AI JTBD / Outcome-Driven Innovation Playbook
以下来源作为 JTBD、ODI 和 AI 风险治理的学习锚点。本文把它们转成金融零售 AI use case 选择、需求证据、自动化边界和产品架构设计语言, 不构成法律、合规、审计、模型验证或监管解释意见。访问日期按 2026-06-29 记录。
AI JTBD / Outcome-Driven Innovation Playbook
面向对象: 已具备 CBAP / 高级 BA / AI PM / Product Architect / Solution Architect / AI Platform PM / 金融零售 AI 转型负责人。 核心问题: 如何用 Jobs-to-be-Done 和 Outcome-Driven Innovation 选择真正值得做的 AI use case, 而不是被 demo、模型能力、供应商话术或内部热点牵着走。 使用方式: 每个候选 AI use case 至少产出一份 AI JTBD Canvas、一张 Outcome Opportunity Scorecard、一份 Automation Boundary Contract、一份 Evidence-to-Architecture Traceability Matrix 和一页 Use Case Selection Memo。 训练定位: 本文不讲基础用户访谈、persona 入门或需求收集套路, 重点训练 CBAP 之后的高级问题定义、outcome 建模、证据驱动取舍、自动化边界和产品架构判断。
Source Anchors
以下来源作为 JTBD、ODI 和 AI 风险治理的学习锚点。本文把它们转成金融零售 AI use case 选择、需求证据、自动化边界和产品架构设计语言, 不构成法律、合规、审计、模型验证或监管解释意见。访问日期按 2026-06-29 记录。
| Anchor | Official / primary source | 本文使用方式 |
|---|---|---|
| HBR Jobs to Be Done - Christensen et al. | https://hbr.org/2016/09/know-your-customers-jobs-to-be-done | 用“客户做选择是为了完成某个 job”的视角, 把 AI idea 从功能愿望改写成稳定的 job、struggle、progress 和 switching trigger。 |
| Strategyn JTBD / ODI official | https://strategyn.com/jobs-to-be-done/ | 用 ODI 的 job map、desired outcome、importance / satisfaction 和 underserved opportunity 思路, 建立 AI use case 选择和排序方法。 |
| NIST AI Risk Management Framework | https://www.nist.gov/itl/ai-risk-management-framework | 用 Govern / Map / Measure / Manage 组织 AI 风险、证据、评估、上线门禁、持续监控和责任边界。 |
| NIST AI RMF Generative AI Profile | https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence | 用 GenAI 特有风险补强 hallucination、data leakage、prompt injection、over-reliance、tool misuse、content provenance 和 misuse 的筛选门槛。 |
| CFPB Chatbots in consumer finance | https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/ | 用于客服 AI 的 outcome 和风险边界: AI 不能以效率为名降低客户获得及时、清晰、人工支持和救济的能力。 |
| CFPB Circular 2022-03 complex algorithms and adverse action | https://www.consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms/ | 用于信贷申请人场景: 复杂算法不免除提供具体、准确 adverse action reason 的义务, 因此 LLM 不能替代可追溯的 reason pipeline。 |
Source-to-artifact mapping:
| Source lens | 可以产出的 artifact | 高级面试表达 |
|---|---|---|
| HBR JTBD | Job definition、switching trigger map、progress narrative | “我不会从 AI 功能开始, 我先定义客户或员工真正要完成的 job, 再判断 AI 是否能让 progress 更快、更可靠或更低风险。” |
| Strategyn ODI | Job map、outcome catalog、importance / satisfaction score、underserved opportunity map | “我用 desired outcome 和 underserved outcome 排序 AI use cases, 避免把资源投到看起来炫但结果已经被满足的场景。” |
| NIST AI RMF | AI risk map、measurement plan、release gate、monitoring gate、evidence binder | “我把 use case selection 同时放进价值和风险闭环, 不把 AI 选择简化成 ROI 或准确率排序。” |
| CFPB consumer finance sources | 客服与信贷客户影响边界、人工升级、具体原因链路 | “金融零售 AI 的好 use case 不是只降低处理成本, 还要保护客户获得解释、人工支持、纠错和救济的能力。” |
1. One-Sentence Positioning / 一句话定位
JTBD / ODI for AI Use Case Selection 的核心不是问“哪里能用 AI”, 而是:
Job to be done
-> job steps and desired outcomes
-> underserved outcomes with evidence
-> AI intervention point
-> automation boundary
-> product architecture and eval gate
-> portfolio decision: scale / pilot / redesign / stop
一句话:
AI JTBD / ODI = 用稳定的 job 和可度量的 desired outcomes 选择 AI 投资点, 用 underserved outcome 证明机会, 用 automation boundary 限定模型权力, 用产品架构和证据链把 idea 变成可上线、可审计、可持续改进的能力。
在金融零售场景里, 一个合格的 AI use case selection 必须同时回答:
- 这个 job 是谁在什么情境下必须完成的, 不是哪个部门想要 AI。
- 当前 job 的哪些 outcome 重要但满足度低, 有什么业务、风险或客户证据。
- AI 介入 job 的哪一步, 是 read、locate、summarize、draft、recommend、monitor、act 还是 decide。
- 哪些 decision、tool action、客户沟通、合规判断和责任边界不能自动化。
- 需要哪些数据、知识、规则、权限、workflow、UI、eval、monitoring 和 audit 组件。
- 什么时候 scale, 什么时候只做受控 pilot, 什么时候用规则/流程/数据质量改造替代 AI。
2. 为什么 CBAP 之后还需要 JTBD / ODI
CBAP 已经证明你掌握 stakeholder、elicitation、requirement lifecycle、strategy analysis、solution evaluation 和 change management。AI 时代的升级点不是再学习“怎么访谈用户”, 而是把已有 BA 能力提升为 outcome-first、evidence-first、architecture-aware 和 risk-aware 的选择系统。
| CBAP 已有能力 | AI use case 选择中的新挑战 | JTBD / ODI 升级方式 |
|---|---|---|
| Business need analysis | AI idea 常由 demo、vendor、内部热点或模型能力驱动 | 用 job 和 desired outcome 过滤 solution-first 需求 |
| Stakeholder requirement | 同一 AI 输出可能影响客户、员工、主管、模型风险、合规和审计 | 区分 job executor、buyer、operator、risk owner、affected customer 和 platform customer |
| Requirements elicitation | 业务方容易说“我要智能助手”, 而不是说清 job struggle | 从 job step、outcome、baseline、exception、evidence 和 switching trigger 提问 |
| Solution evaluation | AI 指标、业务 KPI、风险指标和 adoption 指标可能互相冲突 | 用 ODI opportunity + NIST risk tier + architecture feasibility 做组合判断 |
| Traceability | prompt、RAG、model、tool、policy、knowledge source 都会变化 | 建 requirement -> outcome -> eval case -> architecture component -> release gate 的 lineage |
| Change management | AI 上线后行为会漂移, 组织也会过度信任或绕过流程 | 把 monitoring、feedback、override、incident 和 eval update 写入生命周期 |
| Strategy analysis | AI portfolio 容易碎片化成一堆 point solution | 用 stable job 和 reusable capability 把 use cases 聚合到产品架构和平台路线图 |
高级表达:
CBAP 让我能把业务问题和 stakeholder concern 讲清楚;JTBD / ODI 让我能进一步证明哪个 outcome 真的 underserved, AI 应该介入哪一步, 自动化边界在哪里, 以及需要什么架构和证据才能把 use case 安全地 scale。
3. 核心框架: JOA Selection Loop
JOA = Job, Outcome, Architecture。它把 AI use case selection 从灵感筛选变成可复核的决策系统。
1. Define market as a job executor + job
2. Map the job steps
3. Write desired outcome statements
4. Score importance, satisfaction, evidence and risk
5. Identify underserved outcomes
6. Decide AI intervention point and no-AI alternative
7. Define automation boundary and human accountability
8. Map outcome to product architecture
9. Build eval and release gate
10. Make portfolio decision
3.1 JOA 分层
| Layer | 关键问题 | 主要产出 |
|---|---|---|
| L1 Market / Job Definition | 哪类人要完成哪个稳定 job | Job definition、job executor map、affected stakeholder map |
| L2 Job Map | job 的理想完成过程是什么 | Define / Locate / Prepare / Confirm / Execute / Monitor / Modify / Conclude map |
| L3 Outcome Catalog | 每一步如何衡量“更好地完成” | Desired outcome statements、outcome taxonomy |
| L4 Opportunity Scoring | 哪些 outcome 重要但满足度低 | Importance / satisfaction / evidence / risk-adjusted opportunity score |
| L5 AI Translation | AI 在哪一步提供 read、summarize、draft、recommend、act | AI intervention map、no-AI alternative |
| L6 Automation Boundary | AI 能做什么, 不能做什么, 何时升级人 | Automation Boundary Contract、decision rights matrix |
| L7 Product Architecture | 需要哪些系统、数据、模型、规则、工具和治理组件 | C4 / sequence sketch、control points、component responsibility |
| L8 Eval / Gate | 如何证明可以 pilot、scale 或 stop | Eval contract、release gate memo、monitoring gate |
| L9 Portfolio Decision | 投资、暂停、合并、平台化或退役 | Use case selection memo、portfolio heatmap |
3.2 选择原则
| Principle | 操作化解释 | 反例 |
|---|---|---|
| Job before AI | 先定义稳定 job, 再选择 AI pattern | “我们要做一个财富 GPT” |
| Outcome before feature | 先定义 desired outcome, 再写功能 | “加一个总结按钮” |
| Evidence before prioritization | 没有 baseline、案例、日志、QA、投诉、风险或财务证据, 不进入 scale | 只凭高管兴趣排优先级 |
| Boundary before autonomy | 先定义 AI 不能做什么, 再讨论自动化程度 | 先做 agent, 之后补审批 |
| Architecture before pilot scale | pilot 之前就要知道未来如何接入权限、日志、eval、monitoring 和工具网关 | PoC 能跑, 生产再说 |
| Portfolio before platform | 看到多个 job 共享能力后再平台化 | 第一个 demo 就建设大平台 |
4. JTBD 到 AI 的转译
JTBD 不是 persona, 也不是旅程图。它问的是: job executor 在一个情境中要完成什么进展, 他们用什么替代方案, 现有方案在哪些 outcome 上表现不足。AI 转译的关键是把 job step 改写成 AI 可以帮助或不能帮助的工作类型。
4.1 Job Step 到 AI Intervention
| JTBD step | job executor 在做什么 | AI 可以帮助 | AI 不应直接替代 |
|---|---|---|---|
| Define | 明确目标、限制、成功标准 | 解释政策、识别约束、生成检查清单、提示缺失信息 | 定义受监管决策目标、替代业务 owner 的风险接受 |
| Locate | 找到完成 job 所需输入 | 检索政策、交易、客户资料、案例、产品资料、监管文本 | 绕过权限、使用未经授权或过期来源 |
| Prepare | 整理、清洗、结构化输入 | 摘要、抽取字段、去重、实体匹配、预填表单 | 默默修改正式记录、隐藏不确定性 |
| Confirm | 确认信息足够、证据可信、约束满足 | completeness check、policy check、citation check、risk flags | 把缺证据判断成事实、跳过必要人工确认 |
| Execute | 完成核心任务 | 起草 memo、生成回复草稿、创建 case draft、建议下一步 | 高影响 approve / decline / file / waive / trade / close |
| Monitor | 观察执行是否正常 | 异常检测、队列优先级、SLA 风险、quality drift | 无人监督地改变客户权益或案件结论 |
| Modify | 根据异常调整方案 | 推荐补件、升级、重路由、请求二审、更新草稿 | 自动覆盖人工决定、绕开 maker-checker |
| Conclude | 完成记录、通知、归档、复盘 | 生成 case note、evidence bundle、QA summary、learning loop | 生成不可追溯的正式通知、删除或重写审计证据 |
4.2 AI Pattern 选择
| Outcome 类型 | 更可能适合的 AI pattern | 架构含义 |
|---|---|---|
| 信息查找慢 | RAG / semantic search / policy assistant | source-of-truth、权限过滤、版本生效日期、citation validator |
| 文档和案例太长 | Summarization / structured extraction | schema、evidence reference、human edit log、coverage check |
| 资料缺失多 | Completeness checker / rules + model | rule engine、document parser、missing information workflow |
| 判断不一致 | Recommendation with rubric / decision support | policy rules、case similarity、explanation、reviewer calibration |
| 草稿耗时高 | Drafting copilot | approved templates、tone policy、edit diff、approval before send |
| 队列优先级不清 | Predictive triage / anomaly detection | feature store、risk score、monitoring、bias and drift checks |
| 工具动作频繁且低风险 | Agent with approval or constrained automation | tool gateway、least privilege、idempotency、rollback、kill switch |
| 客户权益影响大 | Human decision with AI evidence support | strong HITL、reason pipeline、formal notice boundary、audit binder |
4.3 JTBD 转译公式
For [job executor],
when trying to [job],
at job step [step],
they struggle to [desired outcome not sufficiently achieved],
because [evidence-backed constraint],
so AI should [read / locate / summarize / draft / recommend / monitor / act-with-approval],
but must not [decision / action / communication boundary],
and success will be evaluated by [outcome metric + risk metric + adoption evidence].
例子:
| 场景 | 弱表达 | JTBD-to-AI 表达 |
|---|---|---|
| AML | 做一个 AML Copilot | 当 AML analyst 调查 alert 时, 在 locate / prepare / confirm 阶段难以及时找到跨账户证据和关键 red flags;AI 应汇总证据、生成引用和起草 narrative, 但不能关闭 alert、决定 SAR 或替代二审。 |
| 客服 | 做更智能的客服机器人 | 当客服处理费用、争议和投诉问题时, 在 define / locate / execute 阶段需要快速识别 regulated intent、引用有效政策并生成可审核回复;AI 可做 routing、知识检索和草稿, 但不能阻断人工支持或承诺未经授权的补偿。 |
| 信贷 | AI 帮客户申请贷款 | 当申请人完成信贷申请时, 在 prepare / confirm 阶段容易漏传资料和误解下一步;AI 可解释材料要求和预填非决策字段, 但不能给出正式批准概率、拒绝理由或替代 adverse action reason pipeline。 |
5. Outcome Statement 设计
ODI 的核心资产不是 pain point list, 而是 desired outcome statement。AI use case 必须从 outcome statement 开始, 否则后续 eval、架构和价值证明都会漂移。
5.1 Outcome statement grammar
建议使用下面语法:
Direction [minimize / increase / reduce / improve]
+ metric object [time, effort, likelihood, accuracy, completeness, risk, rework, variance]
+ job step object [to locate evidence, confirm policy applicability, complete application]
+ context [when / while / before / after]
+ constraints [without increasing customer harm, compliance risk, manual rework, latency, cost]
常用模式:
| Outcome type | Grammar | 金融零售例子 |
|---|---|---|
| Time | Minimize the time it takes to... | 降低 AML analyst 找到关键交易证据和关联实体的时间。 |
| Effort | Minimize the effort required to... | 降低客服为同一客户问题跨系统查找政策、账户状态和历史 case 的努力。 |
| Likelihood | Increase the likelihood that... | 提高信贷申请人在提交前发现资料缺失和字段不一致的可能性。 |
| Completeness | Increase the completeness of... | 提高财富顾问会前准备中 suitability checklist 和风险揭示覆盖度。 |
| Accuracy | Increase the accuracy of... | 提高客服回答中政策引用、费用规则和生效日期的准确性。 |
| Risk reduction | Reduce the risk that... | 降低 AI 或员工遗漏 AML red flag、误导客户、生成不可解释拒贷原因的风险。 |
| Variance | Reduce variance in... | 降低不同分析员、客服或顾问对同类政策问题处理口径的差异。 |
5.2 好 outcome 与弱 outcome
| 弱 outcome | 问题 | 高质量 outcome |
|---|---|---|
| 提升客服效率 | 不知道 job step、指标、风险边界 | 在信用卡费用咨询中, 降低客服查找有效政策和生成可审核回复草稿的时间, 同时 unauthorized fee waiver commitment 为 0。 |
| AML 更准确 | “准确”未定义 | 提高 AML alert investigation 中关键 red flag 被识别并附上 source evidence 的覆盖度, 同时 final disposition 仍由授权 analyst 完成。 |
| 信贷体验更好 | 缺少可测对象 | 提高申请人在提交前发现缺失材料、身份信息不一致和资格限制的可能性, 同时不输出正式批准承诺或拒绝原因。 |
| 财富顾问更会卖产品 | 销售导向且风险高 | 提高顾问在客户会谈前完成 suitability evidence、费用风险解释和替代产品比较的完整性, 同时不让 AI 直接向客户生成个性化交易建议。 |
5.3 Outcome statement 到 eval 的连接
| Outcome | Eval question | Release metric | Evidence |
|---|---|---|---|
| 降低证据查找时间 | AI 是否减少 analyst 跨系统查找次数 | Median evidence collection time、click path length | Case telemetry、pilot shadow study |
| 提高引用准确性 | AI 输出的关键事实是否有有效来源 | Citation precision、unsupported claim rate | Golden set、citation validator |
| 降低误导性客服承诺 | AI 是否生成未授权承诺 | Critical policy violation count | Red-team cases、QA review |
| 提高补件完整性 | 申请人是否更少因缺资料返工 | Missing document return rate | Application funnel log |
| 提高 suitability 覆盖 | 顾问准备是否覆盖风险、费用、客户约束 | Checklist completeness、supervisor QA | Advisor desktop log、review sample |
6. Underserved Outcome: 机会不是痛点, 是高重要低满足
Underserved outcome 是 AI use case selection 的核心。痛点可能很吵, 但不一定重要;重要 outcome 可能已经被规则、流程、培训或现有系统满足;AI 只有在某些 outcome 重要、满足度低、证据充分且风险可控时才值得进入投资队列。
6.1 Opportunity scoring
建议把 ODI 的 importance / satisfaction 思路扩展成 AI 适用的风险调整评分:
AI Opportunity Score =
Importance
+ max(Importance - Satisfaction, 0)
+ Evidence Strength
+ Strategic / Regulatory Relevance
+ Reuse Potential
- Automation Risk
- Data / Architecture Gap
- Adoption Friction
评分用 1-5 分即可, 不要制造虚假的小数精度。
| Dimension | 1 分 | 3 分 | 5 分 |
|---|---|---|---|
| Importance | 边缘便利 | 明显影响团队效率或客户体验 | 影响收入、风险、合规、客户权益或核心运营 |
| Satisfaction | 已被良好满足 | 局部可用但不稳定 | 满足度很低, 返工、投诉、遗漏或延迟明显 |
| Evidence strength | 只有主观意见 | 有案例样本或专家判断 | 有 baseline、日志、QA、投诉、损失、SLA 或审计发现 |
| Strategic / regulatory relevance | 非核心 | 支持部门目标 | 支持企业战略能力或监管关注点 |
| Reuse potential | 单一页面功能 | 可复用到相邻流程 | 可沉淀为企业 AI 平台能力 |
| Automation risk | 低风险可逆 | 有客户或运营影响 | 高影响、不可逆、监管敏感 |
| Data / architecture gap | 源清楚, 接入简单 | 部分源需要治理 | owner 不清、质量差、权限复杂 |
| Adoption friction | 嵌入现有流程 | 需要局部流程改变 | 需要大规模角色、激励或政策改变 |
解释:
| Score pattern | 决策 |
|---|---|
| 高 importance、低 satisfaction、强 evidence、低/中风险 | 优先 pilot, 同步设计 eval 和 architecture runway |
| 高 importance、低 satisfaction、强 evidence、高风险 | 可做 decision support / copilot, 自动化边界必须强, 先影子模式或受控 pilot |
| 高 importance、低 satisfaction、弱 evidence | 先做 measurement sprint, 不直接 build |
| 高 satisfaction、低 friction | 不优先 AI, 可用流程优化或现有系统增强 |
| 高 risk、低 evidence、低 reuse | 停止或重定义 job / outcome |
6.2 Underserved outcome map
| Outcome status | 含义 | AI 策略 |
|---|---|---|
| Underserved | 重要但当前满足度低 | 候选 AI opportunity, 进入证据和边界评审 |
| Overserved | 满足度高但成本或复杂度过高 | 考虑简化、自动化、self-service 或规则化, 不一定需要 LLM |
| Appropriately served | 重要且满足度合理 | 保持现状, 只做小改进或监控 |
| Mis-served | 看似满足, 但通过绕流程、人工补救或风险转移实现 | 优先做流程/数据/控制改造, AI 只能作为辅助 |
| Unserved / non-consumer | 现有方案让某群体无法完成 job | 可探索 disruptive / simpler AI-enabled workflow, 但必须先确认客户影响和合规边界 |
7. AI Automation Boundary
Automation boundary 是把 JTBD / ODI 变成金融零售 AI 产品的关键控制。它定义 AI 可以影响 job 到什么程度, 哪些动作必须由人、规则系统或正式决策引擎完成。
7.1 自动化等级
| Level | AI role | 允许行为 | 金融零售边界 |
|---|---|---|---|
| L0 No AI | 不使用 AI | 用流程、规则、培训、数据质量或系统改造解决 | 高风险且证据不足时优先考虑 |
| L1 Retrieve | 找资料 | 检索政策、案例、产品资料、客户可见 FAQ | 必须权限过滤、版本控制和来源引用 |
| L2 Summarize | 总结资料 | 摘要通话、case、交易、文档 | 不能省略关键风险、反向证据和不确定性 |
| L3 Draft | 草拟内容 | 草拟客服回复、AML narrative、信贷 memo、顾问会前材料 | 外发、归档或提交前必须审核 |
| L4 Recommend | 建议下一步 | 建议补件、升级、二审、调查方向、产品比较维度 | 必须给证据和理由, 高影响建议由人负责 |
| L5 Act with approval | 生成动作待批准 | 创建 case、发内部请求、更新低风险字段 | tool gateway、审批、日志、可撤销 |
| L6 Constrained autonomous act | 在强边界内自动执行 | 低风险、可逆、规则明确、可监控的内部动作 | 不适合自动拒贷、自动投资交易、自动关闭 AML case、自动阻断投诉 |
7.2 Boundary contract 必填项
| Boundary field | 必须写清 |
|---|---|
| Job scope | AI 支持哪个 job、哪一步、哪类用户和哪类案件 |
| Allowed inputs | AI 可访问的数据源、权限范围、版本、生效日期和保留要求 |
| Allowed outputs | AI 可生成的字段、格式、草稿、建议和证据 |
| Blocked outputs | 禁止生成的承诺、判断、建议、通知、原因、法律/合规结论 |
| Tool authority | 可调用工具、读写权限、审批要求、幂等性、回滚方式 |
| Human accountability | 谁审核、谁批准、谁覆盖、谁承担最终业务决定 |
| Escalation trigger | 低置信、缺证据、高风险意图、客户投诉、政策冲突、敏感属性、异常工具动作 |
| Evidence requirement | 每个关键输出需要哪些 citation、trace、score、review record |
| Monitoring trigger | 哪些线上信号触发停用、降级、回归评估或 incident |
| Change control | model、prompt、RAG、policy、tool 或 UI 变更如何触发回归 eval |
7.3 高风险边界红线
| Red line | 原因 | 更稳妥设计 |
|---|---|---|
| LLM 自动批准或拒绝信贷申请 | 客户权益、fair lending、adverse action reason 和模型风险高度敏感 | AI 做资料完整性、政策引用、memo 草稿, 决策由正式模型/规则/人审完成 |
| AI 直接向客户给个性化投资交易建议 | 适当性、销售合规、许可和客户损失风险 | AI 支持顾问准备材料和风险揭示, 客户建议由授权渠道输出 |
| AI 自动关闭 AML alert 或决定不报 SAR | 金融犯罪风险、审计和监管责任 | AI 提供证据包、red flag checklist、narrative draft, final disposition 由 analyst / supervisor 决定 |
| 客服 AI 阻断投诉或人工支持 | 客户救济、服务可得性和合规风险 | regulated intent 检测、显著人工入口、warm handoff、case ID |
| AI 平台默认允许业务 team 接任意模型和任意数据源 | 数据泄露、不可审计、供应商和模型风险 | model gateway、data classification、policy-as-code、eval gate、usage registry |
8. 需求证据与产品架构
高级 AI BA / PM 的差异化能力不是“写得出 AI user story”, 而是能把 outcome evidence 映射到 architecture decision。
8.1 Requirement evidence stack
| Evidence type | 可证明什么 | 例子 |
|---|---|---|
| Workflow telemetry | 哪一步耗时、返工、等待、跳出 | AML case handling time、客服转接次数、申请漏件率 |
| QA / audit sample | 哪类错误真实发生 | 错误政策引用、漏 red flag、误导性客服承诺 |
| Complaint / escalation log | 客户或员工在哪些结果上不满意 | chatbot 无法转人工、信贷申请状态不清、顾问材料解释不足 |
| Policy / control evidence | 哪些边界不能被 AI 越过 | adverse action、suitability、AML escalation、fee waiver authority |
| Financial baseline | outcome 与成本、损失、收入、资本或运营容量的关系 | 平均处理成本、SLA breach penalty、manual review backlog |
| Expert calibration | 专家对 outcome 的重要性和满足度判断 | senior analyst review、credit risk SME、compliance opinion |
| System feasibility evidence | 数据、API、权限、日志、工具是否可用 | source inventory、API readiness、data quality profile |
| Pilot / shadow evidence | AI 在受控环境是否改善 outcome | shadow mode comparison、human override analysis、eval report |
8.2 Evidence-to-architecture traceability
| Outcome evidence | 架构决策 | 需要的 control |
|---|---|---|
| 关键事实经常找不到来源 | RAG + citation + source freshness | source registry、retrieval permission、citation validator |
| 不同员工处理口径不一致 | policy rules + approved response templates | versioned policy, template governance, QA sampling |
| 资料缺失导致返工 | extraction + completeness checker + workflow task creation | schema validation、human confirmation、exception queue |
| 高风险建议需要审核 | copilot + review queue + maker-checker | decision log、override reason、approval authority |
| 工具动作可能造成客户影响 | tool gateway + least privilege + approval | idempotency、rollback、kill switch、action audit |
| 模型行为随版本变化 | model gateway + eval harness + regression suite | model registry、prompt version、release gate |
| 业务方要证明价值 | outcome dashboard + cohort comparison | baseline、pilot design、adoption telemetry |
8.3 Product architecture minimum set
| Component | 作用 | 选择问题 |
|---|---|---|
| Experience surface | 员工 copilot、客户 chatbot、advisor desktop、platform portal | AI 是嵌入工作流还是新入口 |
| Workflow integration | case management、CRM、LOS、advisor platform、ticketing | AI 输出如何进入下一步, 谁负责确认 |
| Knowledge / data layer | policy docs、product master、transaction history、customer profile、case history | source-of-truth、权限、版本和数据最小化是否清楚 |
| Model gateway | model routing、fallback、rate limit、logging、vendor abstraction | 是否需要多模型、成本控制、可替换性 |
| Retrieval / context service | query rewrite、authorization-aware retrieval、citation、freshness | 是否能证明输出基于有效来源 |
| Rules / policy engine | eligibility、authority、regulated intent、sales boundary | 哪些必须由确定性规则控制 |
| Tool gateway | read / write tools、approval、rollback、idempotency | AI 能执行哪些动作, 怎样限制 |
| Eval harness | golden set、rubric、threshold、critical failure | outcome 能否被离线和在线评估 |
| Observability | trace、latency、cost、quality、drift、human override | 上线后如何发现问题 |
| Evidence binder | requirement、source、eval、approval、incident、change record | 审计和复盘如何重放 |
9. 金融零售案例
9.1 AML Analyst: Alert Investigation Copilot
| 维度 | 设计 |
|---|---|
| Job executor | AML analyst / financial crime investigator |
| Core job | 在有限时间内调查 alert, 判断是否有足够证据支持 escalation、closure、SAR consideration 或 further review |
| Job steps | Locate transaction / customer / counterparty evidence -> prepare entity timeline -> confirm red flags and policy applicability -> draft narrative -> supervisor review -> conclude and archive |
| Underserved outcomes | 降低查找跨系统证据时间;提高关联实体、异常交易模式和 prior case 被发现的可能性;降低 narrative 中缺 source evidence 的风险;减少 analyst 之间口径差异 |
| AI intervention | Retrieve、summarize、entity linking、case timeline、red flag checklist、narrative draft、QA pre-check |
| Automation boundary | AI 不关闭 alert, 不决定 SAR, 不替代 supervisor review, 不删除反向证据, 不把未引用事实写入正式 narrative |
| Product architecture | Case management integration + transaction graph + customer / counterparty profile + AML policy RAG + typology library + citation validator + analyst review UI + audit log + eval harness |
| Evidence | Case duration baseline、QA defect types、missed red flag sample、senior analyst calibration、shadow mode comparison、override reason analysis |
| Selection decision | 高价值但高风险;适合 L2-L4 copilot 和受控 pilot, 不适合 autonomous disposition |
Outcome examples:
| Outcome statement | Eval / metric |
|---|---|
| Minimize the time required for an analyst to locate transaction evidence and related entities when investigating an alert, without reducing evidence completeness. | Median evidence collection time, entity recall on golden cases, missing evidence defects |
| Increase the likelihood that key red flags are surfaced with source evidence before narrative drafting. | Red flag coverage, citation validity, senior analyst review score |
| Reduce variance in case narrative quality across analysts while preserving analyst accountability. | Narrative rubric score, edit distance, supervisor rework rate |
9.2 客服: Regulated Service Resolution Copilot
| 维度 | 设计 |
|---|---|
| Job executor | Contact center agent / digital service customer / supervisor |
| Core job | 快速、准确、合规地解决客户问题, 并在投诉、争议、困难援助、信贷、财富、欺诈等高风险意图出现时正确升级 |
| Job steps | Define intent -> authenticate / locate account context -> locate policy -> prepare answer or action -> confirm authority -> execute response / case creation -> monitor resolution -> conclude with record |
| Underserved outcomes | 降低跨系统查找政策和账户状态时间;提高 first-contact resolution;降低错误承诺和错误拒绝人工升级;提高投诉和争议识别率 |
| AI intervention | Regulated intent detection、policy RAG、customer context summarization、reply draft、next best action suggestion、handoff summary |
| Automation boundary | AI 不阻断人工支持, 不承诺未经授权的 fee waiver / credit / compensation, 不关闭投诉, 不替代正式通知, 不在未认证状态披露账户信息 |
| Product architecture | Omnichannel routing + identity / entitlement + CRM + policy RAG + approved response template + regulated intent classifier + human handoff + QA monitoring + complaint trigger log |
| Evidence | Call handle time、transfer rate、complaint keywords missed、QA policy defects、customer effort score、chatbot containment harm review |
| Selection decision | 适合从内部 agent assist 起步;客户可见 AI 必须按风险层级限制输出和人工入口 |
Outcome examples:
| Outcome statement | Eval / metric |
|---|---|
| Reduce the time required for an agent to locate the current policy and account context needed to answer fee and dispute questions. | Average handle time, policy lookup time, citation accuracy |
| Increase the likelihood that complaint, fraud, hardship and regulated credit intents are routed to the correct human or formal workflow. | Intent recall on high-risk class, false negative count, handoff SLA |
| Reduce unauthorized commitments in AI-drafted customer responses. | Critical policy violation count, QA defect rate, red-team cases |
9.3 信贷申请人: Application Completion and Explanation Assistant
| 维度 | 设计 |
|---|---|
| Job executor | Credit applicant / loan officer / credit operations team |
| Core job | 让申请人准确提交申请材料、理解流程状态和下一步, 同时确保正式信贷决策、理由和客户通知来自受控流程 |
| Job steps | Define product / eligibility context -> locate required documents -> prepare application -> confirm completeness -> submit -> monitor status -> modify with additional information -> conclude with decision / next steps |
| Underserved outcomes | 降低资料缺失和字段不一致;提高申请人理解补件要求的可能性;降低重复联系;降低用模糊解释替代正式 adverse action 的风险 |
| AI intervention | Document checklist、field explanation、document extraction draft、missing info detection、status explanation、loan officer memo draft |
| Automation boundary | AI 不承诺批准概率, 不生成正式拒绝原因, 不替代 adverse action reason pipeline, 不使用敏感属性做非授权推理, 不绕过 fair lending 和 credit policy controls |
| Product architecture | Application portal + document ingestion + OCR / extraction + rule-based completeness check + LOS integration + credit policy rules + adverse action reason service + human review + audit evidence |
| Evidence | Application fallout, missing document rate, call reasons, rework rate, fair lending control requirements, credit policy exception sample |
| Selection decision | 客户体验价值高;AI 适合辅助完成申请和解释流程, 不适合 LLM-only decisioning |
Outcome examples:
| Outcome statement | Eval / metric |
|---|---|
| Increase the likelihood that an applicant submits all required documents before formal review begins. | Missing document return rate, first-pass completeness, abandonment rate |
| Reduce applicant effort required to understand what information is missing and why it is needed. | Repeat contact rate, clarification request rate, customer effort score |
| Reduce the risk that customer-facing explanations conflict with formal credit decision reasons. | Reason consistency check, compliance QA, formal notice boundary violations |
9.4 财富顾问: Advisor Suitability Preparation Copilot
| 维度 | 设计 |
|---|---|
| Job executor | Wealth advisor / relationship manager / supervisor / client service associate |
| Core job | 在客户会谈前准备适当性证据、产品资料、风险揭示、费用说明和可讨论选项, 并在合规边界内与客户沟通 |
| Job steps | Define client objective -> locate profile / holdings / constraints -> prepare product comparison -> confirm suitability evidence -> draft meeting notes / disclosure -> execute advisor-led conversation -> monitor follow-up -> conclude record |
| Underserved outcomes | 降低会前准备时间;提高客户限制、风险偏好、集中度、费用、流动性和替代方案覆盖度;降低材料过期或产品风险遗漏 |
| AI intervention | Client profile summarization、portfolio concentration flags、product literature retrieval、suitability checklist、disclosure draft、meeting note draft |
| Automation boundary | AI 不直接向客户给个性化买卖建议, 不承诺收益, 不替代持牌顾问判断, 不生成未经批准的营销材料, 不在缺少适当性信息时推荐产品 |
| Product architecture | Advisor desktop + CRM / portfolio system + risk profile + product master + approved literature RAG + suitability rules + supervision review + communication archive |
| Evidence | Prep time baseline、supervisor QA defects、missing suitability evidence、product literature version issues、client complaint themes |
| Selection decision | 适合顾问内部 copilot;客户可见输出必须经过批准模板、顾问审核和记录保存 |
Outcome examples:
| Outcome statement | Eval / metric |
|---|---|
| Minimize the effort required for an advisor to assemble current product, fee and risk evidence before a client meeting. | Prep time, source freshness, missing document defects |
| Increase the completeness of suitability considerations captured before discussing product options. | Checklist coverage, supervisor QA score, exception count |
| Reduce the risk of unapproved claims or outdated product information entering client communication. | Approved source citation rate, marketing compliance defects |
9.5 AI 平台内客户: Enterprise AI Platform Consumer
| 维度 | 设计 |
|---|---|
| Job executor | Business AI product team / domain PM / solution architect / engineer / risk reviewer |
| Core job | 在符合企业治理、数据、安全、模型风险和审计要求的前提下, 更快交付可复用、可评估、可监控的 AI capability |
| Job steps | Define use case -> locate approved model / data / pattern -> prepare integration -> confirm risk tier and eval -> execute pilot -> monitor usage / cost / quality -> modify prompt / model / knowledge -> conclude scale / stop decision |
| Underserved outcomes | 降低团队找到批准模型、RAG pattern、eval template 和合规门禁的时间;提高复用控制和证据完整度;降低重复建设和 shadow AI |
| AI intervention | Platform catalog assistant、pattern recommendation、eval template generation、architecture checklist, policy-as-code guidance, evidence binder automation |
| Automation boundary | 平台不替业务 owner 接受风险, 不自动批准高风险 use case, 不允许绕过数据权限和模型评审, 不把通用 eval 当成业务验收 |
| Product architecture | Developer portal + model gateway + RAG service + prompt / policy registry + eval harness + tool gateway + observability + cost controls + evidence binder + ARB workflow |
| Evidence | Use case onboarding time、重复组件建设、risk review cycle time、eval adoption、production incident and drift data、platform NPS from internal teams |
| Selection decision | 当多个业务 use case 反复需要相同控制时, 平台化价值高;第一天不应建设全量平台, 应从 repeated control points 开始 |
Outcome examples:
| Outcome statement | Eval / metric |
|---|---|
| Reduce the time required for a domain team to configure an approved model, retrieval source and eval gate for a new AI use case. | Onboarding lead time, configuration defect rate, eval completion rate |
| Increase the likelihood that every production AI use case has traceable requirements, model version, prompt version, data source and release approval. | Evidence completeness, audit sample pass rate, registry coverage |
| Reduce duplicate implementation of model access, retrieval, logging, eval and policy controls across teams. | Reuse rate, redundant service count, platform adoption |
10. Templates
10.1 AI JTBD Intake Canvas
| Field | Guidance |
|---|---|
| Job executor | 谁真正执行 job, 谁受影响, 谁购买/批准, 谁运营/支持 |
| Job | 不写产品或功能, 写要完成的稳定进展 |
| Context | 什么时候、在哪个渠道、什么触发条件下必须完成 |
| Current alternatives | 现在用系统、表格、人工、规则、外包、客户自助或 workaround 完成 |
| Job steps | Define / Locate / Prepare / Confirm / Execute / Monitor / Modify / Conclude |
| Desired outcomes | 每一步如何衡量更快、更准、更低风险、更少努力 |
| Evidence | baseline、日志、QA、投诉、审计、SME calibration、财务影响 |
| Underserved outcomes | 重要高、满足低、证据强的 outcome |
| AI intervention | read、retrieve、summarize、draft、recommend、monitor、act-with-approval |
| No-AI alternative | 流程、规则、系统集成、数据治理、培训、模板是否更合适 |
| Automation boundary | allowed / blocked decisions, actions, outputs, handoff |
| Architecture implication | data、RAG、rules、tool、workflow、eval、monitoring、audit |
| Portfolio decision | scale candidate、controlled pilot、measurement sprint、redesign、stop |
10.2 Job Map to AI Pattern Template
| Job step | Current struggle | Outcome statement | Evidence | AI pattern | Boundary | Architecture component |
|---|---|---|---|---|---|---|
| Define | ||||||
| Locate | ||||||
| Prepare | ||||||
| Confirm | ||||||
| Execute | ||||||
| Monitor | ||||||
| Modify | ||||||
| Conclude |
使用规则:
- 空白单元格在正式交付中必须填入具体内容或明确标记为 not applicable with rationale。
- Outcome statement 不能写成“提升体验”或“更智能”, 必须有 metric object。
- Boundary 必须包含 blocked output 或 blocked action。
- Architecture component 必须能追溯到至少一个真实 outcome 或 control need。
10.3 Outcome Statement Template
Outcome ID:
Job:
Job step:
Job executor:
Outcome statement:
Direction:
Metric object:
Context:
Current baseline:
Target movement:
Risk / compliance constraints:
Evidence source:
Potential AI intervention:
No-AI alternative:
Eval method:
Monitoring signal:
Owner:
10.4 Underserved Outcome Scorecard
| Outcome ID | Importance | Satisfaction gap | Evidence strength | Strategic / regulatory relevance | Reuse potential | Automation risk | Architecture gap | Adoption friction | Decision |
|---|---|---|---|---|---|---|---|---|---|
| O-001 | 5 | 4 | 5 | 5 | 4 | 3 | 2 | 2 | Controlled pilot |
| O-002 | 4 | 2 | 2 | 3 | 2 | 4 | 4 | 3 | Measurement sprint |
Decision vocabulary:
| Decision | 含义 |
|---|---|
| Controlled pilot | 有足够 evidence 和价值, 风险可通过边界和人审控制 |
| Measurement sprint | 价值假设合理但证据不足, 先建立 baseline 和样本 |
| Workflow / rules first | 问题主要来自流程、规则不清或系统断点, AI 不是第一解 |
| Platform candidate | 多个 use case 共享模型接入、RAG、eval、tool gateway 或证据组件 |
| Stop | 低重要、弱证据、高风险或已有满足度高 |
10.5 Automation Boundary Contract
| Section | 内容 |
|---|---|
| Use case name | 清晰命名, 不用“智能助手”泛称 |
| Job and outcome | 绑定到 job、job step 和 underserved outcome |
| AI role | retrieve / summarize / draft / recommend / monitor / act-with-approval |
| Allowed data | source、owner、permission、freshness、retention |
| Blocked data | sensitive attributes、unapproved notes、expired policy、unconsented data |
| Allowed outputs | 结构化摘要、引用、草稿、建议、风险提示、checklist |
| Blocked outputs | 正式信贷理由、个性化投资建议、法律结论、SAR decision、投诉拒绝、未经授权承诺 |
| Allowed tools | read-only API、case draft、task creation、approved workflow transition |
| Blocked tools | approve / decline、file / close、trade、fee waiver、customer lockout、delete evidence |
| Human review | reviewer、approver、override、dual control、SLA |
| Escalation | low confidence、missing source、policy conflict、high-risk intent、customer harm、tool error |
| Eval gate | dataset、rubric、threshold、critical failure |
| Monitoring gate | online signals、alert threshold、rollback、incident owner |
10.6 Evidence-to-Architecture Traceability Matrix
| Outcome | Evidence | Requirement | Architecture component | Eval case | Release gate | Monitoring signal |
|---|---|---|---|---|---|---|
| 降低 AML 证据查找时间 | Case telemetry | AI 必须聚合交易、实体、历史 case 并显示来源 | Entity graph + RAG + case UI | Golden alert set | Evidence completeness >= threshold | Missing citation / analyst override |
| 降低客服错误承诺 | QA defects | AI 草稿不得承诺未经授权的补偿 | Policy rules + approved templates | Red-team fee waiver set | Critical violation = 0 | QA critical defect |
| 提高申请完整性 | Funnel log | 提交前提示缺失材料 | Completeness checker + portal UI | Application test set | Missing doc detection >= threshold | Return-for-incomplete rate |
10.7 One-Page Use Case Selection Memo
Use case:
Job executor:
Job-to-be-done:
Top underserved outcomes:
Evidence summary:
AI intervention:
Automation boundary:
No-AI alternative considered:
Architecture pattern:
Required controls:
Pilot design:
Eval gate:
Monitoring gate:
Risk owner:
Business owner:
Decision:
Rationale:
11. Review Checklist
11.1 JTBD / ODI quality
| Check | Pass signal |
|---|---|
| Job is stable | 写的是要完成的进展, 不是产品、渠道、模型或功能 |
| Job executor is clear | 区分执行者、受影响客户、审批者、平台内客户和风险 owner |
| Job map is complete | 至少覆盖 define、locate、prepare、confirm、execute、monitor、modify、conclude |
| Outcomes are measurable | 每个 outcome 有 direction、metric object、context 和 constraint |
| Underserved is evidenced | 有 importance、satisfaction、baseline 或强专家证据 |
| No-AI alternative considered | 明确为什么不是只做流程、规则、培训或系统集成 |
11.2 AI suitability
| Check | Pass signal |
|---|---|
| AI role is specific | retrieve、summarize、draft、recommend、monitor、act-with-approval 中至少选定一种 |
| AI does not own forbidden decisions | 高影响金融决策仍由正式系统、人或授权流程负责 |
| Boundary is explicit | blocked data、blocked output、blocked tool、escalation trigger 都写清 |
| Human accountability works | reviewer 有证据、权限、时间、训练和覆盖路径 |
| Architecture supports evidence | RAG、rules、tool、eval、observability、audit 能支撑 outcome 和 boundary |
| Eval matches outcome | 不只测模型准确率, 还测业务 outcome、风险 failure 和人审质量 |
11.3 Portfolio quality
| Check | Pass signal |
|---|---|
| Selection is comparable | 每个 use case 用同一评分框架和决策词汇 |
| Scale rule exists | pilot 成功、受限扩展、停止、平台化的条件明确 |
| Reuse is real | 平台候选来自多个 use case 的重复控制点, 不是抽象愿望 |
| Risk is priced | 人审成本、合规成本、监控成本和 incident 成本进入价值判断 |
| Evidence can be audited | 决策、数据、模型、prompt、eval、审批和变更可追溯 |
12. 反模式
| Anti-pattern | 表现 | 修正方式 |
|---|---|---|
| AI feature hunting | 到处问“哪个流程能加 AI” | 从 job 和 outcome portfolio 开始 |
| Persona-only selection | 用“客服小王很忙”决定用例 | 改为 job executor + job step + measurable outcome |
| Pain point inflation | 把所有抱怨都当机会 | 用 importance / satisfaction / evidence 过滤 |
| Demo-driven roadmap | 供应商 demo 好看就排进 roadmap | 要求 evidence、boundary、architecture 和 eval gate |
| Chatbot as default UX | 所有场景都做聊天入口 | 按 job step 选择 search、form assist、copilot、workflow automation 或 agent |
| Automation-first | 先追求无人化 | 先定义 forbidden decisions 和 human accountability |
| Accuracy-only eval | 只看回答准确率 | 增加 outcome、critical failure、citation、human override、customer harm |
| Weak HITL | 写“人工审核”但不给证据和时间 | 设计 reviewer UI、action set、capacity、training 和 audit |
| No baseline | 没有当前耗时、缺陷、投诉或成本 | 先做 measurement sprint |
| Platform theater | 没有重复 use case 就建平台 | 从 repeated controls: model gateway、RAG、eval、tool gateway、evidence binder 开始 |
| Risk transfer to disclaimer | 用免责声明替代控制 | 把边界落实到权限、规则、UI、handoff 和 release gate |
| LLM reason pipeline | 用 LLM 生成正式拒贷原因 | 建正式 reason service, LLM 只解释流程或辅助员工理解 |
| Shadow AI normalization | 允许团队私接模型和数据 | 建 AI registry、gateway、policy-as-code 和 exception process |
13. 30 天训练计划
目标: 30 天内形成一套可展示的 AI JTBD / ODI portfolio pack, 聚焦金融零售高价值场景, 不做基础访谈训练。
| Day | 训练主题 | 产出 |
|---|---|---|
| 1 | 选定 5 个金融零售 AI 候选场景: AML、客服、信贷申请、财富顾问、AI 平台内客户 | Use case inventory |
| 2 | 为每个场景定义 job executor、affected stakeholder、business owner、risk owner | Stakeholder and job executor map |
| 3 | 把每个场景从 solution name 改写成 job-to-be-done | Job definition sheet |
| 4 | 为 AML 场景画 define / locate / prepare / confirm / execute / monitor / modify / conclude job map | AML job map |
| 5 | 为客服场景画 job map, 区分客户可见和 agent assist | Customer service job map |
| 6 | 为信贷申请人场景画 job map, 标出正式决策和原因链路边界 | Credit applicant job map |
| 7 | 为财富顾问和 AI 平台内客户画 job map | Advisor and platform job maps |
| 8 | 为每个 job step 写 3-5 条 outcome statement | Outcome catalog v1 |
| 9 | 清理弱 outcome, 替换“提升体验、提升效率、更准确”等模糊表达 | Outcome quality review |
| 10 | 收集每个 outcome 的 evidence: 日志、QA、投诉、审计、专家校准、成本、SLA | Evidence inventory |
| 11 | 为 outcome 打 importance / satisfaction / evidence strength 分 | ODI scorecard v1 |
| 12 | 加入 automation risk、architecture gap、adoption friction | Risk-adjusted scorecard |
| 13 | 选出每个场景前 2 个 underserved outcomes | Underserved outcome map |
| 14 | 为低证据但高潜力 outcome 设计 measurement sprint | Measurement plan |
| 15 | 将 underserved outcome 转成 AI intervention point | AI intervention map |
| 16 | 为每个 intervention 判断 no-AI alternative | No-AI alternative memo |
| 17 | 为 AML 和客服写 automation boundary contract | Boundary contracts v1 |
| 18 | 为信贷、财富和平台场景写 automation boundary contract | Boundary contracts v2 |
| 19 | 标出 blocked data、blocked output、blocked tool 和 escalation trigger | Boundary redline matrix |
| 20 | 把 outcome evidence 映射到架构组件 | Evidence-to-architecture matrix |
| 21 | 为每个场景画 C4 context / container sketch 或 sequence sketch | Architecture sketches |
| 22 | 设计 eval cases: golden set、rubric、threshold、critical failure | Eval case inventory |
| 23 | 设计 release gate: go / limited go / no-go / rollback | Release gate memo |
| 24 | 设计 monitoring gate: quality、risk、cost、latency、override、incident | Monitoring gate spec |
| 25 | 写 AML use case selection memo | AML selection memo |
| 26 | 写客服和信贷 use case selection memo | Customer service and credit memos |
| 27 | 写财富顾问和平台内客户 selection memo | Advisor and platform memos |
| 28 | 做 portfolio heatmap, 选择 scale、pilot、measurement、workflow-first、stop | Portfolio decision map |
| 29 | 准备 8 个高级面试答案, 每个用 30 秒和 2 分钟版本 | Interview answer pack |
| 30 | 整理作品集: canvas、scorecard、boundary、architecture、eval、memo、executive narrative | AI JTBD / ODI portfolio pack |
每周检查:
| Week | 成功标准 |
|---|---|
| Week 1 | 5 个场景都能从 solution name 改写成 job map |
| Week 2 | 每个场景至少有 8 条高质量 outcome statement 和证据来源 |
| Week 3 | 每个候选 use case 都有 AI intervention、no-AI alternative 和 automation boundary |
| Week 4 | 能用 portfolio heatmap 和 selection memo 向管理层解释为什么做、怎么做、做到哪里停 |
14. 面试答案
14.1 你如何选择一个 AI use case 是否值得做
30 秒版本:
我不会从模型能力或 demo 开始。我会先定义 job executor 和 job-to-be-done, 画 job map, 写 desired outcome statement, 用 importance、satisfaction 和证据找 underserved outcomes。然后判断 AI 适合介入哪一步, 定义 automation boundary, 映射到数据、RAG、rules、tool、eval、monitoring 和 audit 架构。最后用 risk-adjusted scorecard 决定 scale、pilot、measurement sprint、workflow-first 或 stop。
2 分钟版本:
| Step | 说明 |
|---|---|
| Job | 明确谁要完成什么稳定进展, 避免 solution-first |
| Outcome | 把痛点写成可度量 desired outcome, 如降低证据查找时间、提高引用准确性、降低错误承诺 |
| Opportunity | 用 importance / satisfaction / evidence strength 识别 underserved outcome |
| Boundary | 判断 AI 是检索、总结、草拟、建议还是执行, 明确 blocked decisions 和 human accountability |
| Architecture | 反推 source-of-truth、RAG、rules、workflow、tool gateway、eval、monitoring、audit |
| Decision | 用价值、风险、复用、架构可行性和 adoption 共同决定投资方式 |
14.2 JTBD 和传统用户旅程有什么区别
30 秒版本:
用户旅程描述用户今天怎么走流程, JTBD 描述用户真正要完成的稳定 job 和每一步想达到的 outcome。AI 项目里这个区别很重要, 因为如果只优化现有旅程, 很容易给旧流程加 chatbot;如果从 job 和 outcome 出发, 我能判断是 RAG、copilot、规则、流程重构还是不该用 AI。
高级追问:
| 追问 | 回答 |
|---|---|
| 旅程图还要不要 | 要, 但旅程图服务于 workflow integration;JTBD 决定机会是否真实 |
| persona 还要不要 | 要, 但 persona 不能替代 job executor、decision rights 和 risk owner |
| job map 是不是流程图 | 不是。流程图描述组织今天怎么处理, job map 描述理想完成 job 的步骤 |
14.3 什么是好的 outcome statement
30 秒版本:
好的 outcome statement 必须包含方向、度量对象、job step、情境和约束。例如“降低 AML analyst 查找交易证据和关联实体的时间, 同时不降低证据完整性”。它比“提升效率”强, 因为它能直接连接 baseline、eval、架构和上线门禁。
2 分钟版本:
| 要素 | 例子 |
|---|---|
| Direction | reduce、increase、minimize、improve |
| Metric object | time、effort、likelihood、completeness、risk、variance |
| Job step | locate evidence、confirm policy、prepare application |
| Context | investigating an alert、answering fee dispute、submitting application |
| Constraint | without unauthorized commitment、without missing critical evidence |
14.4 如何判断 underserved outcome
30 秒版本:
我看三个核心信号: 重要性高、当前满足度低、证据足够。AI 场景还要加风险和架构可行性。比如客服“错误承诺补偿”虽然频率可能不高, 但客户和合规影响高, 如果 QA 和投诉证据显示真实发生, 就是高优先级 outcome;相反, “回复语气更自然”如果没有客户结果证据, 不应排在前面。
14.5 金融零售 AI 的 automation boundary 怎么定
30 秒版本:
我先把 AI role 分成 retrieve、summarize、draft、recommend、act-with-approval 和 autonomous act。然后对每个 use case 写 blocked data、blocked output、blocked tool 和 escalation trigger。高影响金融决策, 比如拒贷、投资建议、SAR 判断、投诉关闭和 fee waiver, 不让 LLM 单独自动化, 只让 AI 提供证据、草稿、检查和建议, 最终责任在授权人或正式系统。
14.6 如何把 JTBD / ODI 转成产品架构
30 秒版本:
每个 underserved outcome 都会产生架构要求。证据找不到来源, 就需要 source registry、authorization-aware RAG 和 citation validator;错误承诺多, 就需要 policy rules 和 approved templates;工具动作有客户影响, 就需要 tool gateway、approval、rollback 和 audit log。架构不是先画系统图, 而是从 outcome 和 control 反推组件。
14.7 为什么不直接用 LLM 做信贷决策
30 秒版本:
信贷决策影响客户权益, 涉及公平信贷、可解释理由和正式通知。复杂模型也不能免除提供具体、准确原因的义务。LLM 可以帮助申请人理解材料要求、帮助 loan officer 起草 memo、检查资料完整性, 但正式 approve / decline 和 adverse action reason 应来自受控的决策、规则和 reason pipeline。
14.8 什么时候把 AI use case 平台化
30 秒版本:
我不会第一个 demo 就建大平台。平台化要来自重复证据: 多个 use case 都需要模型接入、RAG、权限、eval、tool gateway、observability、cost control 和 evidence binder。只有当这些 control points 重复出现, 并且平台能降低交付时间和风险 review 成本时, 才进入 platform roadmap。
15. 作品集交付物
一套完整的 AI JTBD / ODI portfolio pack 应包含:
| Deliverable | 内容 | 面试价值 |
|---|---|---|
| Source Anchor Memo | HBR JTBD、Strategyn ODI、NIST AI RMF、金融监管锚点如何转成方法 | 证明不是凭感觉设计 AI |
| AI JTBD Canvas | job executor、job、context、alternatives、job steps、outcomes、evidence | 证明能从业务问题而不是模型能力出发 |
| Job Map Pack | AML、客服、信贷申请人、财富顾问、AI 平台内客户五张 job map | 证明具备金融零售场景深度 |
| Outcome Catalog | 每个场景 8-15 条 desired outcome statement | 证明能写可度量需求 |
| Underserved Outcome Scorecard | importance、satisfaction、evidence、risk、architecture、adoption 评分 | 证明能做投资排序 |
| Automation Boundary Contract | allowed / blocked data、outputs、tools、decisions、handoff | 证明知道 AI 风险和控制边界 |
| Evidence-to-Architecture Matrix | outcome evidence 到架构组件、eval、gate、monitoring 的追踪 | 证明能连接 BA、PM 和架构 |
| Use Case Selection Memo | 每个场景一页 scale / pilot / measurement / stop 决策 | 证明能向管理层表达取舍 |
| Eval and Release Gate Pack | golden set、rubric、threshold、critical failure、go / no-go | 证明能把需求转成上线门禁 |
| Portfolio Heatmap | 价值、风险、证据、复用、可行性矩阵 | 证明能管理 AI portfolio, 不是单点项目 |
| Executive Narrative | 为什么选这些 use cases, 为什么现在做, 为什么这个边界 | 证明能进入 AI PM / Architect / Transformation Lead 级别讨论 |
作品集叙事:
I selected AI use cases through a JTBD / ODI lens:
first defining stable jobs and desired outcomes,
then identifying underserved outcomes with evidence,
then deciding the right AI intervention and automation boundary,
then translating the decision into architecture, eval, release gates and monitoring.
This prevents demo-driven AI adoption and creates a portfolio that is valuable, controlled, reusable and auditable.
16. 最终判断准则
一个 AI use case 值得进入 pilot, 至少应满足:
- Job 清楚: 不是功能名, 是稳定 job。
- Outcome 清楚: 不是“更智能”, 是可度量 desired outcome。
- Underserved 清楚: 重要、满足低、有证据。
- AI role 清楚: 介入哪个 job step, 做 retrieve、summarize、draft、recommend、monitor 或 act-with-approval。
- Boundary 清楚: blocked decision、blocked output、blocked tool 和 handoff 写明。
- Architecture 清楚: 数据、知识、规则、工具、workflow、eval、monitoring、audit 能支撑。
- Gate 清楚: pilot、scale、stop 和 rollback 条件可执行。
- Owner 清楚: business、risk、data、platform、model、operations 和 audit responsibility 不混淆。
一句话收束:
好的 AI use case selection 不是证明 AI 能做什么, 而是证明某个重要 job 的 underserved outcome 值得被改善, AI 是合适但受控的介入方式, 并且组织有架构、证据和责任体系把它安全地交付和运营。