返回 Papers
AI 扩展计划 / Playbooks

AI Knowledge Work Redesign / Role-Task Architecture Playbook

以下来源作为本文的设计锚点。本文把它们转成知识工作、角色任务架构、产品控制、流程治理和运行指标语言。访问日期按 2026-06-29 记录。

722AI_KNOWLEDGE_WORK_REDESIGN_ROLE_TASK_ARCHITECTURE_PLAYBOOK.md

AI Knowledge Work Redesign / Role-Task Architecture Playbook

定位: 面向 AI BA / AI PM / Product Architect / Enterprise Architect / Operations Transformation Lead 的知识工作重设计手册。

核心问题: AI 时代的金融零售知识工作, 不能只靠给员工发一个 Copilot。真正要设计的是 role、job、task、decision、artifact、handoff、control、metric 之间的新架构。

适用范围: AML / KYC / credit underwriting / collections / complaints / customer service / branch operations / payment disputes / retail banking operations / insurance and wealth servicing / retail merchandising and supply chain operations。

读者假设: 读者已经具备 CBAP 级业务分析能力。本文不讲基础组织变革、普通培训沟通或通用流程优化, 重点放在 AI-era knowledge work redesign、角色任务边界、责任设计、产品和架构控制、金融零售落地证据。

重要说明: 本文是学习、作品集和内部方案训练材料, 不构成法律意见、合规结论、审计意见、模型验证报告或监管解释。正式项目必须由 Legal、Compliance、Risk、Model Risk、Privacy、Security、Data Owner、Business Owner、Operations Owner、HR、Works Council / Employee Relations 和管理层结合机构类型、司法辖区、员工政策、客户影响和内部控制框架确认。


1. Source Anchors

以下来源作为本文的设计锚点。本文把它们转成知识工作、角色任务架构、产品控制、流程治理和运行指标语言。访问日期按 2026-06-29 记录。

AnchorOfficial source本文使用方式
WEF Jobs of Tomorrow: Large Language Models and Jobshttps://www.weforum.org/publications/jobs-of-tomorrow-large-language-models-and-jobs/用 job exposure / task exposure 视角提醒: AI 冲击的是任务组合, 不是简单替换职位名称。角色重设计必须先做 task decomposition。
Microsoft Work Trend Indexhttps://www.microsoft.com/en-us/worklab/work-trend-index/用 workplace research 视角提醒: AI adoption 的瓶颈不是工具可用性, 而是工作结构、管理节奏、员工负荷、协作方式和组织能否围绕 AI 重构工作。
NIST AI Risk Management Frameworkhttps://www.nist.gov/itl/ai-risk-management-framework用 Govern / Map / Measure / Manage 组织 role-task architecture 的风险识别、控制设计、指标验证、上线门禁和持续改进。
Microsoft Guidelines for Human-AI Interactionhttps://www.microsoft.com/en-us/research/publication/guidelines-for-human-ai-interaction/用 Human-AI Interaction 原则设计预期、证据、不确定性、纠错、反馈、控制、恢复和长期适应, 避免员工对 AI 过度信任或低效抵触。

Source-to-artifact mapping:

Source lens直接产出的 artifact高级表达
WEF task exposureRole-task inventory、AI exposure heatmap、skill shift map“我会先拆任务和判断点, 再决定岗位如何演进, 而不是讨论哪些职位会被 AI 取代。”
Work Trend IndexAdoption operating rhythm、manager dashboard、employee load metric、frontier workflow experiment“AI 转型的核心是重新组织工作, 不是把个人生产力工具分发给员工。”
NIST AI RMFUse case risk tier、control matrix、release gate、monitoring dashboard、evidence binder“我会把任务分配和责任设计纳入 AI 风险管理闭环。”
Microsoft HAI GuidelinesHuman-AI responsibility matrix、review UI requirement、feedback taxonomy、override workflow“我会设计人如何理解、质疑、纠正和覆盖 AI, 而不是只设计 AI 输出内容。”

2. One-Sentence Positioning

AI Knowledge Work Redesign =
用 task decomposition 拆开知识工作,
用 role-task architecture 重新分配人、AI、规则、系统和控制点的职责,
用 accountability、metrics、training、workflow 和 architecture controls
证明工作更快、更准、更可控、员工负荷更健康、客户和监管风险更低。

中文记忆:

AI 时代的工作重设计, 不是“员工 + Copilot”, 而是“任务、判断、证据、责任、交接、控制、指标”的重新架构。

在金融零售里, 这件事尤其重要:

场景不能只看效率的原因需要一起设计的边界
AML alert review错误关闭 alert 可能形成金融犯罪风险和监管暴露investigator 判断权、AI narrative 草稿、证据引用、supervisor review、SAR escalation
Credit underwritingAI 建议可能影响客户授信、拒绝理由、公平信贷和模型风险underwriter accountability、policy reason code、override、adverse action evidence
Customer serviceAI 回复可能成为客户理解中的正式承诺agent assist 边界、客户可见文案、投诉升级、知识库来源、质量抽检
Payment disputes自动化加速不等于争议处理正确scheme rules、case evidence、merchant response、temporary credit approval、reversal path
Collections话术、时机和客户脆弱性有合规和声誉风险contact strategy、hardship routing、human empathy、script control、call audit
Retail operationsAI 推荐补货或促销会影响库存、门店执行和利润buyer / planner 决策权、forecast confidence、exception reason、execution feedback

3. 为什么 AI 转型不是“给员工一个 Copilot”

给员工一个 Copilot 只能改变个人界面, 不能自动改变工作系统。金融零售的知识工作有强约束: 客户影响、监管证据、审批授权、系统记录、员工能力、队列管理、质量抽检、风险升级和审计复盘。AI 工具如果不进入这些工作边界, 只能制造局部效率, 甚至把错误规模化。

3.1 Copilot-only 反模式

反模式表面收益深层问题
Personal productivity only员工写得更快、查得更快没有改变流程瓶颈、审批队列、证据字段和下游系统动作
Shadow workflow员工把 AI 输出复制到正式系统正式系统看不到 AI 来源、版本、证据、人工修改和责任链
Rubber-stamp reviewAI 生成结论, 人点击确认管理指标逼迫员工默认采纳, review 变成行政动作
Task automation without decision designAI 自动分类、摘要、路由分类规则、误判成本、人工覆盖和客户影响没有显式设计
Skill erosion员工越来越依赖 AI 起草和判断新人无法形成业务判断, 资深员工变成低价值复核员
Metric illusionAHT、处理量、文档产出提升质量、风险、返工、客户伤害、员工负荷和长期学习没有被测量
Control theater文档写着“人负责最终决定”系统没有 evidence view、override、escalation、audit log 和 stop control

3.2 真正要重设计的不是工具, 而是工作架构

成熟设计要同时回答 12 个问题:

问题应落到的 artifact
哪个 role 对业务结果负责Role accountability map
哪个 job 被拆成哪些任务Job-task decomposition
哪些 task 适合 AI 辅助、建议、决策或自动化Task allocation matrix
哪些 decision 必须由人作出Decision accountability matrix
AI 生成的 artifact 是否可进入正式记录Artifact lifecycle map
人和 AI 之间如何 handoffHandoff contract
质量和风险控制点在哪里强制执行Control matrix
谁能 override AI 或人工决定Override authority table
证据由谁保存、保存什么、保存多久Evidence retention schema
员工需要什么新技能Skill shift map
管理层如何看采用和负荷Adoption and workload dashboard
架构如何防止越权、泄露、幻觉和不可复盘Architecture control plane

3.3 设计原则

原则含义金融零售落地
Task before title先拆任务, 再谈岗位变化AML analyst 不是被替换, 而是从资料整理转向风险判断、例外处理和监管叙事质量
Decision before automation先定义决策权和证据, 再定义自动化Credit assistant 可以建议 reason code, 但拒贷决定和解释责任仍需授权人员承担
Control before scale先证明控制有效, 再扩大用户Customer service AI 先通过知识来源、合规话术、升级路径和抽检指标, 再开放更多场景
Metrics before claims先建立 baseline, 再声称效率收益不只看处理时长, 还看 first pass quality、override quality、rework、complaint trend
Learning before deskilling设计技能升级路径, 防止判断能力退化新员工用 AI 解释案例, 但必须完成无 AI case drill 和校准评审

4. Role-Task Architecture

Role-task architecture 是把知识工作拆成可分配、可控制、可衡量、可训练、可审计的结构。它连接八个对象:

Role -> Job -> Task -> Decision -> Artifact -> Handoff -> Control -> Metric

4.1 八个核心对象

对象定义设计问题金融零售示例
Role业务责任和权限的承载者谁对结果负责, 谁有授权, 谁接受风险AML investigator、credit underwriter、contact center agent、branch manager
Job一个 role 在组织中的岗位组合这个岗位由哪些任务、目标和约束组成Credit underwriter job 包含资料审查、政策判断、额度建议、拒绝理由、case note
Task可被分配和评估的工作单元任务输入、输出、判断难度、风险和可自动化程度是什么摘要客户收入证明、识别缺失材料、草拟 underwriting memo
Decision对业务结果有影响的判断点决策依据、授权、解释、复核和申诉路径是什么是否升级 EDD、是否批准授信、是否发起 dispute temporary credit
Artifact工作产生或消费的证据对象输出能否进入正式记录, 版本和来源如何追踪AML case narrative、credit memo、customer reply、QA checklist
Handoff人、AI、系统、团队之间的交接交接时传递什么上下文, 谁接收, SLA 是什么Chatbot 转人工时传递身份验证状态、客户诉求、AI 回复、来源
Control防止错误、越权、泄露和不可复盘的机制控制点在流程中是否强制执行tool gateway、maker-checker、policy engine、override reason、audit log
Metric衡量价值、质量、风险、学习和负荷的指标指标是否会鼓励错误行为AHT + QA pass + escalation quality + employee cognitive load

4.2 Role、job、task 的区别

层级错误问法正确问法
Role“AI 会不会替代 AML analyst?”“AML analyst 的哪些任务会被增强, 哪些判断权保留, 哪些新责任出现?”
Job“这个岗位还能剩多少工作量?”“岗位目标、任务组合、绩效指标、技能要求和责任边界如何变化?”
Task“这个任务能不能自动化?”“这个任务的输入证据、输出形式、误判成本、可逆性和复核要求是什么?”

4.3 Task decomposition 层级

层级示例: AML alert reviewAI redesign 关注点
Value stream从 alert 触发到关闭、升级或 SAR 叙事控制金融犯罪风险, 降低低价值整理工作, 提高 narrative 质量
ProcessAlert intake、case enrichment、investigation、decision、supervisor review、filingAI 插入点、人工节点、队列、SLA、例外路径
Activity收集客户资料、分析交易模式、检查名单、写 case note哪些是 evidence gathering, 哪些是 judgment
Task汇总 90 天交易、识别 round-dollar pattern、草拟 narrativeAI extraction / summary / recommendation / drafting
Micro-task提取交易对手、按金额聚合、标记异常描述可自动化但必须记录来源和转换逻辑
Decisionclose、continue monitoring、escalate、SAR consideration人类责任、复核、原因码、证据包
Artifactrisk summary、case note、narrative、supervisor packet版本、引用、人工修改、审批记录

4.4 Work architecture 的四条流

Flow说明典型控制
Work flow工作如何从入口到结果流动BPMN、queue management、SLA、exception routing
Decision flow判断如何被作出、复核和覆盖DMN、policy engine、authority matrix、reason code
Evidence flow事实、来源、模型输出和人工修改如何被保存evidence store、trace id、source metadata、audit trail
Accountability flow谁负责结果、风险接受、复核和持续改进RACI、maker-checker、business owner sign-off、control owner

4.5 Product and architecture controls

Role-task architecture 必须落到产品和架构控制, 否则只是流程图。

User task UI
-> role and entitlement check
-> task context package
-> policy / DMN decision service
-> AI service with retrieval and tool boundaries
-> evidence and uncertainty display
-> human action controls
-> workflow state transition
-> audit event
-> quality, risk and workload monitoring
Control component责任典型 evidence
Identity and entitlement确认员工角色、权限、客户关系、数据用途actor id、role、resource、purpose、access decision
Task context package把 case facts、policy、history、constraints 打包给 AI 和人context version、source list、data freshness
Policy / DMN service把可解释规则和 reason code 从 prompt 中分离decision id、rule id、version、input facts hash
AI orchestration layer限定模型、检索、工具、输出格式和失败路径model id、prompt version、tool call log、guardrail result
Evidence view让人看到 AI 输出背后的来源、缺口和冲突citation、source owner、effective date、missing evidence flag
Human control gateway管理 accept、edit、reject、override、escalate、rollbackaction log、diff、reason code、reviewer id
Workflow engine强制状态转换、审批、SLA 和交接case state、transition rule、queue event
Observability and audit监控质量、风险、采用、成本和员工负荷dashboard、sample review、incident record、audit replay

5. Task Allocation

任务分配不是“AI 能做就让 AI 做”。要按风险、可逆性、数据质量、政策清晰度、时间压力、客户影响、员工能力和控制成本来决定。

5.1 六类 task allocation

Allocation pattern定义适合任务不适合任务控制要求
Human-only人必须完成, AI 不能生成实质性判断或行动高同理心沟通、重大例外、风险接受、监管签字、纪律处分、复杂伦理判断高频低风险资料整理强授权、记录理由、必要时双人复核
AI-assistedAI 帮人检索、摘要、提取、改写, 人作判断case summary、政策查找、证据抽取、文案润色人没有时间或能力检查的高风险结论来源展示、编辑差异、人工确认
AI-recommendedAI 给出建议、排序、下一步, 人选择是否采纳alert prioritization、missing document suggestion、customer intent routing直接影响客户权益且缺少解释的决策解释、替代方案、override reason、采纳率监控
AI-decided with human reviewAI 生成初步决定, 人必须复核后生效低到中风险分类、规则清晰的初审、批量 QA 初筛高影响不可逆动作、歧视敏感决策、正式拒绝原因pre-effect review、sample calibration、false accept / false reject 监控
AI-automatedAI 在授权范围内自动完成任务或系统动作低风险、可逆、规则清晰、高频的状态更新、任务创建、资料分类冻结账户、拒贷、关闭投诉、发正式监管材料tool allowlist、limit、rollback、exception queue、audit trace
AI-monitoredAI 不直接处理任务, 只监控异常、趋势、负荷和控制偏离SLA breach prediction、quality drift、queue anomaly、employee overload signal需要正式个案判断的结论alert owner、threshold governance、no punitive use without review

5.2 Allocation decision criteria

CriterionHuman-only 倾向AI-assisted 倾向AI-automated 倾向
Customer impact高, 影响授信、账户、资金、投诉、法律权利中, 帮助员工解释或准备材料低, 不直接改变客户权益
Reversibility不可逆或补救成本高可人工修正可自动回滚或影响很小
Evidence quality证据冲突、缺失、上下文强有权威来源但需要判断来源结构化、稳定、规则明确
Policy clarity需要专家解释或例外处理规则清楚但有少量灰区规则完全可编码、版本稳定
Bias / fairness sensitivity
Time criticality可等专家判断需要加速但可排队必须实时且低风险
Human skill value需要专业判断、同理心、谈判、风险接受人负责最终判断, AI 减少整理负担人的参与主要增加延迟, 不增加判断质量
Audit expectation需要完整审批和解释需要来源、修改和确认记录需要自动日志和抽样复核

5.3 金融零售任务分配示例

Use caseTaskAllocation原因控制
AML汇总客户 90 天交易和对手方AI-assisted资料多、规则明确, 但可能遗漏异常source-linked summary、investigator confirmation
AML判断是否提交 SAR considerationHuman-only高监管影响, 需要专业判断和责任承担senior review、decision memo、evidence binder
Credit提取收入证明字段AI-assisted文档处理适合 AI, 但字段需验证OCR confidence、source highlight、manual correction
Credit额度审批和拒绝理由确认Human-only 或 AI-recommended影响客户权益和公平信贷reason code validation、underwriter accountability
Customer service根据政策草拟费用解释AI-assisted员工可编辑, 来源可展示approved knowledge source、agent edit log
Customer service自动回答普通营业时间、状态查询AI-automated低风险、可逆、来源稳定intent confidence、fallback、interaction log
Payment disputes路由到 correct dispute typeAI-decided with human review分类可加速, 误分有成本review queue、scheme rule source、override metric
Retail planning异常缺货预警AI-monitoredAI 发现模式, 采购经理决定行动threshold owner、exception annotation、buyer decision log

5.4 Allocation anti-patterns

Anti-pattern问题修正方式
Automate what is broken把混乱流程自动化先做 process mining / workflow baseline 和 decision inventory
Human review as magic所有风险都写“人工审核”明确 reviewer 能看到什么、多久看完、如何覆盖、如何追责
AI for everything low-value员工变成 AI 输出搬运工保留高价值学习任务, 把员工转向 judgment、exception、quality improvement
No task retirementAI 新增任务, 旧任务不减少明确哪些手工整理、重复查询、重复录入被移除
Manager metrics unchanged仍按旧 AHT / volume 管理改为 balanced metrics, 防止员工为速度牺牲质量和复核

6. Accountability Design

AI 可以执行任务, 但不能承担组织责任。Accountability design 的目标是让每个结果都能回答:

谁对业务结果负责?
谁负责复核?
谁能 override?
谁保留证据?
谁接受残余风险?
谁在失效时停机、补救和改进?

6.1 Accountability principles

原则说明失败表现
Accountability stays human and organizational责任落在授权角色和组织治理上“AI 建议错了”被当成事故根因终点
Review authority must match risk复核人必须具备能力、权限和时间junior agent 被要求批准复杂合规结论
Override must be legitimate人可以覆盖 AI, 也可以覆盖低层级人工决定, 但要有理由和审计override 被用来绕过控制或追求 KPI
Evidence must be durable证据能复盘当时谁看见什么、AI 输出什么、人改了什么只保存最终文本, 无来源和修改记录
Risk acceptance is explicit上线和例外风险由授权 owner 接受产品团队默认把风险推给一线员工

6.2 Accountability matrix

Decision / actionAccountable ownerReviewerOverride authorityEvidence ownerEvidence retained
AML alert closeAML operations managerAML investigator / QA sample reviewerAML supervisorCompliance operationsalert evidence、AI summary、investigator rationale、override reason、case status
SAR escalation considerationAML compliance officerSenior investigatorMLRO / BSA officer equivalent by jurisdictionFinancial crime compliancenarrative draft、supporting transactions、policy basis、approval record
Credit approval recommendationCredit policy owner / lending business ownerUnderwriterCredit committee or delegated authorityCredit operationsapplication facts、policy rule id、model score, AI memo、human decision、reason code
Adverse action explanationLending compliance ownerUnderwriter / compliance review sampleCompliance leadCredit compliancefinal reason codes、AI suggestions、human edits、customer communication
Customer service responseService operations ownerFrontline agent / QA reviewerTeam lead / complaint specialistContact center operationstranscript、knowledge sources、agent edits、customer visible message
Payment dispute temporary creditDispute operations ownerDispute analystSupervisor by amount and rulePayment operationsscheme rule source、case evidence、approval, reversal path

6.3 Who is responsible for the result

Result typeResponsible partyDesign implication
Customer-visible communicationBusiness process owner and acting employeeAI draft must be editable, source-linked, reviewed by risk tier
Formal decisionDelegated decision ownerAI can support or recommend, but decision log must name human authority or approved automated policy
Automated low-risk actionProduct / operations owner plus control ownerAutomation boundary, rollback, monitoring and audit evidence must be approved before scale
Model or AI behaviorAI product owner / platform owner / model ownerEval, drift, prompt / model change, incident response and vendor change control are required
Workflow redesign outcomeTransformation lead and business ownerBaseline, adoption, quality, risk and workload metrics must be managed together

6.4 Who reviews

Review design must specify coverage, sampling, skill, SLA and calibration.

Review pattern用法控制
100% pre-effect review高影响、不可逆、客户权益相关输出reviewer checklist、evidence view、approval log
Risk-tiered review中等风险高量任务risk score threshold、queue prioritization、sample expansion trigger
Post-action QA sample低风险、可逆、海量场景statistically meaningful sample、defect taxonomy、feedback loop
Peer calibration专业判断一致性训练blind review、disagreement analysis、rubric refinement
Supervisor exception review例外、高金额、政策冲突、客户投诉escalation packet、decision SLA、override log

6.5 Who can override

Override 不是绕过系统, 而是授权控制。

Override target谁能覆盖必填证据系统动作
AI recommendationActing employee within delegationreason code、free-text rationale for high risklog diff、track adoption and disagreement
AI automated actionSupervisor or control owner by thresholdcustomer impact、rollback feasibility、root causepause automation if pattern repeats
Human frontline decisionSupervisor / specialistpolicy basis、case evidence、customer impactreopen case、notify downstream if needed
Policy rule outputPolicy owner / risk ownerrule exception、approval scope、expiry dateexception registry、sample review
Production AI capabilityBusiness owner / platform owner / risk ownerincident or threshold breachfeature flag off、model route freeze、fallback workflow

6.6 Who retains evidence

Evidence retention is a product requirement, not an afterthought.

Evidence objectWhy it mattersRetention ownerArchitecture requirement
Input facts证明 AI 和人依据什么做判断Data owner / system ownerimmutable snapshot or reproducible reference
Retrieved sources证明引用政策和知识是否有效Knowledge ownersource id、effective date、version
AI output证明模型给出什么建议AI platform ownermodel id、prompt version、output hash
Human edit / decision证明人如何使用或覆盖 AIOperations ownerdiff、reason code、reviewer id
Tool action证明系统状态如何改变Application owneractor、resource、action、before / after state
Monitoring signal证明控制是否持续运行Risk / platform / opsmetric snapshot、threshold、alert disposition

7. Metrics

AI 知识工作重设计不能只用 productivity 指标。金融零售至少要平衡 productivity、quality、risk、learning、adoption 和 employee load。

7.1 Balanced metric model

Metric family问题示例指标反向信号
Productivity是否更快、更少返工、更高吞吐cycle time、touch time、cases per FTE、first contact resolution、automation yield速度提升但 QA defect、complaint、override 同时上升
Quality输出是否更准确、更完整、更一致first pass quality、source-supported claim rate、case note completeness、calibration agreementAI 文本更长但关键证据缺失
Risk是否降低或至少不增加客户、合规、模型和运营风险false close、false approve、under-escalation、policy violation、incident count、audit replay success低风险队列处理快, 高风险积压
Learning员工和系统是否在变聪明calibration improvement、challenge case pass rate、feedback-to-fix cycle、new skill attainment员工无法解释 AI 错误原因
Adoption员工是否在正确任务中正确使用active qualified users、task-level adoption、accepted with edit、rejected with reason、fallback use高 adoption 但低 edit / challenge, 可能是过度信任
Employee load员工负荷是否健康queue load、interruptions、review time sufficiency、after-hours work、cognitive load survey、burnout risk signalAI 增加监控和复核负担, 旧工作没有减少

7.2 Productivity metrics

MetricGood useRisk
Cycle time衡量端到端客户等待时间如果只优化单点, 可能把瓶颈转移到复核队列
Touch time衡量人工实际处理时间不能替代质量和客户影响
First pass yield衡量一次通过率和返工减少需要清楚 defect taxonomy
Automation yield衡量 AI 自动处理的有效比例不能把高风险自动化率当成目标
Queue age衡量积压和 SLA 风险需要按 risk tier 分层

7.3 Quality metrics

MetricDefinition金融零售用法
Source-supported claim rate输出中关键陈述有可追溯来源的比例客服、信贷、AML narrative 必须减少无来源断言
Human edit distance人对 AI 草稿的修改程度高修改可能代表 AI 质量低, 低修改也可能代表 rubber stamp
Calibration agreement多个 reviewer 对同一 case 的一致性underwriter、AML investigator、QA reviewer 训练
Critical omission rateAI 摘要遗漏关键 red flag 的比例AML、投诉、困难援助、欺诈场景
Formal defect rate正式记录中被 QA / audit 发现的缺陷case note、reason code、customer letter

7.4 Risk metrics

MetricDefinitionEscalation trigger
False close / false approve被错误关闭或错误批准的比例高风险队列出现连续缺陷
Under-escalation rate应升级但未升级的 case 比例投诉、欺诈、AML、困难援助显著上升
Override concentration某团队或个人 override 异常集中可能存在培训、激励或控制绕过问题
Policy violation rate输出违反政策或合规话术触发知识源和 prompt / policy review
Audit replay success审计能否复现当时依据和决策低于门槛时暂停 scale
Incident escape rate问题被客户、监管或下游团队先发现说明内部 QA / monitoring 失效

7.5 Learning metrics

MetricWhy it matters
Challenge case pass rate证明员工能识别 AI 错误、缺证据、过度自信和越权建议
Feedback-to-fix lead time衡量一线反馈进入知识、流程、模型或政策修复的速度
Skill mix shift衡量员工从资料整理转向判断、例外处理、客户沟通和质量改进
Reviewer calibration drift发现复核标准随着 AI 输出风格变化而漂移
New hire ramp quality防止新人只会操作 AI, 不懂业务原则

7.6 Adoption metrics

MetricGood interpretationWarning
Task-level adoption哪些任务中 AI 被使用只看登录人数没有意义
Accepted with no editAI 输出可直接使用的比例在高风险场景过高可能是默认采纳
Accepted with edit员工有效利用 AI 并加入判断需要分析修改类型
Rejected with reason员工识别 AI 不适用过低代表不敢拒绝或 UI 不支持
Fallback / escalation use员工知道何时退出 AI过低可能代表隐藏风险, 过高可能代表 AI 适用范围过宽

7.7 Employee load metrics

Metric设计意图
Review time per case复核时间是否足够支持真实判断
AI interruption countAI 建议是否增加干扰和上下文切换
Queue pressure by risk tier高风险任务是否被低风险高量任务挤压
After-hours correction work员工是否把 AI 返工带回非工作时间
Cognitive load pulse员工是否理解新责任、证据和控制界面
Manager coaching time主管是否有时间做校准和反馈, 而不是只追产能

7.8 Metrics anti-patterns

Anti-pattern风险
只看节省小时数会忽略复核、返工、质量、风险和员工负荷
只看 adoption高使用率可能来自管理压力, 不代表价值
只看模型准确率模型指标不能替代流程结果和客户结果
只看平均值高风险尾部和少数客户伤害会被掩盖
用旧 KPI 管新工作员工会用 AI 提速旧流程, 而不是改变工作方式

8. Financial Retail Case: AML Analyst / Credit Underwriter / Customer Service Agent

8.1 AML Analyst 工作重设计

AS-IS pain

Pain point影响
资料收集和系统切换占用大量时间investigator 判断时间被压缩
Case narrative 质量不一致supervisor review 和 audit 压力大
Alert 优先级不清高风险 case 可能被低价值队列拖延
新人依赖 SOP 但缺少判断训练calibration 不稳定
证据链分散复盘和监管询问成本高

Redesigned role-task architecture

TaskAllocationHuman role shiftArtifactControl
Alert intake and enrichmentAI-assisted从手工查资料转向验证证据完整性Enrichment packetsource list、data freshness、missing evidence flag
Transaction pattern summaryAI-assisted从汇总交易转向识别解释和异常Pattern summarylinked transactions、aggregation rule
Risk hypothesis generationAI-recommended从空白分析转向挑战假设Risk hypothesis listalternative hypothesis、unsupported claim flag
Case narrative draftAI-assisted从写初稿转向判断叙事是否完整Draft narrativeedit diff、citation requirement
Close / escalate decisionHuman-onlyinvestigator 承担风险判断Decision memoreason code、supervisor sample
SAR considerationHuman-onlysenior investigator / compliance 承担正式升级Escalation packetapproval workflow、evidence binder
Quality monitoringAI-monitoredQA 从抽样检查转向趋势和校准Quality dashboarddefect taxonomy、calibration session

Metrics

FamilyMetric
Productivityalert cycle time by risk tier、touch time reduction、queue age
Qualitynarrative completeness、source-supported claim rate、supervisor return rate
Riskfalse close sample rate、under-escalation、SAR escalation quality
Learninginvestigator calibration agreement、challenge case pass rate
Employee loadreview time sufficiency、high-risk queue pressure

Architecture controls

AML case UI
-> entitlement check
-> customer / transaction evidence package
-> typology and policy retrieval
-> AI summary and narrative draft
-> investigator evidence view
-> close / escalate decision log
-> supervisor review queue
-> audit evidence binder
-> QA and calibration dashboard

8.2 Credit Underwriter 工作重设计

AS-IS pain

Pain point影响
文档审查和资料缺口识别耗时审批周期长
政策解释和 reason code 使用不一致合规和客户沟通风险
Underwriter 被迫在速度和质量之间权衡可能出现浅层复核
模型分数、政策规则和人工判断分散难以解释最终决定
新人训练依赖经验传授标准化难

Redesigned role-task architecture

TaskAllocationHuman role shiftArtifactControl
Document extractionAI-assisted从录入字段转向验证关键事实Verified fact sheetsource highlight、confidence flag
Missing document detectionAI-recommended从人工逐项查清单转向确认缺口Requirement checklistDMN completeness rule
Policy eligibility checkAI-decided with human review or AI-recommended从查政策转向解释例外Eligibility resultrule id、version、input facts
Credit memo draftAI-assisted从写作转向风险判断和客户解释Underwriting memocitation、edit diff
Approval / decline decisionHuman-only within delegated authority保留最终授信责任Decision recordauthority check、reason code
Adverse action reasonHuman-only with AI assistance确认原因准确、可解释、合规Customer communicationcompliance review sample
Portfolio quality monitoringAI-monitored管理层看趋势而非只追产能Quality dashboardexception trend、fairness review input

Metrics

FamilyMetric
Productivityapplication cycle time、document completeness first pass、touch time
Qualitypolicy rule consistency、reason code accuracy、memo completeness
Riskfair lending review signal、override concentration、appeal / complaint trend
Learningunderwriter calibration、policy quiz and case drill score
Employee loadqueue pressure、manual evidence search time、review interruption count

Architecture controls

Credit workflow
-> applicant data and document package
-> OCR / extraction with source highlight
-> DMN policy decision service
-> AI memo drafting service
-> underwriter decision UI
-> authority and reason-code validation
-> customer communication control
-> audit log and quality sample

8.3 Customer Service Agent 工作重设计

AS-IS pain

Pain point影响
Agent 在多个知识库和系统间切换AHT 高, 体验不一致
话术和政策更新不同步客户可见错误和投诉风险
转人工或转二线时上下文丢失客户重复描述问题
复杂投诉、欺诈、困难援助识别不稳定高风险客户旅程延误
QA 事后发现问题, 修复慢同类错误重复发生

Redesigned role-task architecture

TaskAllocationHuman role shiftArtifactControl
Intent detectionAI-decided with human review for uncertain casesagent 从识别入口转向确认意图和风险Intent tagconfidence threshold、fallback
Knowledge retrievalAI-assisted从搜索知识库转向判断来源适用性Source-linked answerknowledge owner、effective date
Response draftingAI-assisted从写话术转向同理心、客户上下文和合规确认Agent response draftapproved language、edit log
Simple status answerAI-automatedagent 从低价值查询中释放Customer responselow-risk scope、audit transcript
Complaint / fraud / hardship routingAI-recommended with mandatory human confirmationagent 负责识别和保护客户权益Escalation casetrigger list、handoff packet
Handoff summaryAI-assisted二线获得完整上下文Handoff notetranscript summary、unresolved issue
QA monitoringAI-monitoredQA 从随机听录音转向风险抽样和模式修复QA dashboarddefect taxonomy、feedback-to-fix

Metrics

FamilyMetric
Productivityfirst contact resolution、AHT by intent、handoff repeat rate
Qualitypolicy-correct response rate、source-supported answer rate、QA defect rate
Riskcomplaint under-escalation、privacy incident、wrong commitment rate
Learningagent correction quality、knowledge feedback closure
Employee loadcontext switching count、AI suggestion interruption、after-call work

Architecture controls

Contact center desktop
-> customer identity and consent state
-> intent and risk classification
-> RBAC-filtered knowledge retrieval
-> AI draft with source and limitation
-> agent edit / send / escalate controls
-> handoff context packet
-> transcript and audit record
-> QA, knowledge and policy feedback loop

8.4 Cross-case lessons

LessonAMLCreditCustomer service
AI should absorb low-value cognitive loadevidence summarydocument extractionknowledge retrieval
Humans should own judgment and exceptionsclose / escalateapprove / declinecomplaint / fraud / hardship
Controls must be in workflowsupervisor reviewauthority checkescalation trigger
Metrics must balance speed and riskfalse closereason accuracywrong commitment
Training must preserve expertisetypology drillspolicy calibrationcomplex case simulation

9. Artifact Templates

以下模板不是空表, 而是可直接改写的结构和示例。正式项目应把示例行替换为真实 use case 记录, 并保留 owner、版本和批准记录。

9.1 Role-task matrix

用途: 把岗位拆成任务, 标注 AI 分配模式、风险、产物、控制和指标。

RoleJob outcomeTaskCurrent painAI allocationHuman responsibilityArtifactControlMetric
AML investigator正确处置 alert 并形成可审计结论汇总客户交易模式多系统查询耗时AI-assisted验证摘要, 识别异常, 作出处置判断Pattern summarysource-linked transaction listnarrative completeness、false close sample
Credit underwriter在授权范围内作出可解释授信决定识别资料缺口手工 checklist 易漏项AI-recommended确认缺口并决定是否要求补件Completeness checklistDMN rule id、source highlightfirst pass completeness
Service agent快速、准确、可升级地解决客户问题草拟政策解释知识库搜索慢且话术不一致AI-assisted确认来源、语气和客户上下文后发送Customer response draftapproved source、agent edit logQA defect、wrong commitment

字段说明:

Field写法
Role使用真实授权角色, 不使用泛泛的“用户”
Job outcome写业务结果, 不写工具使用目标
Task写可观察的工作单元
Current pain写基于流程事实的痛点
AI allocation从六类 task allocation 中选择
Human responsibility写人保留的判断、复核或责任
Artifact写正式或半正式输出
Control写强制执行的流程、产品或架构控制
Metric同时覆盖效率、质量或风险

9.2 Human-AI responsibility matrix

用途: 明确人、AI、规则服务、系统和控制 owner 的边界。

Work objectAI responsibilityHuman responsibilityRule / policy serviceSystem controlAccountable owner
AML case summary生成来源链接摘要和异常候选验证关键事实、补充上下文、决定处置typology guidance、escalation criteriacitation required、unsupported claim flagAML operations manager
Credit eligibility解释政策条件和候选 reason code确认事实、处理例外、作出授权决定DMN eligibility rulesauthority check、reason-code validationLending business owner
Customer response草拟基于知识库的答复编辑语气、确认适用性、识别升级approved knowledge base、compliance language rulessend control、handoff triggerService operations owner
Automated status update在低风险范围内更新 case 状态处理异常和客户异议workflow transition policytool gateway、rollback logApplication owner / ops owner

责任边界句式:

AI is responsible for producing a bounded, source-linked draft or recommendation.
The human role is responsible for validating evidence, applying judgment, handling exceptions, and accepting the business decision within delegated authority.
The system is responsible for enforcing permissions, workflow transitions, evidence capture, monitoring, rollback and auditability.

9.3 Training / adoption plan

用途: 把 adoption 从“培训完成率”升级为任务级能力迁移和管理节奏。

PhaseAudienceCapability goalTraining methodAdoption evidenceManager cadence
Role readinessAML / credit / service team leads理解 role-task changes、风险边界和指标变化leader workshop with case walkthroughsigned operating model、metric baselineweekly calibration review
Task skill buildfrontline employees能使用 AI 摘要、建议、草稿, 并识别错误guided task simulation、side-by-side old vs new workflowsimulation score、challenge case passdaily floor support for first 2 weeks
Judgment calibrationspecialists and reviewers对高风险 case 保持一致判断blind review、disagreement discussioncalibration agreement、override analysisbiweekly case clinic
Control operationsupervisors / QA / risk liaison会处理 override、incident、rollback 和 quality trendcontrol drill、audit replay exerciseaudit replay success、incident response timemonthly risk / quality forum
Continuous learningproduct, ops, knowledge owner把一线反馈变成知识、流程和模型改进feedback triage、release note reviewfeedback-to-fix lead time、defect reductionmonthly improvement backlog

Adoption 不能只用登录率衡量。每个 phase 都要有:

Evidence type示例
Task usage交易摘要、政策检索、草稿生成、升级包生成的任务级使用
Quality proofQA defect 下降、source-supported answer 上升
Risk proofunder-escalation 没有上升, audit replay 能复现
Learning proofchallenge case pass rate、calibration agreement 提升
Load proof员工没有因复核和反馈增加不可持续负荷

9.4 Workload-risk dashboard

用途: 管理层同时看工作量、风险、采用、质量和员工负荷, 防止 AI 项目只报告效率收益。

Dashboard sectionMetricSliceAction when signal worsens
Workloadqueue age、cases per FTE、touch timerisk tier、team、channel、product调整队列、暂停低价值自动化实验、补充 reviewer capacity
QualityQA defect、critical omission、source-supported claimtask type、AI feature、reviewer group触发知识源修复、prompt / workflow change、retraining
Riskfalse close、under-escalation、policy violation、incidenthigh-risk customer segment、decision type扩大人工复核、降低自动化范围、开 incident review
Adoptionaccepted with edit、rejected with reason、fallbackrole、task、manager group做 floor coaching, 修正 UI 或 incentive
Learningchallenge case pass、calibration agreementtenure、team、role增加 case clinic, 调整培训和 rubric
Employee loadreview time sufficiency、interruptions、after-hours correctionteam、shift、queue移除旧任务, 降低 AI interruption, 调整 staffing
Control healthaudit replay success、override reason completeness、rollback successuse case、release version暂停 scale, 修复 evidence capture 和 workflow gate

Dashboard interpretation rules:

PatternInterpretationResponse
Productivity up, quality downAI 提速但判断或证据质量受损缩小自动化范围, 加强 evidence view 和 QA
Adoption high, override low可能是真正好用, 也可能是过度信任抽样检查 no-edit cases, 做 challenge case drill
Workload down, employee load up旧工作减少但复核、反馈、监控负荷上升重新分配任务, 移除重复记录和无价值审批
Risk defects concentrated by manager group指标、培训或管理压力导致行为差异manager coaching, metric correction, targeted audit
Audit replay failure证据链不完整停止扩大范围, 修复 trace schema 和 evidence retention

10. Interview Answers

10.1 30 秒版本

AI 转型不是给员工一个 Copilot, 而是重新设计知识工作的 role-task architecture。我会先做 job / task decomposition, 判断每个任务适合 human-only、AI-assisted、AI-recommended、AI-decided with human review、AI-automated 还是 AI-monitored。然后明确谁对结果负责、谁复核、谁能 override、谁保留证据, 再用 productivity、quality、risk、learning、adoption 和 employee load 指标验证。金融零售场景尤其要把 workflow、decision accountability、human-AI interaction 和 architecture controls 串起来, 否则 AI 只会制造局部效率和系统性风险。

10.2 2 分钟版本

我不会从“哪个岗位会被 AI 替代”开始, 而会从任务架构开始。一个 AML analyst、credit underwriter 或 service agent 的工作, 其实由资料收集、摘要、判断、客户沟通、系统记录、复核和升级组成。AI 对这些任务的适配程度不同: 摘要、检索、抽取、草拟通常适合 AI-assisted;排序和下一步建议可以是 AI-recommended;低风险、可逆、规则明确的状态更新可以 AI-automated;但高影响决策、风险接受、客户权益和监管签字通常必须 human-only 或至少 human review。

设计时我会把 role、job、task、decision、artifact、handoff、control、metric 连起来。比如信用审批中, AI 可以提取收入证明、发现缺失资料、草拟 underwriting memo, 但 underwriter 仍要在授权范围内确认事实、选择 reason code、作出批准或拒绝决定。系统必须保存来源、模型输出、人工修改、审批人、政策版本和下游动作。管理指标也要从单纯 AHT 或处理量, 扩展到 first pass quality、source-supported claim、override pattern、audit replay success、employee load 和 learning curve。这样 AI 才是工作系统的一部分, 不是旁路工具。

10.3 Transformation Lead 版本

作为 Transformation Lead, 我会把 AI knowledge work redesign 当成 operating architecture 项目, 而不是工具上线项目。第一步是建立 use case portfolio 和流程 baseline: 哪些工作量大、返工多、风险高、员工负荷重, 哪些任务有清晰证据和规则。第二步是 role-task architecture: 拆 role、job、task、decision、artifact、handoff、control、metric, 并为每个任务选择 human-only、AI-assisted、AI-recommended、AI-decided with human review、AI-automated 或 AI-monitored。

第三步是 accountability design: 业务 owner 对结果负责, reviewer 对复核质量负责, policy owner 对规则负责, platform owner 对 AI 行为和日志负责, risk / compliance 对控制有效性提出要求但不替业务接受结果。每个高风险动作都要说明谁能 override、谁保留证据、谁有停机权。第四步是 architecture controls: identity and entitlement、DMN / policy service、AI orchestration、tool gateway、evidence store、workflow engine、audit trail、monitoring dashboard 和 rollback。第五步是 adoption operating rhythm: manager calibration、case clinic、challenge cases、QA feedback-to-fix、workload-risk dashboard。

我会用 balanced metrics 管理收益: productivity 看 cycle time 和 touch time, quality 看 source-supported claim 和 first pass quality, risk 看 false close、under-escalation 和 audit replay, learning 看 calibration 和 challenge case, adoption 看 task-level usage 和 reject / edit 行为, employee load 看 review time、interruptions 和 after-hours correction。这样才能证明 AI 不是把压力转嫁给员工, 而是真正重构了金融零售知识工作。

10.4 常见追问

追问回答要点
如何判断任务能不能自动化看客户影响、可逆性、证据质量、政策清晰度、bias / fairness 敏感性、时间压力、人工判断价值和审计要求。
如何避免员工被 AI 降维成审核按钮保留 judgment、exception、customer empathy、quality improvement 和 risk escalation 任务, 同时减少旧的重复查找和录入任务。
如何防止 adoption 指标误导不只看使用率, 还看 accepted with edit、rejected with reason、fallback、quality defect、override pattern 和 employee load。
如何向高管解释 ROIROI 必须连接 baseline、workflow change、adoption proof、quality proof、risk proof 和 unit economics, 不能只用节省小时数。
如何处理员工担忧不用泛泛安抚, 而是明确任务如何变化、哪些责任保留、哪些技能升级、哪些旧任务移除、绩效指标如何调整。
架构师在这里的价值是什么把角色任务边界转成 workflow gate、policy service、tool gateway、evidence store、audit trail、monitoring 和 rollback, 让责任和控制可执行。

11. 作品集表达方式

如果把本文能力放进作品集, 最好展示一套可交付证据:

Artifact证明的能力
AS-IS / TO-BE role-task architecture能从岗位视角重构知识工作
Task allocation matrix能判断 AI、人、规则和系统的任务边界
Accountability and override matrix能设计责任、授权、复核和证据
Workflow and architecture control diagram能把组织设计落到产品和系统控制
Balanced metric dashboard能证明效率、质量、风险、采用和负荷一起改善
Training and adoption plan能让员工能力迁移, 而不是只完成工具培训
Financial retail case pack能把 AML、credit、customer service 等场景讲成可落地方案

一句 portfolio pitch:

I redesigned AI-assisted financial retail knowledge work by decomposing roles into tasks and decisions, assigning human-AI responsibilities by risk, embedding accountability and evidence controls into workflow architecture, and measuring productivity, quality, risk, adoption, learning and employee load together.