LINE 企業內部助理Internal company assistant on LINE
為虛構的批發商「青禾」設計的 LINE 內部助理概念：員工在 LINE 上詢問訂單與庫存，AI 代理只用唯讀工具查詢，需要對外寄出的通知先存成草稿，由同事確認後才送出。
An internal LINE assistant concept for Qinghe, a fictional wholesaler: staff ask about orders and stock in LINE, the AI agent looks things up with read-only tools, and anything sent to customers is saved as a draft until a colleague approves it.
此為概念設計，用於展示我們的設計與開發能力，並非真實客戶案例。
This is a concept design that demonstrates our design and development capabilities; it is not a real client project.
中小企業的業務與倉管每天都在問同樣的事：某位客戶的訂單出貨了沒、某個品項還有多少庫存。答案散在 ERP、試算表與群組對話裡，查一次就要切換好幾個系統；但直接讓 AI 連上公司系統，又擔心它改錯資料或把客戶資料送出去。
這個概念讓員工在熟悉的 LINE 裡提問，AI 代理依提問者的角色，只能使用管理者在後台開放的技能。查訂單、查庫存都是唯讀；草擬出貨通知可以，但寄出前一定要由主管核准。後台列出技能白名單與每個技能的審查和掃描結果、每個工具的權限，以及模型供應商與資料保留設定，每一步都留下執行紀錄。
實際開發時，我們會依需求評估採用 OpenClaw 這類開源代理框架或自建代理，以 MCP 伺服器把 ERP 與訂單系統包成權限明確的工具，並讓代理在隔離的容器中執行、只能連到白名單內的內部 API。個資欄位在送進模型前先遮罩，所有寫入動作都需真人核准並寫入稽核紀錄。畫面中的公司、人名與資料皆為虛構。
In smaller companies, sales and warehouse staff ask the same questions every day: has this customer’s order shipped, how much of this item is left? The answers sit across the ERP, spreadsheets and group chats, so one question means switching between several systems. Connecting an AI straight to company systems raises a different worry: that it changes the wrong record or sends customer data somewhere it should not go.
In this concept staff ask in LINE, which they already use, and the AI agent can only use the skills an administrator has enabled for the asker’s role. Looking up orders and stock is read-only; drafting a shipping notice is allowed, but a manager must approve it before it is sent. The admin console lists the skills allow-list with each skill’s review and scan result, the permission for every tool, and the model provider and data-retention settings, and every step is logged.
In a real build we would choose between an open-source agent framework such as OpenClaw and a custom agent, wrap the ERP and order system as clearly scoped tools behind MCP servers, and run the agent in an isolated container that can only reach allow-listed internal APIs. Personal data is masked before it reaches the model, and every write action needs human approval and lands in an audit log. All companies, names and data shown are fictional.
