# IL 101 — Industry AI Project Charter：填寫範例（SYN-CS-001）/ Filled Example (SYN-CS-001)

> **合成教學資料 / Synthetic teaching data**：以下範例使用示例組織「示範服務公司 A」與 SYN-CS-001
> 合成案例，不代表任何真實企業。
> / The example below uses the example company "Demo Service Company A" and the SYN-CS-001 synthetic case;
> it does not represent any real company.

版本 / Version: v1.0（2026-09-26）｜對應合成案例 / Linked synthetic case: `SYN-CS-001`

## 中文

| 欄位 | 內容 |
|---|---|
| Project title | 客服知識助理 PoV（示範服務公司 A） |
| Industry domain | 零售／客戶服務 |
| Executive sponsor | （合成角色）客服部門副總 |
| Business / process owner | （合成角色）客服營運經理 |
| Problem statement | 客服人員每次回覆政策類問題（退貨、運費、帳單等）都需要人工搜尋知識庫、核實文件版本，平均每件耗時 15-25 分鐘；部分知識文件已過期或彼此矛盾，導致回覆錯誤與客訴。 |
| Current workflow | 見 Module 2 的現況流程圖：客戶提問 → 排隊 → 客服搜尋知識庫 → 不確定時轉主管確認 → 草擬回覆 → 部分因引用錯誤被退回重寫 → 回覆並結案。 |
| Baseline | 依 `baseline_log.csv`（30 筆合成觀測樣本，2026-01 至 2026-03）：平均總處理時間約 21-30 分鐘／件，依情境類別而異；標記為「估計」，樣本數屬教學規模。 |
| Target outcome | 縮短一般政策問答的平均處理時間，並降低因過期／矛盾文件造成的錯誤回覆比例。 |
| Users | 第一線客服人員；間接使用者為提出問題的客戶。 |
| AI role | 檢索現行有效的知識文件、辨識過期／矛盾來源、草擬回覆、偵測需轉真人的情境。 |
| Human role | 上線前抽樣覆核回覆品質、定期審查限閱文件分類、對客訴升級規則的正式核准、所有實際回覆的最終責任。 |
| Data / context requirements | `knowledge_documents.csv`（22 份合成文件，含 status／role_scope 欄位）及其正文（`documents/`）。 |
| Systems / integration | 現況為獨立知識庫查詢，不整合即時訂單系統（列為 out of scope，見 Module 4）。 |
| Constraints | 不得將限閱文件（KD-022）內容外洩；個資請求須先核實身分（KD-014）。 |
| Risks | 檢索邏輯若優先選語意相似而非現行有效的文件，可能沿用過期規則回覆客戶（見 `failure_log.csv` FL-01、FL-02、FL-09）。 |
| Success metrics | 指標卡「知識問答案件平均處理時間（分鐘／件）」，見 Module 3 範例；資料來源 `baseline_log.csv`／`experiment_results.csv`，期間 2026 Q1，估計值，負責人：專案技術負責人。 |
| Pilot scope | 僅涵蓋 `normal` 與 `missing_context` 兩類情境；`restricted_access` 與 `no_answer` 類別維持全人工，直到累積更多證據。 |
| Estimated effort | 依合成假設，開發與評估約需 2-3 週引導課程時間（IL 201），不含正式企業導入。 |
| Go / No-Go criteria | 依 `decision_examples/proceed.md` 的條件：`conflicting_sources`／`outdated_source` 類別平均品質分數需達 2.0 以上（0-3 尺度），且高風險類別錯誤率為 0。 |

## English

| Field | Content |
|---|---|
| Project title | Customer-Support Knowledge Assistant PoV (Demo Service Company A) |
| Industry domain | Retail / customer service |
| Executive sponsor | (synthetic role) VP of Customer Service |
| Business / process owner | (synthetic role) Customer Service Operations Manager |
| Problem statement | Agents manually search the knowledge base and verify document versions for every policy question (returns, shipping, billing, etc.), averaging 15-25 minutes per case; some documents are expired or contradict each other, causing wrong answers and complaints. |
| Current workflow | See Module 2’s process map: customer asks → queue → agent searches the knowledge base → escalates to a supervisor when unsure → drafts a reply → some are sent back to redo due to a wrong citation → reply sent and case closed. |
| Baseline | Per `baseline_log.csv` (30 synthetic observed samples, 2026-01 to 2026-03): average total handling time is roughly 21-30 minutes/case depending on category; labeled "estimate", teaching-material sample size. |
| Target outcome | Reduce average handling time for general policy Q&A, and reduce the rate of wrong replies caused by expired or contradicting documents. |
| Users | Front-line support agents; indirect users are the customers asking questions. |
| AI role | Retrieve the currently valid document, recognize expired/contradicting sources, draft a reply, and detect scenarios that must hand off to a human. |
| Human role | Sample-review reply quality before launch, periodically review the restricted-document classification, formally approve the escalation rules, and hold final responsibility for every reply. |
| Data / context requirements | `knowledge_documents.csv` (22 synthetic documents with status/role_scope fields) and their bodies (`documents/`). |
| Systems / integration | Currently a standalone knowledge-base lookup; does not integrate with the live order system (listed as out of scope, see Module 4). |
| Constraints | Restricted document content (KD-022) must never be disclosed; a data request must verify identity first (KD-014). |
| Risks | If retrieval ranking favors the semantically closest document over the currently valid one, it may answer with an outdated rule (see `failure_log.csv` FL-01, FL-02, FL-09). |
| Success metrics | The metric card "average handling time for a knowledge Q&A case (minutes/case)" from the Module 3 example; source `baseline_log.csv`/`experiment_results.csv`, period 2026 Q1, an estimate, owner: the project’s technical lead. |
| Pilot scope | Covers only the `normal` and `missing_context` categories; `restricted_access` and `no_answer` stay fully manual until more evidence accrues. |
| Estimated effort | Per the synthetic assumptions, development and evaluation take roughly 2-3 weeks of facilitated course time (IL 201), excluding formal enterprise adoption. |
| Go / No-Go criteria | Per `decision_examples/proceed.md`'s conditions: `conflicting_sources`/`outdated_source` categories must reach an average quality score of 2.0+ (0-3 scale), with a zero error rate in the high-risk categories. |