{
 "reviewed": "2026-07-15",
 "disclosure": "ALLinAI 編輯決策輔助資料，不是通用基準、法定門檻、產品排名或認證。請在測試前依特定部署情境、既有流程與錯誤後果設定門檻；不是每個指標都適合每個系統。",
 "systems": {
  "generation": "文字生成",
  "rag": "RAG 知識庫",
  "extraction": "擷取／結構化",
  "classification": "分類／預測",
  "agent": "AI Agent"
 },
 "dimensions": {
  "quality": "任務品質",
  "grounding": "資料與依據",
  "safety": "安全與權限",
  "operations": "營運與可靠性",
  "cost": "成本與效率",
  "human": "人工與使用影響"
 },
 "sources": {
  "nist_ai_rmf_measure": "https://airc.nist.gov/airmf-resources/airmf/5-sec-core/",
  "nist_ai_rmf_genai_profile": "https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf",
  "nist_tevv": "https://www.nist.gov/ai-test-evaluation-validation-and-verification-tevv",
  "nist_ai_800_2_draft": "https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.800-2.ipd.pdf",
  "owasp_genai_red_teaming": "https://genai.owasp.org/resource/genai-red-teaming-guide/",
  "owasp_excessive_agency": "https://genai.owasp.org/llmrisk/llm062025-excessive-agency/"
 },
 "metrics": [
  {
   "slug": "task-success-rate",
   "dimension": "quality",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "任務成功率",
   "question": "系統是否完成使用者原本要完成的工作？",
   "method": "先寫出每個案例的成功條件，再以成功案例數 ÷ 有效測試案例數；重大失敗另列，不只看平均。",
   "use_when": "所有 AI 應用的主指標，尤其適合有明確工作成果的流程。",
   "trap": "把語句流暢或流程跑完誤當任務成功，忽略結果是否可用。",
   "evidence": "版本化測試集、逐題結果、失敗分類、複核者與執行版本。",
   "related": "/knowledge/enterprise-ai/enterprise-ai-evaluation-plan-guide/"
  },
  {
   "slug": "critical-error-rate",
   "dimension": "quality",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "重大錯誤率",
   "question": "會造成權益、金錢、法遵、安全或不可逆影響的錯誤有多少？",
   "method": "測試前定義重大錯誤類型；重大錯誤案例數 ÷ 有效案例數，並逐案記錄嚴重度與控制是否攔截。",
   "use_when": "任何高影響或對外正式用途。",
   "trap": "用很高的整體正確率掩蓋少量但不可接受的嚴重失敗。",
   "evidence": "錯誤分類表、案例輸入輸出、影響評估、攔截與修正紀錄。",
   "related": "/knowledge/enterprise-ai/ai-acceptance-threshold-guide/"
  },
  {
   "slug": "field-accuracy",
   "dimension": "quality",
   "systems": [
    "extraction",
    "classification"
   ],
   "name": "欄位／標籤正確率",
   "question": "被擷取的欄位或預測標籤是否符合人工確認的標準答案？",
   "method": "依欄位與類別分開計算正確率；必要時同時看 precision、recall、F1 與混淆矩陣。",
   "use_when": "發票、表單、合約擷取，或案件分類與路由。",
   "trap": "只報整體平均，讓少數類別、空值與關鍵欄位的錯誤被隱藏。",
   "evidence": "標註規則、標準答案、分群結果、混淆矩陣與爭議標註處理。",
   "related": "/knowledge/enterprise-ai/llm-evaluation-metrics-guide/"
  },
  {
   "slug": "response-completeness",
   "dimension": "quality",
   "systems": [
    "generation",
    "rag"
   ],
   "name": "回答完整度",
   "question": "回答是否涵蓋任務要求與標準答案中的關鍵資訊？",
   "method": "把必要要點拆成評分規則，由人工或經校準的評測器逐點判斷；同時保留遺漏項目。",
   "use_when": "摘要、研究、客服與需要多項要件的回答。",
   "trap": "用篇幅當完整度；更長的回答也可能漏掉關鍵限制。",
   "evidence": "必要要點 rubric、逐題判斷、遺漏類型與抽樣人工覆核。",
   "related": "/knowledge/enterprise-ai/rag-evaluation-metrics-guide/"
  },
  {
   "slug": "instruction-following",
   "dimension": "quality",
   "systems": [
    "generation",
    "rag",
    "agent"
   ],
   "name": "指令遵循率",
   "question": "輸出是否遵守格式、語氣、範圍、禁止事項與必要步驟？",
   "method": "把指令拆成可判斷條件，逐條記錄通過、失敗或不適用；禁止事項應另設重大失敗。",
   "use_when": "格式固定、需守政策或會觸發後續流程的生成任務。",
   "trap": "只看 JSON 能不能解析，卻不驗證內容與限制是否正確。",
   "evidence": "系統提示版本、條件清單、逐條結果、解析錯誤與例外紀錄。",
   "related": "/knowledge/enterprise-ai/enterprise-ai-evaluation-plan-guide/"
  },
  {
   "slug": "consistency-rate",
   "dimension": "quality",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "重複測試一致性",
   "question": "相同或等價輸入重跑時，關鍵結論與動作是否穩定？",
   "method": "對代表性案例重跑多次，量測關鍵欄位、結論、工具選擇或評分落在容許範圍的比例。",
   "use_when": "模型具隨機性、流程很長，或使用者需要可預期結果時。",
   "trap": "把文字不同都算失敗，或反過來忽略結論與動作已改變。",
   "evidence": "執行參數、重跑次數、差異分類、模型與提示版本。",
   "related": "/knowledge/enterprise-ai/llm-evaluation-metrics-guide/"
  },
  {
   "slug": "retrieval-hit-rate",
   "dimension": "grounding",
   "systems": [
    "rag",
    "agent"
   ],
   "name": "檢索命中率",
   "question": "回答所需的正確來源是否出現在前 k 筆檢索結果？",
   "method": "先標註每題相關文件，再計算 Hit@k 或 Recall@k；依文件類型、權限與題型分群。",
   "use_when": "RAG 回答錯誤時，用來區分是沒找對資料還是生成沒用好資料。",
   "trap": "只測容易搜尋的題目，或在沒有標準相關文件時把主觀相似度當命中。",
   "evidence": "查詢、相關文件標註、前 k 筆結果、索引與切塊版本。",
   "related": "/knowledge/enterprise-ai/rag-evaluation-metrics-guide/"
  },
  {
   "slug": "retrieval-ranking-quality",
   "dimension": "grounding",
   "systems": [
    "rag",
    "agent"
   ],
   "name": "檢索排序品質",
   "question": "最有用的來源是否排在模型實際能使用的位置？",
   "method": "有分級相關標註時使用 MRR 或 NDCG；同時人工檢視前幾筆是否過時、重複或超出權限。",
   "use_when": "命中了正確文件但回答仍常被無關內容干擾時。",
   "trap": "直接比較不同工具顯示的分數，卻忽略 k、標註與計算定義不同。",
   "evidence": "分級相關標註、排名結果、參數、語料與指標定義。",
   "related": "/knowledge/enterprise-ai/rag-evaluation-metrics-guide/"
  },
  {
   "slug": "claim-groundedness",
   "dimension": "grounding",
   "systems": [
    "generation",
    "rag",
    "agent"
   ],
   "name": "主張依據忠實度",
   "question": "回答中的可查證主張是否能由提供來源直接支持？",
   "method": "把回答拆成可查證主張，逐項判斷支持、部分支持、衝突或無依據，再計算支持比例。",
   "use_when": "任何要求依指定文件回答、引用或提出建議的系統。",
   "trap": "把事實正確與依據忠實混為一談；即使剛好正確，也可能不受指定來源支持。",
   "evidence": "主張—來源對照、判斷規則、原文片段位置與複核紀錄。",
   "related": "/knowledge/enterprise-ai/rag-evaluation-metrics-guide/"
  },
  {
   "slug": "citation-correctness",
   "dimension": "grounding",
   "systems": [
    "rag",
    "agent"
   ],
   "name": "引用正確率",
   "question": "引用連結、文件、頁碼或段落是否真的支持前述內容？",
   "method": "抽取每個引用與相鄰主張，逐一檢查來源存在、可存取、版本正確且內容支持。",
   "use_when": "對外研究、客服、法規或內部政策查詢。",
   "trap": "只檢查有沒有引用圖示，沒有檢查引用內容。",
   "evidence": "引用清單、來源版本、支持判斷、失效連結與修正紀錄。",
   "related": "/knowledge/enterprise-ai/rag-evaluation-metrics-guide/"
  },
  {
   "slug": "freshness-coverage",
   "dimension": "grounding",
   "systems": [
    "rag",
    "agent",
    "extraction"
   ],
   "name": "資料新鮮度覆蓋率",
   "question": "需要最新資料的案例，是否使用仍在有效期或最新版的來源？",
   "method": "為受時效影響的來源設定 owner、有效日或更新週期，計算合格來源占比並測試舊版衝突案例。",
   "use_when": "價格、產品、政策、法規、庫存與流程文件等會變動的知識。",
   "trap": "只看索引更新時間，不看原始內容是否已過期。",
   "evidence": "來源 owner、版本、生效失效日、同步紀錄與過期告警。",
   "related": "/knowledge/ai-agent/enterprise-knowledge-base-document-preparation/"
  },
  {
   "slug": "no-answer-behavior",
   "dimension": "grounding",
   "systems": [
    "generation",
    "rag",
    "agent"
   ],
   "name": "無答案處理正確率",
   "question": "資料不足、衝突或超出範圍時，系統能否拒答並提出安全下一步？",
   "method": "建立確定無答案、資訊衝突與超出權限案例，判斷是否誠實表達限制、避免猜測並正確轉人工。",
   "use_when": "資料庫不完整或錯誤後果高的問答。",
   "trap": "把拒答越少當成越好，導致系統在沒有依據時硬答。",
   "evidence": "無答案案例集、拒答規則、錯誤回答、轉人工與使用者提示。",
   "related": "/knowledge/enterprise-ai/rag-evaluation-metrics-guide/"
  },
  {
   "slug": "prompt-injection-resistance",
   "dimension": "safety",
   "systems": [
    "rag",
    "agent",
    "generation"
   ],
   "name": "提示注入抵抗率",
   "question": "惡意輸入或外部內容能否改寫系統規則、洩漏資訊或誘發危險動作？",
   "method": "以直接、間接、多輪與混淆案例測試；分別記錄規則破壞、資料外洩與動作執行是否發生。",
   "use_when": "會讀取外部文件、網頁、郵件或工具輸出的系統。",
   "trap": "只測一句常見 jailbreak，然後宣稱系統安全。",
   "evidence": "攻擊案例版本、輸入輸出、攔截層、工具日誌、修復與重測結果。",
   "related": "/knowledge/enterprise-ai/ai-agent-evaluation-safety-guide/"
  },
  {
   "slug": "unauthorized-action-rate",
   "dimension": "safety",
   "systems": [
    "agent"
   ],
   "name": "未授權動作率",
   "question": "Agent 是否在沒有明確權限或批准時執行寫入、寄送、刪除、付款等動作？",
   "method": "建立允許、需批准、禁止三類案例；計算禁止或未批准動作被執行的比例，重大事件採零容忍門檻。",
   "use_when": "任何能呼叫外部工具或改變真實狀態的 Agent。",
   "trap": "只看 Agent 最後回覆，沒有查實際工具呼叫與外部狀態。",
   "evidence": "權限矩陣、逐次工具呼叫、批准人、外部變更、回復與事件紀錄。",
   "related": "/knowledge/enterprise-ai/ai-agent-evaluation-safety-guide/"
  },
  {
   "slug": "secret-pii-leakage-rate",
   "dimension": "safety",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "敏感資料洩漏率",
   "question": "系統是否在輸出、引用、日誌或工具參數中暴露不該看見的資料？",
   "method": "用假資料與跨角色案例測試輸入、輸出、日誌、快取與錯誤訊息；逐項判斷是否超出最小必要範圍。",
   "use_when": "處理個資、機密、憑證或跨部門知識的系統。",
   "trap": "只掃最終文字的格式，未檢查工具參數、追蹤與支援後台。",
   "evidence": "資料分類、角色權限、測試假資料、日誌檢查、事件與刪除證明。",
   "related": "/knowledge/enterprise-ai/ai-agent-evaluation-safety-guide/"
  },
  {
   "slug": "unsafe-content-rate",
   "dimension": "safety",
   "systems": [
    "generation",
    "rag",
    "agent"
   ],
   "name": "不安全內容率",
   "question": "在指定政策與受眾下，不允許的危害或違法內容是否被生成或放行？",
   "method": "依用途政策建立正常、邊界、規避與多輪案例；記錄攔截、誤擋、漏擋和升級處理。",
   "use_when": "公開服務、未成年或可被大規模濫用的生成系統。",
   "trap": "把供應商的通用安全分類直接當成公司的完整政策。",
   "evidence": "政策版本、測試案例、分類結果、人工複核與申訴修正紀錄。",
   "related": "/taiwan-ai-governance/"
  },
  {
   "slug": "permission-boundary-pass-rate",
   "dimension": "safety",
   "systems": [
    "rag",
    "agent"
   ],
   "name": "權限邊界通過率",
   "question": "不同角色只能檢索、看見或操作被授權的資源嗎？",
   "method": "建立使用者—資源—動作矩陣，測試允許與禁止組合；禁止組合不得因語意相似或跨輪對話繞過。",
   "use_when": "跨部門知識庫、多租戶與使用外部工具的 Agent。",
   "trap": "只測正常角色成功，沒有測禁止角色是否被拒絕。",
   "evidence": "權限矩陣、身分與角色、查詢結果、工具拒絕、變更與例外批准。",
   "related": "/knowledge/ai-agent/enterprise-knowledge-base-access-control-security/"
  },
  {
   "slug": "safe-stop-recovery",
   "dimension": "safety",
   "systems": [
    "agent",
    "rag",
    "generation"
   ],
   "name": "安全停止與回復成功率",
   "question": "超限、異常或人工停止時，系統能否停止且不留下不一致狀態？",
   "method": "在不同步驟注入逾時、拒絕、重複與人工中止；檢查停止時間、已執行動作、補償與資料一致性。",
   "use_when": "長流程、批次、Agent 與會修改外部資料的應用。",
   "trap": "有停止按鈕就算通過，卻沒驗證外部動作是否真的停止或回復。",
   "evidence": "故障注入、狀態轉移、外部紀錄、補償結果與演練報告。",
   "related": "/knowledge/enterprise-ai/ai-agent-evaluation-safety-guide/"
  },
  {
   "slug": "latency-percentiles",
   "dimension": "operations",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "回應延遲 P50／P95／P99",
   "question": "多數與尖峰使用者要等多久，是否符合工作流程時限？",
   "method": "在代表性負載下量測端到端與各步驟延遲，報告 P50、P95、P99，不只看平均。",
   "use_when": "互動服務、尖峰批次或有 SLA 的流程。",
   "trap": "只測開發環境單一請求，忽略併發、冷啟動與供應商限流。",
   "evidence": "負載模型、時間窗、分位數、步驟追蹤、模型版本與異常樣本。",
   "related": "/knowledge/enterprise-ai/ai-production-evaluation-monitoring-guide/"
  },
  {
   "slug": "availability-success-rate",
   "dimension": "operations",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "可用性與有效成功率",
   "question": "在需要的時間內，請求是否完成並產生可用結果？",
   "method": "分開記錄服務可達、技術成功與業務成功；依時間窗計算並標記事故與降級模式。",
   "use_when": "正式上線與依賴外部 API 的服務。",
   "trap": "HTTP 200 就算成功，忽略空白、截斷、錯誤格式或結果不可用。",
   "evidence": "SLO 定義、監控、錯誤分類、事故時間線與降級結果。",
   "related": "/knowledge/enterprise-ai/ai-production-evaluation-monitoring-guide/"
  },
  {
   "slug": "retry-duplicate-rate",
   "dimension": "operations",
   "systems": [
    "agent",
    "extraction",
    "classification"
   ],
   "name": "重試與重複副作用率",
   "question": "逾時或重試是否造成重複寄送、重複寫入或重複交易？",
   "method": "故意製造逾時與重送，計算重複副作用件數；檢查冪等鍵、狀態查詢與人工對帳。",
   "use_when": "任何會寫入外部系統、發訊息或觸發交易的流程。",
   "trap": "只測順利路徑，沒有測回應遺失但動作已完成的情境。",
   "evidence": "請求 ID、冪等鍵、工具與外部系統日誌、對帳及修復紀錄。",
   "related": "/knowledge/enterprise-ai/ai-production-evaluation-monitoring-guide/"
  },
  {
   "slug": "drift-alert-precision",
   "dimension": "operations",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "漂移告警有效性",
   "question": "輸入、資料或結果改變時，告警能否及時指出需要重測的真正變化？",
   "method": "定義分布、錯誤、來源與人工修正的基準；回放已知變化，追蹤告警命中、漏報與誤報。",
   "use_when": "資料、模型、提示、知識庫或使用方式會持續改變的系統。",
   "trap": "只監控 token 或流量，不連到品質與風險結果。",
   "evidence": "基準窗、變更事件、告警、調查、重測與門檻調整批准。",
   "related": "/knowledge/enterprise-ai/ai-production-evaluation-monitoring-guide/"
  },
  {
   "slug": "trace-completeness",
   "dimension": "operations",
   "systems": [
    "rag",
    "agent"
   ],
   "name": "追蹤紀錄完整率",
   "question": "能否從結果追溯到輸入、檢索、模型、提示、工具、批准與版本？",
   "method": "定義必要追蹤欄位，抽樣正式請求檢查可關聯與可讀性；敏感資料應遮罩而非完全失去證據。",
   "use_when": "需除錯、稽核、事件應變或多步 Agent。",
   "trap": "記很多日誌卻沒有共用請求 ID，或把機密原文全部落檔。",
   "evidence": "追蹤欄位規格、抽樣結果、保存權限、遮罩測試與查詢演練。",
   "related": "/knowledge/enterprise-ai/ai-production-evaluation-monitoring-guide/"
  },
  {
   "slug": "fallback-success-rate",
   "dimension": "operations",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "降級與替代流程成功率",
   "question": "模型或依賴失效時，能否安全轉人工、規則流程或替代供應商？",
   "method": "關閉關鍵依賴並演練降級；判斷使用者是否收到正確狀態、工作能否繼續且不重複執行。",
   "use_when": "關鍵流程、外部 API 依賴或不可長時間中斷的服務。",
   "trap": "只在文件寫有備援，從未在接近正式環境演練。",
   "evidence": "演練腳本、RTO/RPO 假設、切換結果、資料一致性與改善項目。",
   "related": "/knowledge/enterprise-ai/ai-production-evaluation-monitoring-guide/"
  },
  {
   "slug": "cost-per-qualified-outcome",
   "dimension": "cost",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "每個合格成果成本",
   "question": "產生一個通過品質與風險門檻的成果，完整成本是多少？",
   "method": "期間內模型、檢索、平台、人工複核與失敗重做成本 ÷ 合格成果數。",
   "use_when": "比較模型、流程、供應商與人工方案。",
   "trap": "只算 token 單價，忽略資料、人工、監控與失敗成果。",
   "evidence": "成本來源、時間窗、合格定義、成果數與分攤規則。",
   "related": "/knowledge/enterprise-ai/ai-project-cost-estimation-guide/"
  },
  {
   "slug": "human-review-minutes",
   "dimension": "cost",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "每件人工複核分鐘",
   "question": "人員要花多少時間確認、修正與完成一件 AI 成果？",
   "method": "從真實工作抽樣，分開量測查核、修改、退回與升級時間，報告中位數與高分位。",
   "use_when": "聲稱節省工時、有人在迴路或需要專業簽核的流程。",
   "trap": "把人工審查視為零成本，或用理想示範取代真實人員。",
   "evidence": "任務計時、角色、修正類型、等待時間與工作量分布。",
   "related": "/knowledge/enterprise-ai/ai-acceptance-threshold-guide/"
  },
  {
   "slug": "token-and-call-efficiency",
   "dimension": "cost",
   "systems": [
    "generation",
    "rag",
    "agent"
   ],
   "name": "Token 與呼叫效率",
   "question": "完成合格任務需要多少輸入輸出 token、模型與工具呼叫？",
   "method": "依合格成果彙總 token、重試、檢索與工具呼叫；比較相同品質門檻下的版本。",
   "use_when": "提示、上下文或 Agent 步驟快速膨脹時。",
   "trap": "追求 token 最少卻讓品質下降、重試增加或人工負荷上升。",
   "evidence": "逐步用量、價格版本、重試、品質結果與成本歸屬。",
   "related": "/tools/ai-project-budget-estimator/"
  },
  {
   "slug": "cache-reuse-rate",
   "dimension": "cost",
   "systems": [
    "generation",
    "rag",
    "agent"
   ],
   "name": "安全快取再利用率",
   "question": "重複工作能否在權限與新鮮度允許時安全重用結果？",
   "method": "計算符合權限、版本與有效期的快取命中；另測跨角色污染、過期與撤銷後失效。",
   "use_when": "高重複查詢、昂貴檢索或大量共同提示。",
   "trap": "只追求命中率，導致舊資料或他人結果被錯誤重用。",
   "evidence": "快取鍵、角色範圍、有效期、失效事件、命中與錯誤重用紀錄。",
   "related": "/knowledge/enterprise-ai/rag-cost-estimation-guide/"
  },
  {
   "slug": "waste-rate",
   "dimension": "cost",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "失敗與未使用成果浪費率",
   "question": "花費中有多少落在失敗、被丟棄或沒有業務用途的結果？",
   "method": "無效請求、失敗重跑、未採用成果與閒置容量成本 ÷ AI 總成本，按原因分類。",
   "use_when": "用量成長但價值沒有同步成長時。",
   "trap": "把所有呼叫都當採用，或只靠供應商帳單無法對應工作成果。",
   "evidence": "請求—成果關聯、採用狀態、失敗原因、帳務標籤與改善紀錄。",
   "related": "/knowledge/enterprise-ai/ai-production-evaluation-monitoring-guide/"
  },
  {
   "slug": "budget-variance",
   "dimension": "cost",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "預算偏差與成本上限事件",
   "question": "實際成本是否在批准情境與用量上限內？",
   "method": "比較週期實際與基準預算，拆解價格、流量、上下文、重試與人工差異；超限事件另記。",
   "use_when": "PoC 轉正式、跨部門擴張或變動計價服務。",
   "trap": "只看月底總額，無法在異常重試或濫用發生時及早停止。",
   "evidence": "預算版本、單價與匯率假設、用量告警、超限批准與根因。",
   "related": "/tools/ai-project-budget-estimator/"
  },
  {
   "slug": "correction-rate",
   "dimension": "human",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "人工修正率",
   "question": "AI 成果有多少需要人員改動才能正式使用？",
   "method": "定義輕微、重大、退回與重做，計算各類比例；同時看修改時間與錯誤種類。",
   "use_when": "草稿、摘要、擷取與分類等由人員接手完成的工作。",
   "trap": "只看有沒有按接受，未量測使用者其實花多久重寫。",
   "evidence": "原始輸出、最終版本、差異分類、修改者與耗時。",
   "related": "/knowledge/enterprise-ai/llm-evaluation-metrics-guide/"
  },
  {
   "slug": "automation-bias-rate",
   "dimension": "human",
   "systems": [
    "generation",
    "rag",
    "classification",
    "agent"
   ],
   "name": "錯誤採納率",
   "question": "AI 明顯錯誤時，人員是否仍未查核就接受或執行？",
   "method": "在知情且合乎倫理的測試中置入可辨識錯誤，量測採納、查核、否決與升級；避免把人員當成責任轉嫁。",
   "use_when": "人在迴路被當作主要安全控制的高影響流程。",
   "trap": "只寫『須人工複核』，沒有驗證人員有時間、資訊、能力與否決權。",
   "evidence": "測試設計、告知與保護、採納行為、訪談、教育與介面改善。",
   "related": "/knowledge/enterprise-ai/ai-acceptance-threshold-guide/"
  },
  {
   "slug": "escalation-accuracy",
   "dimension": "human",
   "systems": [
    "generation",
    "rag",
    "classification",
    "agent"
   ],
   "name": "轉人工正確率",
   "question": "不確定、高風險或超出範圍的案件是否被正確送給適任人員？",
   "method": "建立應轉與不需轉案例，計算 precision、recall；另量測等待時間、資料是否完整與是否找對角色。",
   "use_when": "客服、審查、法規、醫療、財務與例外很多的流程。",
   "trap": "只降低轉人工率，讓困難案件被 AI 強行處理。",
   "evidence": "轉人工規則、案件、接手角色、等待時間、處理結果與漏轉事件。",
   "related": "/knowledge/enterprise-ai/enterprise-ai-evaluation-plan-guide/"
  },
  {
   "slug": "user-comprehension",
   "dimension": "human",
   "systems": [
    "generation",
    "rag",
    "classification",
    "agent"
   ],
   "name": "使用者理解度",
   "question": "使用者是否知道 AI 的角色、限制、來源與下一步責任？",
   "method": "以任務測試與短問卷確認使用者能說明何時要查核、如何轉人工、結果是否為正式決定。",
   "use_when": "對外服務、AI 建議或可能被誤認為專業結論的介面。",
   "trap": "只確認畫面上有警語，沒有測使用者是否看懂並採取正確行動。",
   "evidence": "研究腳本、受試者範圍、理解題、觀察、介面版本與改善。",
   "related": "/knowledge/enterprise-ai/ai-transparency-labeling-taiwan-guide/"
  },
  {
   "slug": "workload-change",
   "dimension": "human",
   "systems": [
    "generation",
    "rag",
    "extraction",
    "classification",
    "agent"
   ],
   "name": "工作負荷與品質變化",
   "question": "導入後是否真的減少總負荷，或只是把工作改成更難的查錯與例外處理？",
   "method": "比較導入前後任務時間、等待、返工、例外、認知負荷與工作品質；依角色分群。",
   "use_when": "任何以提升效率、減少人力或改善體驗為目標的專案。",
   "trap": "只算模型生成速度，忽略查核、改錯、申訴與技能流失。",
   "evidence": "基準期、角色別工時、返工、錯誤、訪談與改善追蹤。",
   "related": "/knowledge/enterprise-ai/ai-adoption-kpi-metrics/"
  },
  {
   "slug": "appeal-remedy-success",
   "dimension": "human",
   "systems": [
    "classification",
    "agent",
    "generation",
    "rag"
   ],
   "name": "申訴與救濟成功率",
   "question": "受影響者能否理解、質疑、改正錯誤並取得人工處理？",
   "method": "測試告知、申訴入口、案件路由、處理時限、結果更正與後續系統修正；重大案例逐案檢討。",
   "use_when": "結果會影響資格、價格、服務、權益或正式紀錄時。",
   "trap": "只有客服信箱就宣稱可申訴，未驗證案件能否真正改正結果。",
   "evidence": "告知版本、案件紀錄、時限、處理人、結果更正與根因改善。",
   "related": "/taiwan-ai-governance/"
  }
 ]
}