The trend of implementing AI agents (OpenClaw) is blowing too strong. Not only are individuals experimenting, but companies are also rushing to deploy them, fearing to fall behind.
But the common situation is:Easy to install, hard to use effectively. An e-commerce company asks the agent about last month’s return rate, and it goes searching online; ask it to draft a payment reminder email, and it doesn’t even know how much the client owes. Such an agent is no different from a pet in a fish tank, only good for show.
Many say the reason enterprise AI agents aren’t performing well is that the models aren’t powerful enough or the Skills are poorly written. That’s not entirely correct.Even the smartest agent, even if it masters the art of dragon-slaying, can’t accomplish much if it lives in a fish tank.
Where the agent resides is more important than the agent itself.
In this article, I will discuss how to make AI agents truly valuable from the perspective of enterprise Agent implementation. I’ll use Feishu as a case study, not because it’s perfect, but because among the enterprise-level solutions I’ve seen, Feishu has done the most complete job in “providing a working environment for the agent.”
Personal vs. Enterprise AI Agent Implementation: Two Completely Different Things
For my personal use of an AI agent, it’s just about sending a link for it to scrape and translate an article, or generating an image for an article. I don’t use it much. This is the personal playstyle, and it’s sufficient.
Enterprises are completely different.
Enterprises need the agent to understand the business. Not “help me write an email,” but “draft a notice to Client A regarding the Q2 delivery delay according to the framework agreement signed last quarter, referencing the tone of the previous email template sent to Client B.” To complete this task, the agent needs to know the content of the framework agreement, Client A’s delivery timeline, and the tone/style of Client B’s email. What’s needed behind this is not a smarter model, butan entire chain of data, relationships, permissions, and business rules.
Enterprises need the agent to be secure. For personal use, guarding bank cards and private information, confirming critical steps, can prevent most issues. Enterprises are different: data needs isolation, permissions need hierarchy, operations need logging, and issues need traceability.
Enterprises also need the agent’s experience to be transferable. Today we use Claude, next year we might switch to GPT-6 or some domestic model. If business knowledge is scattered in prompts and people’s minds, changing personnel or models resets all accumulated knowledge to zero.
In a nutshell: Personal implementation is installing a tool; enterprise implementation is building a knowledge system.
What Should Enterprises Actually Prepare for the Agent?
Knowing that enterprise implementation is different from personal use, the next question is: What exactly does a knowledge system refer to?
For example. You ask the agent to “check the progress of Manager Zhang’s project.” What does the agent need to do? It first needs to know who “Manager Zhang” is (data), then find the projects Manager Zhang is responsible for (relationships between data), then check the progress in the project management system (operation), and finally organize the results into your desired format and send them to you (skill).
Four steps. Missing any step makes it impossible. The knowledge an enterprise needs to accumulate, broken down, consists of these four layers:
-
• First layer: Data.
Customer information, product materials, business documents, historical messages. This is the most basic raw material. -
• Second layer: Relationships between data.
“Which contract did this customer sign?” “Which factors influence this metric?” “Which departments does this approval process depend on?” Isolated data isn’t very useful; connected data has value. -
• Third layer: Operations.
Query, update, approve, send notifications. Different roles can perform different operations. These operational rules themselves are also part of enterprise knowledge. -
• Fourth layer: Agent Skills.
Encapsulating the first three layers into a form that AI can understand and execute. “Upon receiving a payment reminder email, automatically query the corresponding contract amount and draft a reply” – this is a skill.
I tend to include tools within skills, because an Agent needs Skills to know how to use tools.Skills are the tip of the iceberg; the bottom three layers are the main body of the iceberg.
This also explains why many enterprises struggle to write good Skills alone. The quality ceiling of a Skill depends on the completeness of the underlying three layers. The Agent cannot directly access the underlying data and relationships; it can only reach them through Skills and tools.If the foundation is empty, even the most exquisite Skill will just spin its wheels.
Three Things Enterprises Must Do for AI Agent Implementation, Regardless of Platform
Having covered the framework, let’s talk about implementation. Whether you use Feishu, DingTalk, or a self-built system, three things are common.
First: Extract the enterprise’s data from people’s minds
Critical information in many enterprises resides in three places: someone’s mind, some chat history, or an Excel file on someone’s computer. This information doesn’t exist for AI. The first step is to gather scattered business data into a place accessible to AI.
No need to achieve everything at once. Start with a specific scenario, such as customer information lookup, contract management, or project progress tracking. Structure and organize the data involved in this scenario. Once one scenario works, expand to the next.
Second: Make operational rules explicit
“Who handles this approval?” “Who can view this data?” “What process is followed in this situation?” These rules often only exist in the experience of veteran employees. If they are not turned into configurable permissions and executable workflows, the Agent will never know what it can and cannot do.
This isn’t preparation for AI; it’s something the enterprise should have done anyway. AI is just forcing you to catch up on what you should have done.
Third: Detach knowledge from specific models and specific people
The prompts written today, the Skills configured, the workflows tuned – if they are all tied to one specific model, or only one person knows how to maintain them, that’s fragile. Knowledge should reside in a system with version control, permission management, and traceable operations, usable regardless of model or personnel changes.
Models are rented; knowledge is your own. Models are already very intelligent and have sufficient capability to help you structurally solidify your enterprise’s knowledge. Once knowledge accumulation is in place, the requirements for the model actually decrease. It’s like having a detailed city map drawn; anyone can drive and find the way.
Why Feishu Happens to Fit Well
After discussing the general methodology, let’s talk about Feishu. Why are OpenClaw developers spontaneously flocking to Feishu? Yang Mingfeng, founder