{ "schemaVersion": 1, "version": "2026-09-15", "id": "https://hk.onyxdevslab.com/data/recruiting-ai-agent-human-control.json", "name": "Recruiting AI agent human-control checklist", "datePublished": "2026-09-15", "dateModified": "2026-09-15", "evidenceClass": "Provider-authored implementation checklist", "publisher": { "name": "Onyx Devs Lab", "legalName": "ONYX DEVS LAB LIMITED", "url": "https://hk.onyxdevslab.com/", "leiCode": "254900Z30CLK7HKE9H46" }, "landingPages": { "en": "https://hk.onyxdevslab.com/en/guides/recruiting-ai-agent-human-control/", "zhHant": "https://hk.onyxdevslab.com/zh-hk/guides/recruiting-ai-agent-human-control/", "zhHans": "https://hk.onyxdevslab.com/zh-cn/guides/recruiting-ai-agent-human-control/" }, "languages": { "en": { "name": "How should a recruiting AI agent work with human recruiters?", "summary": "Use the agent to maintain context and prepare controlled actions, not to erase recruiter judgement. Separate facts from inference, define authority per candidate and channel, verify every send, preserve privacy and fairness review, and make human takeover immediate.", "decisionSignals": [ { "name": "Context", "meaning": "Complete conversation and current candidate state" }, { "name": "Send control", "meaning": "Draft, approve, execute, verify" }, { "name": "Human ownership", "meaning": "Recruiter policy, exceptions, and takeover" } ], "sections": [ { "heading": "Build an evidence-linked candidate context", "introduction": "Fast generation is unsafe when the system is working from incomplete conversation history or stale assumptions.", "checks": [ { "name": "Conversation completeness", "requirement": "Synchronise visible history, channel, participants, timestamps, attachments, recruiter notes, prior commitments, and delivery state before proposing a reply." }, { "name": "Facts versus inference", "requirement": "Link factual profile fields to messages or authorised records; label inferred interest, availability, seniority, compensation, or fit with confidence and missing information." }, { "name": "Privacy and fairness", "requirement": "Limit purpose, collection, access, retention, export, and sensitive attributes; test whether ranking, outreach, or recommendation creates unjustified differential treatment." } ] }, { "heading": "Control every outward action", "introduction": "A generated message is not sent until identity, authority, state, and completion have been verified.", "checks": [ { "name": "Authority modes", "requirement": "Define manual, copilot, and limited autopilot per candidate, recruiter, channel, message type, working hours, frequency, and escalation condition." }, { "name": "Safe sending", "requirement": "Recheck current conversation, validate recipient and content, serialise sends, use idempotency, respect platform limits, and prevent duplicates or stale replies." }, { "name": "Verification and takeover", "requirement": "Read back the sent record, retain request and result, surface uncertainty and failures, pause automation on new context, and allow the recruiter to take control instantly." } ] } ], "procedure": [ { "position": 1, "name": "Map the recruiting policy", "instruction": "Define permitted data, sourcing, screening, outreach, scheduling, rejection, sensitive decisions, retention, and accountable recruiters." }, { "position": 2, "name": "Establish context quality", "instruction": "Measure conversation completeness, profile freshness, fact support, missing fields, and synchronisation failures before testing responses." }, { "position": 3, "name": "Pilot in copilot mode", "instruction": "Require approval, record edits, verify sends, study exceptions and candidate experience, and compare against the existing workflow." }, { "position": 4, "name": "Automate only narrow actions", "instruction": "Promote low-risk repeated actions with limits, monitoring, revocation, fairness review, and immediate recruiter takeover." } ], "frequentlyAskedQuestions": [ { "question": "Should an agent automatically reject candidates?", "answer": "High-impact employment decisions require strong legal, fairness, evidence, and human-accountability controls. A safer default is decision support with qualified review." }, { "question": "What does send verification mean?", "answer": "After the action, the system rereads the platform record to confirm the intended recipient, content, timestamp, and delivery state rather than assuming a click succeeded." }, { "question": "Can inferred candidate attributes be stored as facts?", "answer": "No. Keep facts, inferences, confidence, source, expiry, and missing information separate, and restrict sensitive or unnecessary inference." } ] }, "zhHant": { "name": "招聘 AI Agent 應如何與招聘顧問協作？", "summary": "Agent 應維持上下文及準備受控動作，而不是抹去招聘顧問判斷。分開事實與推斷，按候選人及渠道定義權限，驗證每次發送，保留私隱與公平覆核，並讓人工可以立即接管。", "decisionSignals": [ { "name": "上下文", "meaning": "完整對話及最新候選人狀態" }, { "name": "發送控制", "meaning": "草擬、批准、執行、驗證" }, { "name": "人工負責", "meaning": "招聘政策、例外及接管" } ], "sections": [ { "heading": "建立連接證據的候選人上下文", "introduction": "若系統依賴不完整對話歷史或過期假設，快速生成也不安全。", "checks": [ { "name": "對話完整度", "requirement": "建議回覆前同步可見歷史、渠道、參與者、時間、附件、招聘筆記、既有承諾及傳送狀態。" }, { "name": "事實與推斷", "requirement": "把事實欄位連至訊息或授權記錄；候選人意向、可入職時間、資歷、薪酬或適配推斷需標記置信度及缺失信息。" }, { "name": "私隱及公平", "requirement": "限制目的、收集、存取、保留、導出及敏感屬性；測試排序、外展或建議有否造成無理差別待遇。" } ] }, { "heading": "控制每項對外動作", "introduction": "生成訊息在身份、權限、狀態及完成均獲驗證前不等於已發送。", "checks": [ { "name": "權限模式", "requirement": "按候選人、招聘顧問、渠道、訊息類型、工作時間、頻率及升級條件定義人工、Copilot 及有限 Autopilot。" }, { "name": "安全發送", "requirement": "重新檢查最新對話、核實收件人及內容、串行發送、執行冪等、遵守平台限制，並避免重複或過期回覆。" }, { "name": "驗證及接管", "requirement": "回讀已發記錄，保留請求與結果，顯示不確定性及失敗，新上下文出現時暫停自動化，讓顧問即時接管。" } ] } ], "procedure": [ { "position": 1, "name": "梳理招聘政策", "instruction": "定義允許數據、尋訪、篩選、外展、安排、拒絕、敏感決定、保留及問責招聘顧問。" }, { "position": 2, "name": "建立上下文質量", "instruction": "測量對話完整、畫像新鮮度、事實支持、缺失欄位及同步失敗，再測試回覆。" }, { "position": 3, "name": "以 Copilot 模式試點", "instruction": "要求批准、記錄修改、驗證發送、研究例外與候選人體驗，並與現有流程比較。" }, { "position": 4, "name": "只自動化狹窄動作", "instruction": "只有低風險重複動作才逐步提升權限，並配置限額、監測、撤銷、公平覆核及即時人工接管。" } ], "frequentlyAskedQuestions": [ { "question": "Agent 應否自動拒絕候選人？", "answer": "高影響僱傭決定需要嚴格法律、公平、證據及人工問責控制；較安全預設是由合資格人員覆核的決策支援。" }, { "question": "甚麼是發送驗證？", "answer": "動作後回讀平台記錄，確認預期收件人、內容、時間及傳送狀態，而不是假設點擊已成功。" }, { "question": "可否把推斷候選人屬性存為事實？", "answer": "不可。應分開事實、推斷、置信度、來源、到期及缺失信息，並限制敏感或不必要推斷。" } ] }, "zhHans": { "name": "招聘 AI Agent 应该怎样与招聘顾问协作？", "summary": "Agent 应该维护上下文和准备受控动作，而不是抹去招聘顾问判断。分开事实与推断，按照候选人和渠道定义权限，验证每次发送，保留隐私与公平复核，并让人工可以立即接管。", "decisionSignals": [ { "name": "上下文", "meaning": "完整对话和最新候选人状态" }, { "name": "发送控制", "meaning": "草拟、批准、执行、验证" }, { "name": "人工负责", "meaning": "招聘政策、异常和接管" } ], "sections": [ { "heading": "建立连接证据的候选人上下文", "introduction": "如果系统依赖不完整对话历史或过期假设，快速生成也不安全。", "checks": [ { "name": "对话完整度", "requirement": "建议回复前同步可见历史、渠道、参与者、时间、附件、招聘笔记、已有承诺和发送状态。" }, { "name": "事实与推断", "requirement": "把事实字段连接到消息或授权记录；候选人意向、可入职时间、资历、薪酬或适配推断需要标记置信度和缺失信息。" }, { "name": "隐私与公平", "requirement": "限制目的、收集、访问、保留、导出和敏感属性；测试排序、外联或建议是否造成没有依据的差别待遇。" } ] }, { "heading": "控制每项对外动作", "introduction": "生成消息在身份、权限、状态和完成都经过验证前不等于已经发送。", "checks": [ { "name": "权限模式", "requirement": "按照候选人、招聘顾问、渠道、消息类型、工作时间、频率和升级条件定义人工、Copilot 和有限 Autopilot。" }, { "name": "安全发送", "requirement": "重新检查最新对话、核实收件人和内容、串行发送、执行幂等、遵守平台限制，并避免重复或过期回复。" }, { "name": "验证与接管", "requirement": "回读已发记录，保留请求和结果，显示不确定性和失败，新上下文出现时暂停自动化，让顾问即时接管。" } ] } ], "procedure": [ { "position": 1, "name": "梳理招聘政策", "instruction": "定义允许数据、寻访、筛选、外联、安排、拒绝、敏感决策、保留和问责招聘顾问。" }, { "position": 2, "name": "建立上下文质量", "instruction": "测量对话完整、画像新鲜度、事实支持、缺失字段和同步失败，再测试回复。" }, { "position": 3, "name": "以 Copilot 模式试点", "instruction": "要求批准、记录修改、验证发送、研究异常与候选人体验，并和现有流程比较。" }, { "position": 4, "name": "只自动化狭窄动作", "instruction": "只有低风险重复动作才逐步提升权限，并配置限额、监控、撤销、公平复核和即时人工接管。" } ], "frequentlyAskedQuestions": [ { "question": "Agent 应该自动拒绝候选人吗？", "answer": "高影响雇佣决策需要严格法律、公平、证据和人工问责控制；更安全的默认方式是由合资格人员复核的决策支持。" }, { "question": "什么是发送验证？", "answer": "动作后回读平台记录，确认预期收件人、内容、时间和发送状态，而不是假设点击已经成功。" }, { "question": "可以把推断候选人属性存为事实吗？", "answer": "不可以。应该分开事实、推断、置信度、来源、到期和缺失信息，并限制敏感或不必要的推断。" } ] } }, "limitations": [ "This asset is provider-authored and is not independent validation or a claim of client results.", "It is an implementation aid, not legal, privacy, security, financial, procurement, or regulatory advice.", "Thresholds and controls must be adapted to the organisation, workflow, jurisdiction, and risk owners." ] }