot fit.
In other words, the context of any enterprise with a bit of history is a literal, overwhelmingly stacked historical burden. Asking employees to proactively organize massive, scattered data into structured data—let alone whether it can be completed—if they had such perseverance and boldness, they could succeed without needing an Agent.
Addressing these two issues, Qwen Office's enterprise context product has prescribed a targeted remedy: connection and compression.
The intention behind connection is to integrate the knowledge accumulated by the enterprise with Qwen Office, systematically connecting the know-how in documents, meetings, and chats, turning it into an enterprise context that can be flexibly called by Agents.
After the two are integrated, Qwen Office can retrieve data from the context according to permissions, truly understanding enterprise knowledge and completing tasks like a human.
Guming processes raw materials structurally through Enterprise Context
The purpose of compression, on the other hand, is to systematically store fragmented knowledge, establish relational links between data that can be understood by Agents, and further improve calling efficiency.
Data is the cornerstone of all intelligence. Whether data is stored on servers or on A4 paper directly determines the digitalization level of an enterprise. The digitalization construction of moving organizational and business processes online has only been around for a few years at most in the industry.
What Qwen Office's enterprise context product actually accomplishes is the "prerequisite" for Agents entering the company—the construction of infrastructure.
Once the infrastructure is built, can enterprises just move in? The answer is not yet.
Who Let It In?
Grab any AI trendsetter at random and ask them what Agent capability they care about most, and the answer is highly likely to be the model.
When individual users use Agents, if it answers incorrectly, they can just rephrase the question and start over at almost zero cost, so the evaluation focus is on the product and the model; however, enterprise context is scattered across five departments and three systems, and getting a single number wrong is a compliance incident, so what enterprises really care about is reliability.
This reliability has two very specific measurement indicators:
First is Token quota consumption. A bizarre trend of Token leaderboards swept through big Silicon Valley companies; Meta once launched a leaderboard to internally compare who uses AI the most aggressively. As a result, the top user burned through USD 500,000 in a single month.
Operating cost is a key metric for enterprises. If employees use the highest-intensity reasoning mode of GPT-6 Astra just to polish copy, which capitalist wouldn't have a heart attack looking at that?
Second is data security. Any legitimate enterprise will strictly manage employees' use of personal AI tools, and the reason is data security. You might be willing to spark a vigorous AI revolution in the company, but the company's risk control department might not be willing to join you in the detention center.
Therefore, a standalone Agent product cannot meet the needs of enterprises; what they truly need is a complete and reliable solution. Without safety valves, letting Agents in at will—no matter how many points are given away, bosses wouldn't dare to accept them.
Therefore, beyond Enterprise Context, Qwen Office has prepared solutions composed of different products. Only after setting up access control and installing cameras will bosses feel safe enough to foot the bill.
For example, Managed Agents are responsible for configuring models, instructions, tools, and runtime environments, enabling enterprises to uniformly define and manage different Agents; the Security Center is used to uniformly manage Agent data access, tool invocation, and task execution, keeping AI from running wild in the database.
The Digital Employee and Collaboration features point to specific business scenarios; the former gives Agents clear identities, positions, permissions, and responsibilities, while the latter brings together employees, digital employees, and project materials in the same space.
Qwen Office launches the "Collaboration" feature
In this way, the individual-Agent interaction becomes organization-Agent, and the positioning of the Agent can shift from assistant to colleague. As for whether it needs to become a boss, let's not go there just yet.
Paired with the QwenNote A2 hardware, it can directly connect with enterprise Agents in offline scenarios such as meetings and client visits, automatically collecting context and expanding the boundaries of Agent execution capabilities.
Based on the context infrastructure, applications connect to business systems, digital employees connect to organizational positions, collaboration spaces connect interpersonal collaboration, and QwenNote hardware connects offline scenarios.
A press conference at the Apsara Conference has given enterprise Agents a real prototype.
Holding the Position
The AI office war is a vigorous industry war that China's Internet industry has reignited after many years. The reason is not hard to understand: the market is just too big.
Specifically, Agents will reshape the enterprise market landscape in China, which is basically an industry consensus. Earlier this year, the US capital market was discussing "SaaS is dead," while China was as quiet as a mouse. The reason is simple: China's SaaS has been lying in the ICU and has never really been alive.
Historically, domestic companies have kept their distance from annual subscriptions for standardized products. But in the Agent era, the industry has generally formed an understanding: when Agents can call business systems and participate in organizational collaboration, the business model may shift to paying based on the number of Agents, task execution volume, or business outcomes.
At the same time, the cooperation between Qwen and SAP and Salesforce also indicates that enterprise Agents and SaaS software may form a new way of collaboration: professional software handles the business, while Agents bridge cross-system tasks, allowing employees to avoid repeatedly switching software, searching for information, and transferring data, and instead advance work and obtain results more directly.
The status of SaaS software in the Agent era will not decline, and may even continue to rise.
As more and more SaaS vendors and service providers connect, enterprise Agents may also form an application and service ecosystem, opening up new revenue spaces through the exchange of models, tools, digital employees, and industry capabilities.
On the other hand, the stickiness of the enterprise market cannot be compared to that of the consumer market.
Generally speaking, large enterprises do not easily change suppliers, especially when Agents have accumulated a large amount of enterprise context and deployed extensive collaboration between the organization and Agents. When AI is deeply linked with the enterprise's way of working, the stickiness of enterprise users will become stronger and stronger.
Therefore, for tech giants, losing an enterprise order very likely means losing revenue for many years to come. The cost is extremely high, leading to the AI office war being fiercely contested.
But one thing is certain: to impress enterprise bosses, you cannot rely on handing out red envelopes or giving away points. It requires testing product capabilities, ensuring data security, and deeply integrating with the enterprise's existing systems (such as ERP and CRM).
As various tech giants successively unveil solutions from their arsenals, the office war may have truly entered the main battlefield. To paraphrase Churchill:
This is not the end. It is not even the beginning of the end. But it is, perhaps, the end of the beginning.
References
[1] SaaS Applications “Will Collapse” In The AI Agent Era: Microsoft CEO Satya Nadella, S&P Global
[2] AI Productivity Tools Market, Market.us
[3] Superagency in the workplace: Empowering people to unlock AI’s full potential, McKinsey