最近有两条新闻放一起看，挺有意思。
One is Meituan. Its internal AI workbench called CatPaw, tested by 90,000 people over three years, was officially upgraded to a full-scenario AI Agent platform on July 27. There was a sentence in the report that I read several times: "A dialog box cannot contain real work".
The other track is the Xinchuang circle. Pulian Software has implemented its intelligent agent development platform — knowledge Q&A assistants and document intelligent processing assistants are directly embedded into the three Xinchuang products: ERP, EAM, and EPM, and all have produced demonstration projects.
One focuses on local life services, the other on enterprise management software—two businesses that seemingly have nothing to do with each other, yet they arrive at the same conclusion: for AI to truly get things done, being able to chat is far from enough. What exactly is holding it back behind the scenes is worth breaking down and discussing.
1. Four Hurdles: Why Chatbots Can't Get the Job Done
Meituan broke down the gap between conversational AI and execution-oriented agents into four pain points. After reading them, I felt these four points apply not only to local life services but to any enterprise scenario:
First, lack of context. A general-purpose large model stuffed into a chat window has no awareness of your business—store data, historical reviews, operational rules, inventory limits—it knows none of it. Whatever you ask, it answers, but it is always guessing your business from scratch. The same applies to ERP: the customers, orders, materials, and process rules in the system are the prerequisites for AI to answer "should this order be approved." Without these, AI is just an intern with no memory.
Second, it cannot call tools. Conversational AI can only talk, but cannot use its hands. To check inventory, initiate approvals, or generate documents — these actions it cannot perform, and humans still have to do them. The essential difference between an Agent and a chatbot lies here: an Agent can call tools and interfaces within the system, turning "saying it will be done" into "actually doing it."
Third, there is a lack of sustained execution capability. Generating contracts, reviewing risks, organizing feedback, and outputting improvement plans—these are not tasks that can be completed with a single sentence; they require multi-step, cross-module, and even cross-system advancement. Conversational AI stops after answering, but an Agent can stay focused on a goal from start to finish, only coming to ask for instructions when encountering anomalies.
Fourth, results are difficult to reuse. Once a conversation ends, the results are lost. Real work outcomes need to be deposited back into the system—documents entered into accounts, processes approved, data incorporated into reports—so that next time, work can continue on that basis. The product of Chat is conversation; the product of Agent is business records. This is a qualitative difference.
These four pain points are precisely the root cause of "AI looks good on the surface but is hard to use in practice." Many enterprises fail in AI projects not because the models are not powerful enough, but because they simply fail to realize that for AI to get things done, four things must be in place: business context, tool interfaces, execution orchestration, and result accumulation. Missing any one of these, and AI degenerates into an advanced chatbot.
II. Why ERP is Actually the Main Arena for AI
聊到这你可能会问：那到底什么样的系统，AI 才能真正上手干活？答案可能出乎意料——恰恰是那些被吐槽"又重又复杂"的 ERP。
The reason is simple: ERP inherently possesses the four things mentioned above. It has full business context (customers, orders, inventory, finance), complete tool interfaces (business processes are the toolchain), built-in execution orchestration (approval flows and workflows are ready-made), and comprehensive result accumulation (all operations are recorded). What ERP lacks is not the soil for work, but the brain to do the work. Connecting an Agent to ERP is equivalent to installing an upper layer that can "understand, judge, and execute" onto a system that originally could only record.
The implementation path of Puliang Software serves as a standard model: the knowledge Q&A assistant is responsible for "guiding employees on how to use the system," while the intelligent document processing assistant handles "digesting unstructured materials such as contracts, documents, and regulations." Both are directly embedded in the Xinchuang ERP. For enterprises pursuing domestic substitution, this signal is particularly important — in the past, there were concerns that "after replacing foreign systems, AI capabilities might shrink accordingly," but now Xinchuang vendors are answering through action: AI follows the business, not the brand.
Odoo actually touched this line earlier. At the end of July, Odoo officially released a batch of overseas cases, among which the precision electromechanical company Motek was notable not for AI, but for "data integration" itself—orders, inventory, procurement, manufacturing, and finance all running on one unified base, with the error shipment rate dropping by 60%. Why mention this? Because this is the prerequisite for AI to work. If data isn't on the same line, AI can't even calculate the "error shipment rate," let alone help you reduce it. Only after "integration" can we talk about "intelligence."
III. From "Asking AI" to "Directing an AI Team": A Four-Level Climbing Path
The technology foundation is a prerequisite, but the organizational side is equally challenging. TopTech released a model a couple of days ago, viewing the implementation of enterprise intelligent agents as a progressive course from 101 to 401. I quite agree with this, so I'll relay it directly to you:
| Stage | The Relationship Between Employees and AI | Typical actions |
|---|---|---|
| 101 | Seek help from AI | Q&A, search, document summarization |
| 201 | Collaborate with AI to complete tasks | Human-machine division of labor, result verification |
| 301 | Authorize AI to independently complete work | Preset rules, exception reporting |
| 401 | Coordinate multi-agent orchestration | Cross-system, cross-process collaborative operations |
What are the benefits of this framework? It turns "AI transformation" from a muddled account into a checklist that can be ticked off level by level. I have seen too many enterprises that want to jump straight to 301 and 401—AI working independently and multi-agent collaboration—only to find that no one in the organization knows how to verify AI output, the rules are not clearly defined, and failure is inevitable. This is the same principle as the "Three-Step Method for AI Deployment" (knowledge foundation → process embedding → continuous growth) shared by NetEase Zhiqi at WAIC in July: First, teach employees to use it, then let AI learn to do the work, and only then talk about letting the AI team operate independently.
One-sentence reminder: The more capable AI becomes, the greater the destructive power of its mistakes. Before the 301 stage, be sure to install the three gates: permission boundaries, exception reporting, and manual approval. This has nothing to do with the system—it's about management. Gartner predicts that by the end of 2027, over 40% of agentic AI projects will be cut, and most of those cut are not due to technical shortcomings, but because the gates weren't installed properly.
Conclusion
At this point, let me distill the core conclusions of these two news items:
For platform-based companies (like Meituan and Alibaba), the decisive factor for Agent lies in "business context"—whoever has a deeper grasp of scenario data, tools, and execution chains will take the lead. For enterprise management software (ERP vendors and domestic IT vendors), the decisive factor for Agent is "connecting the underlying data foundation first"—Motek's 60% reduction in misdelivery rates was not driven by AI, but by a unified data pipeline. As for business leaders planning AI projects, what truly belongs on the timeline, beyond system upgrades, is a roadmap for elevating employee capabilities from level 101 to 401.
The real work can't fit in a dialog box. That may sting to hear, but for any company looking to boost efficiency with AI, it's a particularly sobering starting point. First, figure out where your business context lives, whether your data flows smoothly, and whether your employees know how to use it—once you've thought through these three things, deploying Agents will come naturally; if you haven't, it won't matter how many models you switch between.
