不要让 Agent 一开始就拥有全部权限
企业 Agent 的第一原则是可控。它可以读取资料、生成建议、创建草稿，但不应该一开始就直接修改客户数据、下单、退款或发送外部邮件。高风险动作应当默认进入人工确认。
常见接入对象
- CRM：读取客户信息、沟通记录、销售阶段，生成跟进建议。
- 订单系统：读取订单状态、异常原因、物流进度，生成处理摘要。
- 文档系统：读取制度、产品说明、合同模板，回答内部问题。
- 客服系统：整理工单、识别重复问题、生成回复草稿。
工具调用设计
每个工具都应该有清晰边界，例如“搜索客户资料”“创建邮件草稿”“查询订单状态”。工具返回结构化数据，而不是把整个系统暴露给模型。
建议把 Agent 动作分成三档：只读、草稿、需确认执行。第一阶段优先做只读和草稿。
评估指标
- 任务完成率：Agent 是否能给出可用结果。
- 人工采纳率：人是否愿意使用它的建议。
- 节省时间：每个工单、客户或订单减少多少人工处理时间。
- 错误率：事实错误、格式错误和越权尝试出现频率。
Do not give the agent full permissions on day one
The first principle of enterprise agents is control. An agent can read data, generate suggestions and create drafts, but it should not immediately modify customer records, place orders, issue refunds or send external emails. High-risk actions should go through human approval.
Common systems to connect
- CRM: read customer data, conversation history and sales stage to draft follow-up suggestions.
- Order systems: read order status, exception reasons and logistics updates to create summaries.
- Document systems: read policies, product documents and contract templates for internal Q&A.
- Support systems: organize tickets, identify repeated issues and draft replies.
Tool-call design
Each tool should have a clear boundary, such as searching customer records, creating an email draft or checking order status. Tools should return structured data instead of exposing the whole system to the model.
Classify agent actions into read-only, draft and approval-required execution. Phase one should focus on read-only and draft workflows.
Evaluation metrics
- Task completion rate: whether the agent produces usable outputs.
- Human adoption rate: whether people accept the agent's suggestions.
- Time saved: how much manual work is reduced per ticket, customer or order.
- Error rate: factual errors, format errors and permission violations.
需要把 Agent 接入 CRM、订单或客服系统？Need to connect agents to CRM, order or support systems?
讨论接入方案Discuss integration