Support Agent Pattern(客服智能体模式)
知识库位置:
FDE Knowledge Base/12 Patterns/02-Support-Agent-Pattern.md
1. Problem Signature(问题签名)
Section titled “1. Problem Signature(问题签名)”Support(客服)团队需要不断:
- 阅读 Ticket(工单)
- 搜索 Knowledge(知识)
- 查询 CRM(Customer Relationship Management,客户关系管理)
- 查询 Billing(计费系统)
- 执行简单操作
- 写回复
导致:
- Handle Time(处理时长)高
- 多系统切换
- 新员工培训慢
- 重复任务多
2. Business Signals(业务信号)
Section titled “2. Business Signals(业务信号)”客户可能说:
“客服人太多。”
“重复问题很多。”
“客服要切很多系统。”
“我们希望 Agent 自动解决 Ticket。”3. Typical Users(典型用户)
Section titled “3. Typical Users(典型用户)”- Customer Support Agent(客服坐席)
- Support Manager(客服主管)
- Customer(客户)
4. Current Workflow(现状工作流)
Section titled “4. Current Workflow(现状工作流)”Ticket(工单)↓Read(阅读)↓Classify(分类)↓Search KB(搜知识库)↓Search CRM(搜客户系统)↓Search Billing(搜计费)↓Draft Reply(起草回复)↓Approve(审批)↓Send(发送)5. Pain Points(痛点)
Section titled “5. Pain Points(痛点)”- Search(搜索)
- Context switching(上下文切换)
- Repetitive questions(重复问题)
- Manual notes(手工记录)
- Long resolution(解决耗时长)
- Inconsistent answers(答案不一致)
6. Root Causes(根本原因)
Section titled “6. Root Causes(根本原因)”常见:
- Knowledge fragmented(知识碎片化)
- Customer context fragmented(客户上下文碎片化)
- Stable workflow not automated(稳定流程未自动化)
- Too many manual checks(过多人工核对)
7. Target Workflow(目标工作流)
Section titled “7. Target Workflow(目标工作流)”建议从 Agent Assist(坐席辅助)开始:
Ticket(工单)↓Intent Classification(意图分类)↓Context Retrieval(上下文检索)├─ RAG(检索增强生成)├─ CRM└─ Billing↓Suggested Resolution(建议解决方案)↓Policy Check(策略检查)↓Human Review(人工审核)↓Send(发送)低风险场景成熟后:
Low Risk(低风险)→ Auto Resolution(自动解决)
High Risk(高风险)→ Human(人工)8. Value Hypothesis(价值假设)
Section titled “8. Value Hypothesis(价值假设)”Lower Handle Time(更低处理时长)+Higher Resolution Rate(更高解决率)+Less Manual Search(更少人工搜索)+Faster Training(更快培训)9. When to Use(适用场景)
Section titled “9. When to Use(适用场景)”适合:
- Ticket volume(工单量)高
- 大量重复问题
- Knowledge(知识)较成熟
- CRM / API 可访问
- Support 流程明确
10. When NOT to Use(不适用场景)
Section titled “10. When NOT to Use(不适用场景)”如果:
- 每个 Case(案件)极度复杂
- 数据缺失严重
- 业务流程本身未定义
- 高风险操作占大多数
不应一开始追求 Full Autonomy(完全自主)。
11. Architecture(架构)
Section titled “11. Architecture(架构)”组合:
Workflow(工作流)+Enterprise RAG(企业级检索增强生成)+Tool Calling(工具调用)+Policy Layer(策略层)+Human-in-the-Loop(人在环)Agent(智能体)仅用于动态部分。
12. Required Data(所需数据)
Section titled “12. Required Data(所需数据)”- Ticket(工单)
- Customer(客户)
- Subscription(订阅)
- Product(产品)
- Knowledge(知识)
- Account state(账户状态)
13. Integrations(所需集成)
Section titled “13. Integrations(所需集成)”- CRM
- Support Platform(客服平台)
- Billing(计费)
- Knowledge Base(知识库)
- Identity(身份)
14. AI Role(AI 角色)
Section titled “14. AI Role(AI 角色)”- Intent classification(意图分类)
- Search(搜索)
- Summarization(摘要)
- Suggested response(建议回复)
- Tool selection where needed(按需选择工具)
15. Deterministic Logic(确定性逻辑)
Section titled “15. Deterministic Logic(确定性逻辑)”- Refund threshold(退款阈值)
- Permission(权限)
- Account eligibility(账户资格)
- Policy(策略)
- Approval routing(审批路由)
16. Human Role(人工角色)
Section titled “16. Human Role(人工角色)”负责:
- Review(复核)
- High-risk decisions(高风险决策)
- Complex escalation(复杂升级)
- Customer judgment(客户判断)
17. Security & Risk(安全与风险)
Section titled “17. Security & Risk(安全与风险)”重点:
- Customer PII(Personally Identifiable Information,个人身份信息)
- User-context authorization(用户上下文授权)
- Tool write operations(工具写操作)
- External messaging(对外消息)
- Audit(审计)
18. Engineering Considerations(工程考量)
Section titled “18. Engineering Considerations(工程考量)”- Tool design(工具设计)
- CRM API reliability(CRM 接口可靠性)
- RAG quality(RAG 质量)
- Latency(延迟)
- Agent evaluation(智能体评估)
- Workflow state(工作流状态)
19. Delivery Strategy(交付策略)
Section titled “19. Delivery Strategy(交付策略)”Historical Ticket Analysis(历史工单分析)↓Assist Mode(辅助模式)↓Pilot with Support Agents(客服坐席试点)↓Low-risk Automation(低风险自动化)↓Expand Automation Boundary(扩大自动化边界)20. Pilot Design(试点设计)
Section titled “20. Pilot Design(试点设计)”首期不要追求 Ticket 100% 自动解决。
建议测试:
- 1–3 Intent Types(意图类型)
- Read-only Tooling(只读工具)
- Suggested Replies(建议回复)
- Human confirmation(人工确认)
21. Metrics(指标)
Section titled “21. Metrics(指标)”- Average Handle Time(平均处理时长)
- First Contact Resolution(首次接触解决率)
- Suggested Reply Acceptance(建议回复接受率)
- Escalation Rate(转人工率)
- Customer Satisfaction(客户满意度)
- Automation Rate(自动化率)
22. Failure Modes(失败模式)
Section titled “22. Failure Modes(失败模式)”- Agent tool overlap(工具职责重叠)
- Full autonomy too early(过早完全自主)
- No policy layer(没有策略层)
- Customer context stale(客户上下文过期)
- Human review becomes rubber stamp(人工复核沦为橡皮图章)
23. Troubleshooting(问题排查)
Section titled “23. Troubleshooting(问题排查)”- Agent Chooses the Wrong Tool(Agent 选错工具)
- Agent Loops Repeatedly(Agent 反复循环)
- API Integration Is Unstable(API 集成不稳定)