cs-analytics
用 CSAT、NPS、FCR 和工單文字探勘衡量並改善客服績效
中文翻譯
---
name: "cs-analytics"
description: "用 CSAT、NPS、CES、首次接觸解決率（First Contact Resolution）和客服工單文字探勘，衡量並最佳化客服績效。當使用者需要評估客服團隊表現、找出最主要的客訴原因、最佳化人力配置，或建立客服儀表板時使用。就算使用者只說「我們客服做得好嗎」、「客戶都在抱怨什麼」、「需要幾個客服」或「做一個客服 dashboard」也要用。"
metadata:
category: "WP-06 Agent通訊+客服"
tags: ["customer-service", "analytics", "nps", "csat"]
---
# 客服分析
## 框架
```
IRON LAW: Measure Satisfaction AND Efficiency — Never Just One
High CSAT with terrible resolution time = unsustainable (agents spend
too long per ticket). Fast resolution with low CSAT = cutting corners.
Both dimensions must be tracked and balanced.
```
### 關鍵指標
**滿意度指標**
| 指標 | 衡量什麼 | 如何蒐集 | 基準 |
|--------|-----------------|---------------|-----------|
| **CSAT** | 對特定互動的滿意度 | 互動後問卷（1-5 分） | > 4.0/5 |
| **NPS** | 推薦意願 | 「你有多可能推薦我們？」（0-10） | > 30 |
| **CES** | 解決問題所需的力氣 | 「解決這個問題有多容易？」（1-7） | > 5.0/7 |
**效率指標**
| 指標 | 公式 | 基準 |
|--------|---------|-----------|
| **首次接觸解決率（FCR）** | 首次接觸就解決的數量 / 總接觸數 | > 70% |
| **平均處理時間（AHT）** | 總處理時間 / 總接觸數 | 5-8 分鐘（依產業而異） |
| **平均回應時間** | 從工單建立到第一次回覆的時間 | < SLA 目標 |
| **積壓量（Backlog）** | 未結工單 / 每日處理量 | < 1 天 |
| **升級率** | 升級的工單 / 總工單 | < 20% |
| **重開率** | 重開的工單 / 已解決的工單 | < 5% |
**營運指標**
| 指標 | 公式 | 用途 |
|--------|---------|-----|
| **工單量** | 每日/每週/每月工單數 | 人力規劃 |
| **管道組成** | 各管道占比（email、聊天、電話、LINE） | 資源配置 |
| **尖峰時段** | 各小時的工單量 | 排班 |
| **類別分布** | 各問題類型占比 | 流程改善的優先順序 |
### 分析流程
**1. 主要來電原因分析**
- 把所有工單依原因分類（自動標籤或手動）
- 畫柏拉圖：前 5 大原因通常占 60-80% 的量
- 對每個主要原因：能不能自助解決？能不能自動化？能不能從源頭消除？
**2. 工單文字探勘**
- 從工單描述中擷取高頻關鍵字/片語
- 分群成主題（LDA、BERTopic，或簡單的 TF-IDF）
- 找出新興問題（最近幾週才冒出來的新主題）
- 對客戶訊息做情緒分析
**3. 人力最佳化**
```
Required Agents = Peak Hour Volume × AHT / (60 × Utilization Target)
Example: 50 tickets/hour × 8 min AHT / (60 × 0.75 utilization) = 8.9 → 9 agents
```
要再加上休息、會議和訓練的緩衝（約 15-20%）。
**4. 客服人員績效**
| 指標 | 比較 | 行動 |
|--------|---------|--------|
| 個人 CSAT vs 團隊平均 | 找出需要輔導的人 | 低於平均的人給予訓練 |
| 個人 AHT vs 團隊平均 | 找出效率落差 | 跟著高績效者學習 |
| 各人員的 FCR | 找出知識缺口 | 改善知識庫 |
### VOC（客戶之聲）追蹤
| 訊號 | 來源 | 頻率 |
|--------|--------|-----------|
| 新興客訴 | 工單文字探勘 | 每週 |
| 功能需求 | 標記過的工單 + 問卷 | 每月 |
| 流失訊號 | 「取消」意圖的工單、低 CSAT 模式 | 每週 |
| 讚美模式 | 高 CSAT + 正面評語 | 每月（與團隊分享） |
## 輸出格式
```markdown
# CS Analytics Report: {Period}
## Summary Dashboard
| Metric | Current | Prior | Target | Status |
|--------|---------|-------|--------|--------|
| CSAT | {X}/5 | {X}/5 | >4.0 | 🟢/🟡/🔴 |
| FCR | {%} | {%} | >70% | 🟢/🟡/🔴 |
| Avg Response Time | {hrs} | {hrs} | <{X}hrs | 🟢/🟡/🔴 |
| Ticket Volume | {N} | {N} | — | ↑/↓ |
## Top Contact Reasons (Pareto)
| # | Reason | Volume | % | Self-Servable? |
|---|--------|--------|---|---------------|
| 1 | {reason} | {N} | {%} | Y/N |
## Emerging Issues
{New topics detected in text mining this period}
## Staffing
- Current agents: {N}
- Required (based on volume): {N}
- Gap: {over/under-staffed by N}
## Recommendations
1. {highest-impact improvement}
```
## 常見陷阱
- **CSAT 的回應偏誤**：只有 10-20% 的客戶會回問卷，而且通常是非常滿意或非常不滿的人。沉默的大多數的體驗是未知的。要用行為資料（重複聯絡、流失）補充。
- **NPS 是策略層，CSAT 是戰術層**：NPS 衡量整體品牌忠誠度（長期），CSAT 衡量特定互動的品質（短期）。不要用 NPS 來評估個別客服人員。
- **最佳化 AHT 可能傷害品質**：要求降低 AHT 的壓力可能讓客服趕進度，反而拉低 FCR 和 CSAT。先最佳化 FCR，再看 AHT。
- **工單分類會漂移**：隨著產品演進，分類會過時。每季檢視並更新分類體系。
- **客服資料中相關不等於因果**：「用比較多範本的客服 CSAT 比較高」可能代表範本有幫助，也可能只是資深客服（剛好愛用範本）本來就比較強。
## 參考資料
- NPS 問卷設計，見 `references/nps-methodology.md`
- 客服工單的文字探勘，見 `references/ticket-text-mining.md`英文原文
---
name: "cs-analytics"
description: "Measure and optimize customer service performance using CSAT, NPS, CES, First Contact Resolution, and text mining on support tickets. Use this skill when the user needs to evaluate CS team performance, identify top complaint drivers, optimize staffing, or build CS dashboards — even if they say 'is our CS team doing well', 'what are customers complaining about', 'how many agents do we need', or 'build a CS dashboard'."
metadata:
category: "WP-06 Agent通訊+客服"
tags: ["customer-service", "analytics", "nps", "csat"]
---
# Customer Service Analytics
## Framework
```
IRON LAW: Measure Satisfaction AND Efficiency — Never Just One
High CSAT with terrible resolution time = unsustainable (agents spend
too long per ticket). Fast resolution with low CSAT = cutting corners.
Both dimensions must be tracked and balanced.
```
### Key Metrics
**Satisfaction Metrics**
| Metric | What It Measures | How to Collect | Benchmark |
|--------|-----------------|---------------|-----------|
| **CSAT** | Satisfaction with specific interaction | Post-interaction survey (1-5 scale) | > 4.0/5 |
| **NPS** | Likelihood to recommend | "How likely to recommend?" (0-10) | > 30 |
| **CES** | Effort required to resolve | "How easy was it to resolve?" (1-7) | > 5.0/7 |
**Efficiency Metrics**
| Metric | Formula | Benchmark |
|--------|---------|-----------|
| **First Contact Resolution (FCR)** | Resolved on first contact / Total contacts | > 70% |
| **Average Handle Time (AHT)** | Total handle time / Total contacts | 5-8 min (varies by industry) |
| **Average Response Time** | Time from ticket creation to first response | < SLA target |
| **Backlog** | Open tickets / Daily throughput | < 1 day |
| **Escalation Rate** | Escalated tickets / Total tickets | < 20% |
| **Reopen Rate** | Reopened tickets / Resolved tickets | < 5% |
**Operational Metrics**
| Metric | Formula | Use |
|--------|---------|-----|
| **Ticket Volume** | Tickets per day/week/month | Staffing planning |
| **Channel Mix** | % by channel (email, chat, phone, LINE) | Resource allocation |
| **Peak Hours** | Volume by hour-of-day | Shift scheduling |
| **Category Distribution** | % by issue type | Process improvement priority |
### Analysis Workflows
**1. Top Contact Reason Analysis**
- Categorize all tickets by reason (auto-tag or manual)
- Pareto chart: top 5 reasons usually account for 60-80% of volume
- For each top reason: can it be self-served? Automated? Eliminated at source?
**2. Text Mining on Tickets**
- Extract frequent keywords/phrases from ticket descriptions
- Cluster into topics (LDA, BERTopic, or simple TF-IDF)
- Identify emerging issues (new topics appearing in recent weeks)
- Sentiment analysis on customer messages
**3. Staffing Optimization**
```
Required Agents = Peak Hour Volume × AHT / (60 × Utilization Target)
Example: 50 tickets/hour × 8 min AHT / (60 × 0.75 utilization) = 8.9 → 9 agents
```
Add buffer for breaks, meetings, and training (~15-20%).
**4. Agent Performance**
| Metric | Compare | Action |
|--------|---------|--------|
| Individual CSAT vs team avg | Identify coaching needs | Training for below-average |
| Individual AHT vs team avg | Identify efficiency gaps | Shadow high-performers |
| FCR by agent | Identify knowledge gaps | Knowledge base improvements |
### VOC (Voice of Customer) Tracking
| Signal | Source | Frequency |
|--------|--------|-----------|
| Emerging complaints | Ticket text mining | Weekly |
| Feature requests | Tagged tickets + surveys | Monthly |
| Churn signals | "Cancel" intent tickets, low CSAT patterns | Weekly |
| Praise patterns | High CSAT + positive comments | Monthly (share with team) |
## Output Format
```markdown
# CS Analytics Report: {Period}
## Summary Dashboard
| Metric | Current | Prior | Target | Status |
|--------|---------|-------|--------|--------|
| CSAT | {X}/5 | {X}/5 | >4.0 | 🟢/🟡/🔴 |
| FCR | {%} | {%} | >70% | 🟢/🟡/🔴 |
| Avg Response Time | {hrs} | {hrs} | <{X}hrs | 🟢/🟡/🔴 |
| Ticket Volume | {N} | {N} | — | ↑/↓ |
## Top Contact Reasons (Pareto)
| # | Reason | Volume | % | Self-Servable? |
|---|--------|--------|---|---------------|
| 1 | {reason} | {N} | {%} | Y/N |
## Emerging Issues
{New topics detected in text mining this period}
## Staffing
- Current agents: {N}
- Required (based on volume): {N}
- Gap: {over/under-staffed by N}
## Recommendations
1. {highest-impact improvement}
```
## Gotchas
- **CSAT response bias**: Only 10-20% of customers respond to surveys, usually the very happy and very unhappy. The silent majority's experience is unknown. Supplement with behavioral data (repeat contact, churn).
- **NPS is strategic, CSAT is tactical**: NPS measures overall brand loyalty (long-term). CSAT measures specific interaction quality (short-term). Don't use NPS to evaluate individual agents.
- **AHT optimization can hurt quality**: Pressure to reduce AHT may cause agents to rush, reducing FCR and CSAT. Optimize FCR first, then look at AHT.
- **Ticket categorization drift**: Categories become outdated as products evolve. Review and update the category taxonomy quarterly.
- **Correlation ≠ causation in CS data**: "Agents who use more templates have higher CSAT" might mean templates help, OR that experienced agents (who happen to use templates) are just better.
## References
- For NPS survey design, see `references/nps-methodology.md`
- For text mining on support tickets, see `references/ticket-text-mining.md`