AI Wardley Mapping Product Strategy Playbook
这些来源用于校准术语、地图方法和治理语言。本文不是法律、合规、采购、投资或监管意见;正式项目必须由 legal、compliance、risk、security、privacy、procurement、data owner、model risk、architecture board 和 business owner 审查。
Wardley Mapping for AI Product / Platform Strategy Playbook
适用对象:高级 AI PM、AI Platform PM、Product Architect、AI Architect、Enterprise Architect、金融零售 AI 转型负责人。 核心问题:如何用 Wardley Mapping 把 AI 产品机会、平台边界、build / buy / partner、竞争地形和架构投资路线放到同一张战略地图里,而不是用孤立 PRD、供应商清单或技术架构图做决策。 学习目标:训练战略态势感知能力。重点不是基础产品管理,而是从 user need、value chain、evolution axis、doctrine、climatic patterns 和 strategic gameplay 出发,设计 AI 平台与产品组合路线。 作品集定位:本文可直接转化为 AI strategy portfolio evidence,包括 AI Wardley Map、value chain template、evolution heatmap、build-buy-partner memo、platform boundary ADR、competitive landscape map、architecture investment thesis 和金融零售案例包。
Source Anchors
这些来源用于校准术语、地图方法和治理语言。本文不是法律、合规、采购、投资或监管意见;正式项目必须由 legal、compliance、risk、security、privacy、procurement、data owner、model risk、architecture board 和 business owner 审查。
| Anchor | Official / primary source | 在本 playbook 中的用法 |
|---|---|---|
| Simon Wardley - Writings / Books | https://www.swardleymaps.com/writings-books | 作为 Wardley Mapping 原始材料入口,用于 user need、value chain、evolution、doctrine、climatic pattern 和 gameplay 的方法锚点 |
| Wardley Mapping book by Simon Wardley | https://learnwardleymapping.com/book/ | 作为 Wardley Mapping 书籍章节入口,用于训练从地图看战略、组织、机会、浪费和竞争态势 |
| Wardley Maps community resources | https://list.wardleymaps.com/ | 用于工具、社区资料、案例和地图实践参考 |
| NIST AI RMF | https://www.nist.gov/itl/ai-risk-management-framework | 用 Govern、Map、Measure、Manage 组织 AI risk、eval、monitoring、control、portfolio governance 和 evidence |
| Team Topologies | https://teamtopologies.com/ | 用 stream-aligned team、platform team、enabling team、complicated subsystem team 和 interaction modes 设计 AI 平台组织边界 |
| Team Topologies - Key Concepts | https://teamtopologies.com/key-concepts | 用 four team topologies 和 collaboration / X-as-a-Service / facilitation 解释平台能力如何进入业务流 |
| Team Topologies - Platform Manifesto | https://teamtopologies.com/platform-manifesto | 用 internal customer、self-service pattern、adoption、engagement 和 platform as product 约束 AI 平台建设 |
Source anchors 到作品集证据的映射:
| Source lens | 可以产出的 evidence | 面试表达 |
|---|---|---|
| Wardley Mapping | AI Wardley Map、evolution heatmap、strategic playbook、investment thesis | “我不是按 vendor hype 排路线,而是先把 user need、value chain 和组件演进位置画出来,再决定哪里 build、哪里 buy、哪里平台化。” |
| NIST AI RMF | AI risk tier、control pack、eval gate、monitoring plan、evidence binder | “地图告诉我哪里投资,AI RMF 告诉我哪些风险必须被治理、评估和持续管理。” |
| Team Topologies | platform boundary、team interaction model、cognitive load assessment、X-as-a-Service contract | “平台不是共享代码库,而是让 stream-aligned teams 更快、更安全交付 AI 产品的内部产品。” |
One-Sentence Positioning
Wardley Mapping for AI Strategy = 用 user need 锚定价值链,用 evolution axis 判断组件成熟度,用 doctrine 建立基本作战纪律,用 climatic patterns 识别不可逆趋势,用 strategic gameplay 决定 AI 产品、平台、供应商和架构投资的取舍。
1. 为什么 AI PM / 架构师需要战略地图
AI 产品和平台最容易出现的失败不是“模型不够强”,而是团队没有态势感知:
- 把所有 AI 机会都当成同一种 chatbot。
- 把供应商能力表当成战略。
- 把 foundation model 选择当成护城河。
- 把第一个成功 pilot 扩成平台。
- 把所有平台能力都集中建设,导致业务团队等待和绕行。
- 把高风险 decisioning 和低风险 knowledge retrieval 放进同一套上线门禁。
- 把 eval 当测试活动,而不是产品契约和治理控制面。
Wardley Map 的价值在于把讨论从“我们喜欢什么方案”拉回“用户需要什么、价值链依赖什么、每个组件处于什么演进阶段、哪些变化由市场驱动、哪些投资能形成真实优势”。
1.1 AI 战略地图回答的高级问题
| 决策问题 | 没有地图时的典型错误 | 用 Wardley Map 后的判断方式 |
|---|---|---|
| 哪些 AI use case 值得做 | 看高层偏好、供应商 demo 或团队热情 | 从 user need 和 value chain 识别高可见、高痛点、高杠杆节点 |
| 哪里 build | 看到关键能力就自研 | 只在差异化、控制权、规模经济或监管责任明确的组件 build |
| 哪里 buy | 看到成熟 SaaS 就采购 | 对 commodity / product 化组件优先 buy,同时保留出口、审计和数据边界 |
| 哪里 partner | 复杂就外包 | 在 domain know-how、实施加速、监管解释或组织能力迁移上 partner,避免核心判断外包 |
| 何时平台化 | 第一个场景就建平台 | 当多个 use case 共享低可见但高复用组件,并且 Team Topologies 的 platform-as-product 条件成立时平台化 |
| 如何投资架构 | 按技术潮流投资 | 围绕 value chain bottleneck、risk control point、evolution movement 和 competitive gap 投资 |
| 如何看竞争 | 列 competitor feature | 比较竞争者在哪些 value chain 节点拥有速度、数据、分发、合规、生态或成本优势 |
1.2 地图与传统 artifacts 的关系
| Artifact | 适合回答 | 不适合回答 | 与 Wardley Map 的关系 |
|---|---|---|---|
| PRD | 用户流程、范围、验收 | 为什么这个产品路线比另一个更值得投资 | 地图先定义战略位置,PRD 再定义可交付产品 |
| C4 架构图 | 系统边界、组件、部署 | 哪些组件应该商品化、外购或形成护城河 | 地图决定架构投资方向,C4 展示实现结构 |
| Capability map | 企业能力覆盖 | 能力组件的市场演进和竞争动作 | capability map 可作为 value chain 的上层输入 |
| Vendor scorecard | 供应商能力比较 | 供应商能力是否战略上重要 | 地图先定义需要买什么,scorecard 再比较谁更合适 |
| Risk register | 风险和控制 | 哪些风险来自演进趋势和平台边界 | 地图把风险放回价值链和组织边界 |
| Roadmap | 时间安排 | 为什么这个顺序正确 | 地图用 evolution、dependency 和 gameplay 解释 roadmap 顺序 |
2. Wardley Mapping 核心语言转成 AI 战略语言
| Wardley concept | 原始含义 | AI 产品 / 平台转译 | 产出 |
|---|---|---|---|
| User need | 地图锚点,用户真正要完成的事 | 客户、员工、风控、合规、开发者或监管审查者的高价值需求 | User need statement |
| Value chain | 满足用户需求所需的组件和依赖 | 体验、流程、决策、知识、数据、模型、工具、评估、治理、基础设施 | AI value chain |
| Visibility | 对用户价值的可见程度 | 越靠上越接近业务体验和决策影响,越靠下越接近平台和基础设施 | Map y-axis |
| Evolution axis | Genesis -> Custom -> Product -> Commodity | 从新颖实验到定制建设、产品化服务、标准化基础设施 | Map x-axis |
| Doctrine | 普遍适用的基本原则 | 先有 user need、owner、数据边界、风险分层、eval、可逆性和反馈机制 | AI doctrine checklist |
| Climatic patterns | 外部不可逆趋势 | 模型能力商品化、agent 风险上升、eval 标准化、监管增强、成本下降但用量上升 | Trend assumptions |
| Strategic gameplay | 基于地图的行动 | build、buy、partner、平台化、开源、生态、标准、差异化、快速跟随、退出 | Strategy moves |
3. AI Value Chain 模板
AI value chain 不从“模型”开始,而从 user need 开始。模型通常在价值链中段或底层;真正决定成败的是业务流程、数据治理、控制点和用户信任。
3.1 通用模板
User / actor
-> user need
-> experience / workflow surface
-> decision or task boundary
-> domain policy / business rule
-> AI interaction pattern
-> orchestration / workflow engine
-> tool access / system integration
-> knowledge / retrieval / memory
-> data products / metadata / lineage
-> model access / model gateway
-> eval / monitoring / incident loop
-> identity / entitlement / audit
-> infrastructure / cloud / runtime
3.2 AI value chain 表格模板
| Value chain node | User-visible? | Strategic question | Typical AI components | Evidence required |
|---|---|---|---|---|
| User need | Very high | 谁的哪个高价值问题被解决 | Persona、workflow pain、outcome metric | 用户证据、业务基线、风险影响 |
| Experience / workflow | High | AI 在流程哪里出现,改变什么行为 | Copilot UI、case workbench、advisor desktop、agent console | TO-BE workflow、adoption signal、human role |
| Decision boundary | High | AI 是 inform、draft、recommend、approve 还是 act | HITL、approval queue、policy engine | Decision rights、risk tier、exception path |
| Domain policy | Medium-high | 哪些政策和业务规则约束输出 | Policy RAG、rule engine、reason codes | Policy owner、effective date、exception rule |
| AI pattern | Medium | RAG、agent、classification、extraction、forecasting、decisioning 如何组合 | RAG、tool agent、classifier、scoring model | Pattern ADR、failure taxonomy |
| Orchestration | Medium | 多步骤流程如何稳定、可回滚、可审计 | Workflow engine、state machine、event bus | State model、retry、compensation、SLA |
| Tool / integration | Medium-low | AI 可读写哪些系统,权限如何控制 | Tool gateway、API adapter、CRM/core/case system connector | AuthZ、idempotency、tool audit |
| Knowledge / retrieval | Medium-low | 知识从哪里来,如何证明有效 | Ingestion、chunking、metadata、reranker、citation | Source owner、lineage、freshness、retrieval eval |
| Data product | Low-medium | 特征、标签、历史案例是否可复用 | Feature store、golden dataset、entity graph | Data contract、quality SLA、PII control |
| Model access | Low | 如何选择、路由、替换和降级模型 | Model gateway、routing、fallback、quota | Provider policy、latency、cost、version |
| EvalOps | Low but critical | 如何证明质量、风险和回归 | Dataset registry、eval runner、judge、release gate | Eval report、critical failure、slice analysis |
| Security / audit | Low but mandatory | 如何防泄露、越权、滥用和审计失败 | IAM、policy-as-code、audit log、SIEM export | Control pack、evidence binder |
| Infrastructure | Low | 运行成本、性能、可靠性如何管理 | Cloud、container、queue、vector DB、observability | SLO、DR、cost model、capacity plan |
3.3 AI value chain 使用规则
| Rule | 说明 |
|---|---|
| 从上往下拆 | 先写 user need,再写 AI pattern 和技术组件 |
| 每个节点必须有 owner | 没有 owner 的知识、数据、eval、prompt、工具权限会变成生产风险 |
| 越靠上越关注差异化 | 用户体验、决策边界、领域政策和工作流往往是差异化来源 |
| 越靠下越关注复用和标准化 | model gateway、EvalOps、audit、observability 更适合平台化 |
| 高风险节点不能只靠供应商承诺 | AML、信贷、财富建议、客户可见输出必须有内部控制和证据 |
| 地图要反映依赖 | 如果 value chain 中某个低层组件不成熟,上层产品路线必须降速或改变路径 |
4. Evolution Mapping:把 AI 组件放到演进轴
演进轴不是技术成熟度评分,而是市场与组织对某个组件的共同认知:它是否稀缺、不可预测、需要定制,还是已经产品化、标准化、可替换。
4.1 演进阶段定义
| Stage | 特征 | AI 场景信号 | 管理方式 |
|---|---|---|---|
| Genesis | 新颖、不确定、少数团队探索 | 新 agent pattern、新交互范式、新监管解释、新数据产品假设 | 小实验、强学习、少承诺、快速关闭 |
| Custom-built | 价值明确但高度定制 | 领域 workflow、专有 policy、复杂 entitlement、业务特有 eval | 内部设计、架构 ADR、领域团队深度参与 |
| Product / Rental | 多家供应商可提供,配置仍重要 | RAG 平台、LLM app framework、eval SaaS、workflow automation、AI客服套件 | buy / partner / hybrid,重视集成和控制 |
| Commodity / Utility | 标准化、低差异、高可替换 | 云算力、基础模型 API、日志管道、身份认证、对象存储、基础向量检索 | 优先消费成熟服务,控制成本、SLO 和 exit path |
4.2 AI 组件演进热力图
| Component | Current evolution | 可能变化 | 战略含义 |
|---|---|---|---|
| Foundation model API | Product -> Commodity | 主流模型能力继续趋同,价格和上下文窗口竞争加剧 | 不把单一模型当护城河,投资 model gateway 和评估能力 |
| Domain-specific small model | Custom -> Product | 金融、客服、合规、文档理解模型逐渐产品化 | 对高价值领域可 hybrid,保留训练数据和验证能力 |
| Prompt engineering practice | Custom -> Product | 模板、registry、policy、testing 工具成熟 | 不靠个人 prompt 技巧,转向 prompt/config lifecycle |
| RAG ingestion pipeline | Custom -> Product | 连接器、解析、chunking、rerank 产品化 | buy 基础能力,build 权限、元数据、source governance |
| Knowledge governance | Genesis -> Custom | 企业内部知识 owner、有效期、权限、引用标准仍不成熟 | 这是高杠杆架构投资,不是简单工具采购 |
| EvalOps | Genesis -> Product | eval platform、judge、dataset registry 正在标准化 | 早建内部 eval contract 和 golden set,工具可买 |
| AI observability | Product | tracing、token、latency、feedback 工具成熟 | 平台统一 telemetry,接入 risk 和 incident loop |
| Tool gateway for agents | Genesis -> Custom | 权限、审批、idempotency、audit 模式还在形成 | 对金融动作必须内部控制,不能完全托管给 agent vendor |
| Policy-as-code for AI | Genesis -> Custom | 合规规则、输出约束、审批策略逐渐工程化 | 建立 policy owner 和 machine-enforceable controls |
| AI gateway | Custom -> Product | 多模型路由、审计、成本、guardrail 产品化 | 关键控制面可 hybrid:买底层,保留内部 policy abstraction |
| Customer-facing AI safety | Custom -> Product | guardrails、red-team、content safety 逐渐成熟 | 金融零售仍需本地化合规话术、适当性和升级机制 |
| Wealth suitability logic | Custom | 监管、产品、客户画像和建议责任高度本地化 | 不外包最终 suitability boundary,AI 只做受控辅助 |
| Credit decisioning policy | Custom | 部分自动化平台成熟,但风险偏好和解释责任本地化 | build / hybrid policy layer,供应商可提供模型或平台组件 |
4.3 演进位置到动作的转换
| Map position | 推荐动作 | 例子 | 风险 |
|---|---|---|---|
| 高可见 + Genesis | 探索、学习、差异化试验 | AI-native advisor experience、agentic AML investigation | 过早规模化,合规边界不清 |
| 高可见 + Custom | 核心产品投资 | 信贷审批工作台、财富顾问 copilot、case narrative workbench | 低估组织 adoption 和流程改造 |
| 高可见 + Product | 选择性购买并深度配置 | 客服 AI 套件、contact center AI | 被 vendor UI 限制体验和数据 |
| 低可见 + Genesis | 小团队研究,不进入主路线 | 新 agent memory、自治工具策略 | 技术驱动,缺业务证据 |
| 低可见 + Custom | 可能是平台投资窗口 | EvalOps contract、tool gateway、knowledge governance | 平台团队孤岛化 |
| 低可见 + Product | buy / hybrid | model gateway、RAG platform、observability | 锁定、成本、审计缺口 |
| 低可见 + Commodity | 直接消费,避免自研 | 云基础设施、基础日志、对象存储 | 忽略成本治理和可用性 |
5. AI Doctrine:地图之前先建立作战纪律
Doctrine 是不依赖具体场景、相对普遍适用的基本原则。AI 领域的 doctrine 不是口号,而是每张地图都要默认检查的前置纪律。
5.1 AI strategy doctrine
| Doctrine | 操作化解释 | 证据 |
|---|---|---|
| Know your users | 明确 customer、frontline、risk reviewer、developer、data owner、auditor 的不同需求 | Persona / actor map、workflow observation、adoption evidence |
| Focus on user need | 不以模型能力或供应商 demo 定义路线 | User need statement、business baseline |
| Understand the value chain | 把体验、流程、数据、模型、eval、治理和组织依赖画出来 | AI value chain map |
| Use appropriate methods | Genesis 用探索,custom 用工程和治理,commodity 用采购和自动化 | Evolution heatmap、sourcing memo |
| Challenge assumptions | 地图节点、位置、依赖和趋势必须能被不同角色挑战 | Assumption log、map review notes |
| Design for reversibility | 模型、供应商、prompt、index、policy、workflow 都要有 rollback / exit | Reversal trigger、exit plan、fallback design |
| Make risk measurable | 关键风险必须转成 eval、monitoring、control 和 release gate | NIST AI RMF aligned control pack |
| Keep platforms as products | 平台必须服务 internal customer adoption,不是架构团队的标准化欲望 | Platform PRD、team interaction contract、adoption metrics |
| Reduce cognitive load | 业务团队不应理解每个模型和 eval 细节才能安全交付 | Golden path、self-service API、enablement material |
| Preserve decision rights | 高风险金融决策必须明确人、规则、模型和系统的责任边界 | Decision boundary map、RACI、audit trail |
5.2 AI doctrine review questions
| Question | Good answer signal | Weak answer signal |
|---|---|---|
| User need 是否足够具体 | 能说出角色、任务、风险、频率、价值和失败成本 | “提升效率”“用 AI 降本” |
| Value chain 是否完整 | 包含知识、数据、权限、eval、audit、incident | 只有 UI 和模型 |
| Evolution 位置是否可辩护 | 每个关键组件都有市场证据、内部能力证据、供应商证据 | 凭团队直觉打点 |
| Build / buy 是否由地图推导 | 决策与组件可见性、演进位置、风险和差异化一致 | 先有偏好再补理由 |
| 平台边界是否清楚 | 平台提供 X-as-a-Service,业务保留 domain workflow | 所有能力都塞进平台 |
| 风险是否进入路线 | 高风险节点有 eval gate、HITL、monitoring、kill switch | 风险放到上线前一次性评审 |
6. Climatic Patterns:AI 领域的不可逆趋势
Climatic patterns 是不受单个企业意愿控制、但会改变战略地形的外部力量。AI PM / 架构师要把这些趋势写进地图假设,而不是等趋势发生后被动重排 roadmap。
6.1 AI climatic patterns
| Pattern | AI 领域表现 | 战略影响 |
|---|---|---|
| Everything evolves | 模型、RAG、agent、eval、gateway 都会从新颖走向产品化和商品化 | 不把今天的定制实现当长期架构,路线要预留替换 |
| Commoditization accelerates adoption | 模型 API、托管向量检索、AI observability 降低门槛 | 竞争差异转向 workflow、data、trust、distribution、governance |
| Higher-order systems emerge | 当模型和工具调用商品化,会出现更复杂 agentic workflow 和 AI operating model | 提前投资 tool gateway、state management、eval replay 和 incident response |
| Inertia is predictable | 老系统、数据 owner、合规流程、采购合同会抵抗变化 | 地图中要标出 inertia nodes,安排 migration 和 change strategy |
| Co-evolution of practice | 技术成熟会推动组织、流程和监管实践演进 | EvalOps、AI risk review、model registry 会从“创新团队做法”变成标准生产纪律 |
| Efficiency creates more demand | 模型价格下降不等于总成本下降,使用量会增加 | 必须投资 cost governance、quota、routing、caching 和 value attribution |
| Data gravity remains | 模型可换,但受控数据、标签、知识、事件和反馈难迁移 | 数据产品、knowledge governance 和 eval dataset 是长期资产 |
| Regulation follows harm | 客户可见、信贷、财富、AML、支付等高影响场景监管压力更高 | 低风险 copilot 与高风险 decisioning 不能共享同一控制强度 |
| Ecosystems beat isolated tools | 单点 AI 工具容易被平台、工作流系统或 hyperscaler 吸收 | 需要生态策略、API strategy、exit path 和 partner model |
6.2 趋势到投资主题
| Climatic pattern | Architecture investment | Product strategy |
|---|---|---|
| Foundation models commoditize | Model gateway、model evaluation、routing、fallback | 多模型策略,不卖“某模型能力”,卖业务 outcome |
| RAG tooling productizes | Knowledge governance、metadata、authorization-before-retrieval | 用可信知识体验做差异化 |
| Agentic workflows rise | Tool gateway、workflow state、approval queue、idempotency、kill switch | 从 read-only copilot 逐步走向 bounded action |
| Eval becomes standard | Dataset registry、release gate、production sampling、evidence binder | 把 eval 报告作为产品发布凭证 |
| AI cost becomes board-level | FinOps、quota、cost per successful task、budget guardrails | 用 unit economics 管 portfolio,不按 token 消耗讲价值 |
| Regulatory scrutiny grows | Risk tiering、audit trail、control pack、human oversight | 高风险路线先设计控制,再扩展功能 |
7. Strategic Gameplay:从地图到行动
Strategic gameplay 不是抽象战略词汇,而是对地图上具体节点采取行动。AI 领域尤其需要避免“全都做”“全都买”“全都平台化”的懒惰决策。
7.1 常用 AI strategic plays
| Play | 何时使用 | AI 示例 | 成功条件 |
|---|---|---|---|
| Exploit commodity | 组件已商品化,差异化低 | 使用云模型 API、托管日志、对象存储 | 严控成本、SLO、数据边界和 exit |
| Build where differentiated | 高可见、强领域差异、强控制需求 | 财富顾问 suitability workflow、信贷 exception reason engine | 有 domain owner、数据资产、eval 和长期维护能力 |
| Buy mature components | Product / commodity 组件成熟 | Contact center AI、OCR、managed vector DB、eval dashboard | 合同可审计、集成可控、数据可导出 |
| Partner for capability transfer | 需要领域实施或监管解释,但不能外包核心能力 | AML taxonomy setup、credit policy mapping、knowledge migration | 明确 knowledge transfer、internal owner 和退出条件 |
| Platform the invisible repeated capabilities | 多业务重复且低可见组件 | model gateway、EvalOps、tool gateway、audit、cost dashboard | 平台团队以 internal customers 和 adoption 为指标 |
| Differentiate through workflow | 底层模型相似,流程体验可形成优势 | AML investigator workbench、advisor desktop、客服坐席 copilot | 深入真实流程,减少切换成本和认知负荷 |
| Use ecosystem leverage | 市场产品化快,自己无法覆盖 | 与 hyperscaler、SI、domain SaaS、data vendor 形成组合 | 保留架构控制面和数据资产 |
| Attack inertia | 竞争者被遗留系统或采购锁定 | 用 thin layer、API facade、shadow mode、parallel run 快速落地 | 避免制造新一代不可替换 legacy |
| Open standards / portability | 供应商锁定风险高 | 日志 schema、eval dataset、policy config、prompt registry 可导出 | 把 portability 写进架构和合同 |
| Fast follower | 技术演进快但差异化不高 | 不自研通用 agent framework,等待成熟产品 | 保留实验窗口和替换能力 |
7.2 AI 竞争地形地图
竞争不是“谁的模型分数高”,而是谁控制了 value chain 中最有杠杆的节点。
| Competitor type | 控制的节点 | 优势 | 弱点 | 我方打法 |
|---|---|---|---|---|
| Hyperscaler / cloud AI platform | Model access、infra、security、data platform、observability | 集成深、规模大、采购便利 | 领域工作流浅,容易锁定 | 利用基础能力,保留 domain workflow、eval contract 和 policy layer |
| Foundation model provider | Model quality、API、tool ecosystem | 能力迭代快,开发者心智强 | 企业治理、workflow、行业责任有限 | 多模型路由,避免单模型依赖 |
| Enterprise SaaS vendor | CRM、contact center、case management、workflow | 已在业务流程内,有分发 | 跨系统平台能力弱,定制受限 | 让其服务特定流程,不让其吞掉全局 AI 控制面 |
| Domain AI vendor | AML、KYC、credit、wealth 等行业流程 | 领域模板和实施经验 | 数据适配、审计、模型透明度不一 | 做 vendor diligence,内部掌握 policy、eval、audit 和 exit |
| SI / consulting partner | 交付能力、流程梳理、变更管理 | 能快速铺项目 | 容易制造定制债和知识外流 | partner for acceleration,不外包核心架构判断 |
| Open-source ecosystem | Framework、模型、eval、observability 工具 | 灵活、透明、成本可控 | 运维和安全责任在内部 | 用于可控组件和防锁定,生产前补治理 |
| Internal platform team | 组织上下文、权限、审计、风险责任 | 可对齐企业控制和复用 | 可能脱离业务、过度抽象 | 用 Team Topologies 的 platform-as-product 约束 |
7.3 Strategic gameplay selection matrix
| Map observation | Recommended gameplay | 例子 |
|---|---|---|
| 高价值节点在 Genesis,竞争者也不成熟 | 探索 + 小规模差异化 | AI-native relationship manager assistant |
| 高价值节点在 Custom,我方有领域数据 | Build / hybrid | 信贷 policy exception reasoning、AML entity graph |
| 低可见节点在 Product,多个团队重复建设 | Platform + buy / hybrid | EvalOps、model gateway、RAG ingestion |
| 低可见节点在 Commodity | Consume + govern | Cloud runtime、basic logging、storage |
| Vendor 控制高风险决策边界 | Rebalance boundary | 把决策、政策和审计层拉回内部 |
| 业务团队绕过平台 | Improve platform UX | 增加 golden path、SDK、self-service、support model |
| 地图显示多个 use case 共享同一知识治理瓶颈 | Invest in architecture runway | Knowledge source registry、metadata、permission model |
| 成本节点快速右移但用量上升 | FinOps gameplay | quota、routing、cache、unit economics |
8. Build / Buy / Partner Decision
Build / buy / partner 不是采购问题,而是地形问题。先看组件在地图上的位置,再看差异化、风险、控制权、组织能力和经济性。
8.1 决策原则
| Option | Best when | Avoid when | Evidence |
|---|---|---|---|
| Build | 组件靠近用户价值、高差异、强监管责任、需要内部控制、长期规模经济成立 | 市场已有成熟产品且差异化低;内部没有运行能力 | Capability thesis、TCO、architecture ADR、operating model |
| Buy | 组件成熟、标准化、非差异化、time-to-value 关键、供应商有审计证据 | 高风险决策边界被供应商黑箱控制;数据和日志不可导出 | Vendor scorecard、security review、exit plan、contract terms |
| Partner | 需要领域知识、迁移能力、实施速度或 change management;内部必须学习接管 | Partner 变成永久外包核心能力 | Knowledge transfer plan、RACI、handover criteria |
| Hybrid | 底层或工具成熟,但 policy、eval、audit、domain workflow 必须内部控制 | 边界复杂到团队无法运营 | Boundary ADR、integration contract、support model |
8.2 AI build-buy-partner matrix
| Component | Default stance | Why | Boundary guardrail |
|---|---|---|---|
| Foundation model | Buy / multi-provider | 快速商品化,单模型不是长期护城河 | 通过 model gateway 管理版本、日志、成本和 fallback |
| Model gateway | Hybrid | 可买底层能力,但企业 policy、routing、audit、cost attribution 是控制面 | 内部保留 policy abstraction 和 logs export |
| RAG platform | Hybrid | parsing、index、retrieval 工具成熟;知识治理差异化 | 内部掌握 source registry、permission、freshness、eval |
| Domain ontology | Build / partner | AML、信贷、财富和客服术语高度本地化 | Partner 可加速,但 owner 必须在内部 |
| EvalOps tooling | Buy / hybrid | 工具产品化,但 eval contract 和 golden set 是核心资产 | 数据集、rubric、release thresholds 内部控制 |
| Tool gateway | Build / hybrid | 金融动作权限、审批、幂等和审计责任高 | 不允许 vendor agent 直接绕过企业授权 |
| Customer service copilot UI | Buy / configure | contact center AI 已产品化,价值在 adoption 和知识治理 | 客户可见回答规则、升级路径、日志内部可审计 |
| AML investigation workbench | Hybrid | 领域流程、证据链、SAR 边界强差异 | 可买 extraction / summarization,内部控制 case narrative 和 approval |
| Credit decisioning | Build / hybrid | 风险偏好、政策例外、解释和模型风险责任本地化 | AI 不应黑箱替代审批责任 |
| Wealth advisor assistant | Hybrid | suitability、建议边界、客户信任高风险 | 供应商可供内容和工具,内部保留建议策略和合规审核 |
8.3 Decision memo 模板
# AI Build / Buy / Partner Decision Memo
## 1. Decision
- Use case: AML investigation copilot
- Component: Evidence retrieval and SAR narrative drafting
- Recommendation: Hybrid
- Decision owner: Head of Financial Crime Product
- Review date: 2026-09-30
## 2. Map Context
- User need: Investigator needs faster evidence gathering and regulator-ready narrative draft without losing human accountability.
- Value chain position: High-visible workflow and narrative layer depend on lower-visible retrieval, entity graph, LLM summarization, EvalOps and audit.
- Evolution stage: Narrative boundary and red flags are custom; LLM summarization is product / commodity; retrieval tooling is product; AML eval dataset is custom.
- Dependencies: Case management, transaction history, entity graph, sanctions / watchlist data, policy documents, model gateway, evidence binder.
- Competitor / vendor landscape: Domain AML vendors offer workflow accelerators; hyperscalers offer model and search infrastructure; internal risk team controls policy and SAR accountability.
## 3. Strategic Rationale
- Differentiation: Case evidence quality, investigator workflow and SAR narrative evidence chain.
- Control requirement: SAR decision, red flag taxonomy, audit schema and release thresholds remain internally controlled.
- Risk tier: High, because wrong recommendations can affect regulatory reporting and investigation quality.
- Time-to-value: Buy mature extraction / summarization components to shorten pilot, build internal evaluation and approval boundary.
- Scale economics: Shared retrieval, model gateway and EvalOps can support KYC, fraud and compliance operations.
- Reversibility: Model and retrieval provider can be replaced if eval regression, audit gaps or cost thresholds are breached.
## 4. Options Considered
| Option | Benefits | Costs | Risks | Reversal trigger |
|---|---|---|---|---|
| Build | Maximum control over workflow, policy and evidence chain | Slower delivery and higher operating burden | Reinventing mature LLM / search capabilities | If delivery misses pilot gate by two review cycles |
| Buy | Faster pilot and vendor domain accelerators | Contract, integration and data governance effort | Vendor controls too much of SAR boundary | If audit logs, eval export or policy control are insufficient |
| Partner | Speeds taxonomy, workflow migration and change management | Requires strong knowledge transfer management | Long-term dependency on external experts | If internal owners cannot operate without partner after agreed handover |
| Hybrid | Uses mature components while retaining decision boundary and evidence control | Boundary design and integration complexity | Ownership confusion between vendor, platform and AML team | If support model or control ownership remains unclear after pilot |
## 5. Required Controls
- Data boundary: PII and investigation data only through approved gateway, no provider training use, retention aligned to compliance policy.
- Eval gate: Critical failures for fabricated evidence, stale policy citation, missing adverse signal and unauthorized disclosure must be zero.
- Audit evidence: Store input metadata, retrieved sources, model version, prompt version, narrative draft, reviewer approval and final action.
- Human oversight: Investigator remains accountable for SAR narrative and escalation decision.
- Incident response: Wrong evidence, leakage, unsafe recommendation, tool failure and cost runaway enter AI incident taxonomy.
- Exit path: Export prompts, eval datasets, logs, retrieval index metadata and case evidence schema.
## 6. Decision
- Selected option: Hybrid.
- Why now: LLM summarization and retrieval tooling are mature enough, while AML evidence and audit controls remain strategic.
- What would make this decision wrong: Vendor cannot expose required logs, internal team cannot maintain eval dataset, or investigator adoption remains below agreed threshold.
- Next governance checkpoint: Pilot gate review after 200 sampled cases and model risk signoff.
9. Platform vs Product Boundary
AI 平台边界的核心不是“共享能力越多越好”,而是让 stream-aligned product teams 更快、更安全、更低认知负荷地交付 AI 产品。
9.1 平台应该承担的能力
| Platform capability | 为什么平台化 | Internal customer |
|---|---|---|
| AI use case intake + risk tiering | 统一进入 portfolio,避免暗箱 pilot | AI PM、business owner、risk reviewer |
| Model gateway | 模型路由、日志、成本、fallback、allowlist 一致 | AI app team、security、FinOps |
| Prompt / config registry | 版本、审批、回滚、环境隔离 | Product team、risk、ops |
| Knowledge ingestion baseline | 连接器、metadata、权限、freshness 基线 | Knowledge owner、domain team |
| Retrieval evaluation starter kit | 统一衡量 citation、grounding、stale source | RAG product team |
| Tool gateway | 工具权限、审批、审计、幂等、kill switch | Agent product team、security |
| EvalOps platform | dataset、runner、judge、release gate、evidence | Product, architecture, model risk |
| AI observability | trace、latency、cost、quality、feedback、incident | SRE、ops、risk |
| Policy guardrail service | 统一底线:PII、越权、禁止建议、升级规则 | Legal、compliance、product |
| Cost governance | quota、budget、cost per successful task | Sponsor、FinOps、platform PM |
9.2 产品团队应该保留的能力
| Product / domain capability | 为什么不应平台吞掉 |
|---|---|
| User need and workflow ownership | 业务流程、采用方式和价值指标必须由 stream-aligned team 负责 |
| Domain policy interpretation | AML、信贷、财富、客服政策不能由平台团队代替业务责任人解释 |
| Product experience | 不同角色需要不同 UX、信任信号、升级路径和反馈机制 |
| Domain eval cases | golden set、failure taxonomy、rubric 需要领域专家维护 |
| Business outcome accountability | 平台负责加速和控制,产品负责业务结果 |
| Adoption and change management | 用户行为改变必须在业务团队里发生 |
9.3 Team Topologies 边界设计
| Team type | AI strategy role | Interaction mode |
|---|---|---|
| Stream-aligned team | 负责 AML、客服、信贷、财富等具体 AI 产品价值流 | 主要消费平台 X-as-a-Service,也与平台短期 collaboration |
| Platform team | 提供 model gateway、EvalOps、tool gateway、knowledge baseline、observability | X-as-a-Service;用 platform roadmap 管 internal customer |
| Enabling team | 帮助业务团队掌握 eval、RAG、risk controls、AI-native workflow design | Facilitation;短期嵌入,降低认知负荷 |
| Complicated subsystem team | 深度模型、优化、实体图谱、特征工程、高风险 decisioning engine | Collaboration + internal service;处理专业复杂度 |
9.4 Boundary ADR 模板
# AI Platform Boundary ADR
## Context
- Product / use case: Customer service knowledge platform.
- User need: Agents need current, authorized and cited answers across products and customer states.
- Shared platform capabilities consumed: Model gateway, knowledge ingestion baseline, retrieval evaluation, prompt registry, guardrail service, AI observability.
- Domain capabilities retained locally: Product answer policy, customer eligibility rules, escalation path, tone policy, training and adoption metrics.
## Decision
- Platform owns: Gateway, metadata standard, retrieval baseline, eval runner, trace schema, cost attribution and guardrail primitives.
- Product team owns: Customer journey, answer policy, source owner workflow, launch sequencing, agent coaching and outcome metrics.
- Shared governance owner: CX product lead, compliance reviewer, platform PM and enterprise architect.
- Interaction mode: X-as-a-Service / Collaboration / Facilitation
## Rationale
- Reuse evidence: AML, KYC and service use cases all need retrieval, citation, prompt versioning and trace review.
- Cognitive load reduction: Product teams consume approved ingestion and eval patterns instead of building search and model controls.
- Risk control: Platform enforces permission-before-retrieval and release gates; CX owns answer accuracy and escalation rules.
- Differentiation: Customer trust and agent workflow remain product-specific.
- Cost / speed impact: Shared gateway and retrieval reduce duplicate engineering while allowing domain teams to move independently.
## Consequences
- Benefits: Faster pilots, consistent audit evidence, clearer source ownership and lower duplicate platform spend.
- Trade-offs: More upfront boundary design and shared prioritization.
- Operational responsibilities: Platform SLO for shared services; CX owns knowledge freshness and adoption.
- Metrics: Time-to-safe-pilot, retrieval groundedness, agent adoption, answer correction rate, cost per resolved contact.
- Revisit trigger: Two product teams require unsupported retrieval pattern or platform adoption falls below agreed target.
10. 架构投资路线:从地图到 Architecture Runway
AI 架构投资必须绑定地图上的 bottleneck、control point 和 evolution movement。不要因为某项技术热门就进入 runway。
10.1 投资主题
| Investment theme | Map signal | Architecture deliverable | Business payoff |
|---|---|---|---|
| AI gateway | 多团队接模型、成本不可见、供应商依赖上升 | Model gateway、routing policy、quota、fallback、logs | 降低接入成本,提升可替换性和治理 |
| EvalOps | 多 use case 上线靠主观体验,缺 release evidence | Dataset registry、eval runner、release gate、production sampling | 让质量和风险可决策 |
| Knowledge governance | 多个 RAG 共享低质量知识源 | Source registry、metadata standard、permission filter、freshness SLA | 提高引用可信度和复用 |
| Tool gateway | agent 需要调用内部系统 | Tool catalog、authZ、approval、idempotency、kill switch | 支持 bounded action,控制越权和误操作 |
| Domain evidence graph | AML、欺诈、信贷需要实体和关系证据 | Entity graph、case link、evidence lineage | 提高调查质量和解释力 |
| Policy-as-code | 合规规则散落在文档和 prompt | Machine-enforceable policies、approval workflow | 减少违规输出和变更不可追溯 |
| AI observability | 生产问题难定位 | End-to-end trace、quality signal、incident taxonomy | 提高可运营性和审计能力 |
| Cost / FinOps | 使用增长快,账单不可解释 | Cost attribution、budget caps、cache、unit economics | 防止规模化后 ROI 失真 |
| Developer golden path | 业务团队重复造轮子 | SDK、reference app、templates、docs、support model | 降低认知负荷,加快安全交付 |
10.2 投资优先级评分
| Dimension | 1 分 | 3 分 | 5 分 |
|---|---|---|---|
| Value chain leverage | 只影响单点小任务 | 改善一个流程节点 | 支撑多个高价值 AI value streams |
| Reuse potential | 单一团队 | 两个 use case | 多业务、多产品、多模型复用 |
| Risk reduction | 风险影响不清 | 降低部分上线风险 | 形成正式 release gate 或 control plane |
| Competitive impact | 内部效率 | 改善 time-to-market | 形成领域体验、数据、合规或成本优势 |
| Evolution timing | 市场不成熟且需求弱 | 可试点 | 组件正在产品化,适合 hybrid 抢占窗口 |
| Operating readiness | owner 不清 | 有候选 owner | 平台、业务、风险、SRE owner 明确 |
分数解释:
- 24-30:进入 architecture runway,并配 funding gate。
- 18-23:进入 pilot runway,限定范围和复盘点。
- 12-17:先做 discovery、供应商验证或组织能力补强。
- 低于 12:不进入路线,保留观察。
11. 金融零售案例
以下案例用同一套地图语言:user need、value chain、evolution、platform boundary、sourcing、architecture investment、strategic gameplay。
11.1 AML Copilot
User need
AML investigator 需要在有限时间内识别高风险可疑活动,快速聚合证据,生成可审查的 narrative draft,同时保留人工判断、监管证据链和 SAR 决策责任。
Value chain
| Node | Evolution | Strategic note |
|---|---|---|
| Investigator case outcome | Custom | 高可见、高责任,必须按机构 AML 操作模型设计 |
| Case triage workbench | Custom | 体验和流程差异化明显 |
| SAR narrative draft | Custom | AI 可 draft,但 SAR 决策必须保留给授权人员 |
| Red flags taxonomy | Custom | 机构政策、地区监管和风险偏好本地化 |
| Entity / transaction graph | Custom -> Product | 图谱工具可买,但实体解析、风险解释和 lineage 是核心资产 |
| Evidence retrieval | Product -> Custom | 检索工具成熟,证据来源、权限、引用要内部治理 |
| LLM summarization | Product -> Commodity | 不形成长期差异 |
| Eval cases | Custom | 历史 case、false positive、missed risk、regulator feedback 是核心资产 |
| Audit / evidence binder | Custom -> Product | 工具可买,证据 schema 和监管响应由内部控制 |
Build / buy / partner
| Component | Decision | Rationale |
|---|---|---|
| Foundation model | Buy via gateway | 摘要和文本生成不是差异化 |
| AML workbench UX | Build / hybrid | 工作流、证据展示、升级路径与机构流程强相关 |
| Transaction / entity graph | Hybrid | 图数据库和 entity resolution 可买,风险语义和数据 lineage 内部掌握 |
| Red flags taxonomy | Build + partner | 可让专家顾问加速,但 owner 必须在 financial crime team |
| Eval dataset | Build | 历史 case 和质量门槛是核心控制资产 |
| SAR decision boundary | Build governance | 不允许供应商或 AI 黑箱做最终监管判断 |
Strategic gameplay
- 用 commodity LLM 加速文本处理。
- 在 case evidence graph、workflow UX、eval set 和 audit binder 上 build。
- 通过 partner 加速 taxonomy 和流程迁移,但在 2-3 个季度内完成内部接管。
- 将 evidence retrieval、model gateway、EvalOps 纳入平台,AML-specific red flags 留在产品域。
11.2 客服知识平台
User need
客服坐席和数字渠道需要基于最新、授权、可引用的产品政策回答客户问题,降低平均处理时长、升级率和错误承诺风险。
Value chain
| Node | Evolution | Strategic note |
|---|---|---|
| Customer trust and answer quality | Custom | 品牌、合规话术、客户状态决定体验 |
| Agent assist / customer-facing UI | Product -> Custom | contact center AI 可买,但语气、升级、限制条件要本地化 |
| Product policy knowledge | Custom | source owner、effective date、eligibility 和 jurisdiction 是核心 |
| Retrieval / citation | Product | 可买工具,但 eval 和权限过滤要统一 |
| Knowledge ingestion | Product -> Custom | 连接器可买,metadata 和 governance 是架构投资 |
| Conversation analytics | Product | SaaS 成熟 |
| Guardrails | Product -> Custom | 通用安全可买,金融承诺、费用、豁免、适当性规则需本地化 |
Platform boundary
| Platform owns | Product / CX owns |
|---|---|
| Knowledge ingestion baseline、metadata schema、retrieval eval、model gateway、conversation trace、guardrail service | Answer policy、customer journey、handoff rules、tone、training、adoption、business metrics |
Strategic gameplay
- Buy contact center AI capabilities where mature。
- Build knowledge governance and answer policy。
- Platformize retrieval、EvalOps、guardrails、observability。
- 用 A/B 或 shadow mode 比较 AI-assisted answer quality,但高风险话术先走坐席 copilot,不直接 customer-facing。
11.3 信贷 Decisioning
User need
信贷团队需要更快、更一致、更可解释地完成申请评估、政策例外判断、材料缺口识别和审批 memo,但不能把受监管的最终授信责任交给不可解释的 AI。
Value chain
| Node | Evolution | Strategic note |
|---|---|---|
| Credit decision accountability | Custom | 风险偏好、监管、模型风险和审批责任高度本地化 |
| Underwriter workflow | Custom | 体验和例外处理可差异化 |
| Policy eligibility check | Custom -> Product | 规则引擎成熟,但政策语义和 exception reason 本地化 |
| Document extraction | Product | OCR / extraction 可买 |
| Scorecard / ML risk model | Custom -> Product | 传统模型成熟,GenAI 不替代核心评分责任 |
| Explainability / adverse action reasons | Custom | 高监管敏感,必须内部控制 |
| Eval / backtesting | Custom | 历史申请、拒绝理由、违约结果、fair lending slices 是核心资产 |
Decision boundary
| AI level | Allowed | Not allowed |
|---|---|---|
| Inform | 政策引用、材料缺口、历史相似案例 | 生成未经验证的政策解释 |
| Draft | 贷审 memo 草稿、question list、exception summary | 自动给客户发送拒绝或批准理由 |
| Recommend | 低风险补件建议、人工复核优先级 | 黑箱替代审批人 |
| Act | 创建内部任务、预填字段并等待审批 | 自动批准、自动拒绝、自动调额 |
Strategic gameplay
- Build policy and decision boundary。
- Buy document extraction and workflow acceleration。
- Hybrid scorecard / ML platform,但 model risk、fairness slices、reason code 和 approval responsibility 内部控制。
- Invest in backtesting, adverse action evidence, policy-as-code and audit trail。
11.4 财富顾问
User need
财富顾问需要在客户目标、风险承受能力、持仓、产品适当性和市场信息之间形成更高质量的建议准备材料,同时保持 suitability、披露、审批和客户沟通责任清晰。
Value chain
| Node | Evolution | Strategic note |
|---|---|---|
| Client trust / suitability | Custom | 高可见、高风险,是差异化和监管责任中心 |
| Advisor desktop | Custom / Product | CRM 和财富平台可买,顾问工作流需定制 |
| Portfolio analysis | Product | 分析工具成熟 |
| Product universe and restrictions | Custom | 机构产品、限制名单、客户资格本地化 |
| Meeting prep summary | Product -> Custom | LLM 摘要可买,合规边界需内部设计 |
| Recommendation drafting | Custom | 必须受 suitability、审批和 disclosure 控制 |
| Market commentary | Product | 可买内容,但引用和许可要清楚 |
| Audit and supervision | Custom -> Product | 监管和机构政策强约束 |
Platform boundary
| Shared platform | Wealth product domain |
|---|---|
| Model gateway、retrieval、content licensing metadata、prompt registry、EvalOps、audit logging | Suitability policy、advisor workflow、recommended action boundary、disclosure language、supervision rules |
Strategic gameplay
- 不把“客户面对面的自动投资建议”作为第一阶段目标。
- 先做 advisor copilot:prepare、summarize、compare、draft with approval。
- Build suitability boundary and disclosure controls。
- Partner with content / market data providers,但保留建议策略和客户沟通责任。
11.5 AI Gateway / EvalOps 平台
User need
业务 AI 团队需要快速、安全、可评估、可审计地构建 AI 应用;风险、合规、架构和平台团队需要统一控制模型接入、成本、质量、变更和证据。
Value chain
| Node | Evolution | Strategic note |
|---|---|---|
| AI product teams shipping safely | Custom | internal customer need,要用 platform-as-product 管 |
| Golden paths / SDK | Custom -> Product | 组织上下文强,先内部建 |
| Model gateway | Product | 可 buy / hybrid |
| Prompt / config registry | Product -> Custom | 工具可买,审批和 policy 适配内部 |
| EvalOps | Genesis -> Product | 工具成熟中,eval contract 和 dataset 内部建 |
| AI observability | Product | buy / integrate |
| Cost attribution | Product | buy / integrate,和 internal chargeback 适配 |
| Evidence binder | Custom -> Product | 审计证据 schema 要结合金融治理 |
Platform strategy
| Capability | MVP stance | Scale stance |
|---|---|---|
| Model gateway | Hybrid: approved models, logs, cost attribution | Multi-provider routing, fallback, risk-based policy |
| EvalOps | Build eval contract + buy runner if useful | Dataset registry, release gates, production sampling |
| Knowledge governance | Start with 2 source systems and metadata standard | Enterprise source registry and permission-before-retrieval |
| Tool gateway | Start read-only and approval-only actions | Risk-tiered tool catalog, idempotency, kill switch |
| Developer experience | Reference apps and templates | Self-service platform with SLO and support model |
Strategic gameplay
- 平台从 2-3 个高价值 use case 的重复痛点中抽象,不从空白架构蓝图开始。
- 低可见、重复、控制强的能力平台化。
- 高可见、领域差异的工作流保留给 stream-aligned teams。
- 平台指标用 time-to-safe-pilot、eval coverage、reused components、cost per successful task 和 incident rate,不只看 API calls。
12. 模板
12.1 AI Wardley Map Canvas
# AI Wardley Map Canvas
## User Need
- Primary actor: AML investigator.
- Need: Prepare evidence-backed case narrative faster while preserving human accountability.
- Frequency: Daily case review and escalation workflow.
- Business value: Lower investigation cycle time, better consistency and stronger regulatory evidence.
- Risk / harm if wrong: Missed suspicious activity, fabricated evidence, privacy breach or poor regulator response.
## Value Chain
| Node | Dependency | User visibility | Owner | Evidence |
|---|---|---:|---|---|
| Case narrative | Evidence retrieval, red flags, model gateway | High | AML product owner | Reviewer quality score and SAR audit sample |
| Evidence retrieval | Source registry, permissions, entity graph | Medium | Knowledge / data owner | Citation accuracy and stale-source eval |
| EvalOps | Golden set, failure taxonomy, release gate | Low | EvalOps lead | Critical failure report and approval record |
## Evolution Placement
| Node | Genesis | Custom | Product | Commodity | Rationale |
|---|---:|---:|---:|---:|---|
| SAR narrative boundary | Low | High | Low | Low | Institution policy and accountability are local |
| LLM summarization | Low | Low | High | Medium | Multiple providers can perform generic summarization |
| Audit logging | Low | Medium | High | Low | Tools mature, but evidence schema is institution-specific |
## Climatic Patterns
| Pattern | Impact on this map | Timing | Confidence |
|---|---|---|---|
| Model commoditization | Model choice becomes less differentiating than evidence and eval | 6-12 months | High |
| Regulatory scrutiny | SAR-related AI output needs stronger audit and human oversight | Continuous | High |
## Strategic Gameplay
| Node | Play | Build / buy / partner | Architecture investment | Reversal trigger |
|---|---|---|---|---|
| Evidence retrieval | Hybrid platform capability | Buy search components, build source governance | Knowledge registry and retrieval eval | Citation accuracy below release threshold |
| SAR narrative boundary | Build domain control | Internal policy and approval design | Workflow audit and human review queue | Reviewer override rate remains excessive |
12.2 Evolution Heatmap
| Component | Current stage | Evidence | Movement expected | Action | Owner |
|---|---|---|---|---|---|
| Foundation model access | Product -> Commodity | Multiple providers, falling prices, API parity | More commodity | Buy through gateway | Platform |
| Domain policy | Custom | Institution policy, jurisdiction, risk appetite | Slow movement | Build / govern | Domain |
| Eval dataset | Custom | Internal historical cases and expert labels | Becomes strategic asset | Build | EvalOps + domain |
12.3 Strategic Gameplay Register
| Play ID | Map node | Move | Why now | Expected advantage | Risk | Control | Review date |
|---|---|---|---|---|---|---|---|
| PLAY-001 | Model gateway | Hybrid platform | Multiple teams integrating models separately | Cost, audit, routing, replacement | Platform friction | Golden path + support | 2026-09-30 |
12.4 Competitive Landscape Canvas
| Competitor / vendor | Value chain nodes controlled | Evolution stage | Strength | Weakness | Our response |
|---|---|---|---|---|---|
| Domain AML vendor | Case workbench, typology templates, investigation accelerators | Product / custom | Fast pilot and domain workflow patterns | May not expose full eval, audit and policy control | Use for acceleration, keep SAR boundary and evidence schema internal |
| Hyperscaler AI platform | Model access, search, observability, security primitives | Product / commodity | Scale and enterprise integration | Limited domain workflow | Consume infrastructure, build domain controls and eval assets |
12.5 Architecture Investment Thesis
# Architecture Investment Thesis
## Investment
- Name: Enterprise AI Gateway and EvalOps control plane.
- Value chain bottleneck: Multiple AI product teams need safe model access, release evidence, traceability and cost control.
- Components affected: Model gateway, prompt registry, eval runner, dataset registry, audit log, cost attribution.
- Risk tier: Medium to high, because the platform supports both employee copilot and regulated workflow use cases.
## Strategic Reason
- User need improved: Stream-aligned teams can launch AI pilots faster without rebuilding model, eval and audit controls.
- Competitive gap closed: Reduces time-to-safe-pilot while competitors remain blocked by fragmented PoCs or vendor lock-in.
- Platform reuse: AML, KYC, service, credit and wealth use cases share gateway, eval and observability.
- Control / compliance benefit: Creates consistent release gate, model log, prompt version, dataset lineage and evidence binder.
## Evidence
- Current pain: Teams connect models directly, release quality is reviewed inconsistently and cost attribution is unclear.
- Pilot signal: Two or more use cases require the same model logging, eval runner and retrieval trace capabilities.
- Cost of delay: Duplicate integrations, delayed risk signoff, weak audit reconstruction and rising token spend.
- Alternatives considered: Allow local implementations, buy a full AI suite, or build a lightweight internal control plane.
## Delivery Shape
- Build / buy / partner: Hybrid, using mature gateway / observability components while building internal policies, eval contracts and evidence schema.
- Platform boundary: Platform owns shared controls; product teams own domain workflow and outcome metrics.
- Teams involved: Platform, security, model risk, EvalOps, enterprise architecture and two pilot product teams.
- Release gates: Gateway onboarding, eval threshold approval, production sampling and quarterly map review.
## Success Metrics
- Adoption: Percent of new AI use cases using gateway and EvalOps before pilot.
- Quality: Percent of releases with passing eval report and zero critical failures.
- Risk: Audit reconstruction completeness and incident closure time.
- Cost: Cost per successful task and model spend by use case.
- Business outcome: Time-to-safe-pilot and reuse across business teams.
## Stop / Revisit Rule
- Stop if: Product teams bypass the platform because golden paths are slower than local implementation.
- Revisit if: Vendor gateway becomes insufficient for internal policy, audit or data boundary requirements.
- Scale if: Three high-value use cases pass release gates and show measurable reduction in integration and review time.
13. 评审清单
13.1 Map quality checklist
| Check | Pass signal |
|---|---|
| User need 清晰 | 能说清 actor、task、outcome、risk、frequency |
| Value chain 完整 | 包含 workflow、decision boundary、domain policy、data、model、eval、audit、infra |
| Evolution placement 可挑战 | 每个关键节点有供应商、市场、内部能力或案例证据 |
| 依赖关系真实 | 上层体验依赖的底层知识、权限、eval 和控制被画出 |
| 趋势假设明确 | 包含 commoditization、regulation、cost、ecosystem、inertia |
| Gameplay 可执行 | 每个动作有 owner、投资、边界和复盘条件 |
13.2 Build / buy / partner checklist
| Check | Pass signal |
|---|---|
| 不是先有结论 | 决策由 map position、risk、differentiation、control 和 economics 推导 |
| Vendor risk 可管理 | 数据、日志、模型版本、subprocessor、SLA、exit path 清楚 |
| 内部能力不空心化 | 核心 policy、eval、decision boundary 和 audit owner 留在内部 |
| Partner 有退出机制 | 明确 knowledge transfer、handover 和内部 owner |
| Build 有运行能力 | 团队、SRE、security、risk、support 和 funding 可持续 |
13.3 Platform boundary checklist
| Check | Pass signal |
|---|---|
| Internal customer 明确 | 平台服务哪些 stream-aligned teams 和哪些任务 |
| Cognitive load 降低 | 平台提供 golden path、SDK、reference app、support model |
| 复用证据充分 | 至少 2 个高价值 use case 共享能力或控制点 |
| 不吞掉 domain ownership | 业务 team 保留 workflow、domain eval、adoption、business outcome |
| Interaction mode 清楚 | Collaboration、X-as-a-Service、Facilitation 被明确选择 |
| Metrics 不虚 | 使用 safe pilot speed、reuse、eval coverage、cost、incident,而不是只看调用量 |
13.4 Financial retail AI risk checklist
| Area | Pass signal |
|---|---|
| Customer impact | 明确客户权益、资金、信用、隐私和投诉影响 |
| Human oversight | 高风险 decisioning 有人工责任、审批和 override |
| Explainability | 信贷、财富、AML 输出有 reason、evidence、citation 或 policy reference |
| Fairness / bias | 信贷、营销、服务优先级有 slice analysis 和监控 |
| Audit | 能重建 input、context、model、prompt、tool、output、approval、final action |
| Incident | 有 unsafe output、data leakage、wrong advice、cost runaway、tool misuse 的处理流程 |
| Reversibility | 可以 rollback、disable、switch model、re-index、replay eval、通知 owner |
14. 反模式
| Anti-pattern | 表现 | 后果 | 纠正动作 |
|---|---|---|---|
| Chatbot-as-strategy | 所有需求都变成聊天入口 | 高风险流程被浅层 UI 掩盖 | 回到 user need 和 workflow insertion point |
| Vendor-first strategy | 先选供应商再定义问题 | 锁定、范围膨胀、价值不清 | 先画 map,再做 vendor scorecard |
| Model-as-moat | 把某个 foundation model 当长期优势 | 快速被商品化 | 投资 workflow、data、eval、trust、distribution |
| RAG as data governance | 把文档丢进向量库就算知识平台 | 过期、越权、错误引用 | 建 source registry、metadata、permission、freshness |
| Eval theatre | 上线前跑几个 happy path | 生产风险不可见 | 建 failure taxonomy、golden set、critical gates |
| Platform maximalism | 平台团队试图统一所有业务流程 | 业务绕行,平台 adoption 低 | 平台只做低可见、重复、控制强能力 |
| Product team local optimum | 每个团队自建模型、RAG、日志、eval | 成本高、治理不一致 | 共享 model gateway、EvalOps、audit 和 guardrails |
| Partner dependency trap | SI / vendor 永久掌握核心知识 | 内部能力空心化 | 设定 knowledge transfer 和 internal owner |
| Automation before decision rights | 先让 agent 执行动作,再补审批 | 越权、事故、审计失败 | 先画 decision boundary 和 tool permission |
| All-build ideology | 认为自研才安全 | 慢、贵、重复造轮子 | 商品化组件 buy,控制面和差异化 build |
| All-buy ideology | 认为 vendor handles everything | 核心责任外包、审计薄弱 | 内部保留 policy、eval、audit、decision boundary |
| Static map | 地图画完不更新 | 路线和市场脱节 | 每季度复盘 evolution movement 和 gameplay |
15. 30 天训练计划
目标:30 天内形成一套可展示的 AI Wardley Mapping portfolio,不训练基础 PM 技能,只训练战略地图、平台边界、build / buy / partner、竞争地形和架构投资判断。
| Day | Focus | Output | Review question |
|---|---|---|---|
| 1 | 阅读 Wardley Mapping book 目录和核心概念 | 1 页方法摘要 | user need、value chain、evolution 如何互相约束 |
| 2 | 拆解 NIST AI RMF | AI risk 到 map 节点映射表 | 哪些地图节点必须进入 Govern / Measure / Manage |
| 3 | 拆解 Team Topologies | AI 平台团队边界草图 | 平台 team 服务谁,降低什么认知负荷 |
| 4 | 选择 5 个金融零售 AI 场景 | Scenario shortlist | 每个场景的 actor 和 risk 是否不同 |
| 5 | AML Copilot user need | AML user need + value chain v1 | 哪些节点高可见、哪些节点低可见 |
| 6 | AML evolution mapping | AML evolution heatmap | 哪些 build,哪些 buy,哪些 partner |
| 7 | AML strategic gameplay | AML build-buy-partner memo | SAR 决策边界是否清楚 |
| 8 | 客服知识平台 value chain | Service knowledge map v1 | 知识治理瓶颈在哪里 |
| 9 | 客服平台边界 | Platform vs CX boundary ADR | 平台和 CX 团队 owner 是否分清 |
| 10 | 客服竞争地形 | Vendor / SaaS landscape table | 哪些 vendor 控制流程入口 |
| 11 | 信贷 decisioning map | Credit decisioning map v1 | AI 允许影响哪些决策,禁止哪些动作 |
| 12 | 信贷 eval / fairness | Credit eval slice plan | 哪些 failure 是 release blocker |
| 13 | 信贷 sourcing memo | Credit hybrid decision memo | 哪些政策和 reason code 必须内部掌握 |
| 14 | 财富顾问 map | Wealth advisor copilot map | suitability 和 disclosure 边界是否显式 |
| 15 | 财富顾问 gameplay | Advisor copilot phased strategy | 为什么先 advisor copilot 而不是客户自动建议 |
| 16 | AI gateway map | AI gateway value chain | gateway 是产品、平台还是控制面 |
| 17 | EvalOps map | EvalOps platform map | eval contract 和工具边界是否清楚 |
| 18 | Knowledge governance map | Enterprise knowledge map | source owner、freshness、permission 是否可执行 |
| 19 | Tool gateway map | Agent tool gateway control map | read、draft、act 的权限是否分层 |
| 20 | Competitive landscape | 6 类竞争者地形图 | 我方在哪些节点能赢 |
| 21 | Climatic pattern synthesis | AI climatic pattern memo | 哪些趋势会改变 12 个月路线 |
| 22 | Architecture investment scoring | 10 个投资主题评分 | 哪些进入 runway,哪些等待 |
| 23 | Portfolio roadmap | 4 季度 AI strategy roadmap | 顺序是否由依赖和演进推导 |
| 24 | Platform operating model | Team Topologies interaction model | collaboration 何时转 X-as-a-Service |
| 25 | Executive memo | 2 页高管决策备忘录 | 是否能解释 trade-off 而非堆术语 |
| 26 | ARB review pack | 架构评审包 | 是否包含 map、ADR、risk、eval、exit |
| 27 | Interview answers | 10 道高级面试答案 | 是否能用地图回答 build / buy / platform |
| 28 | Portfolio packaging | 作品集目录和 artifact links | 每个 artifact 是否证明一种能力 |
| 29 | Mock review | 自我挑战假设清单 | 哪些 map node 证据不足 |
| 30 | Final synthesis | AI Wardley Mapping case portfolio | 能否 15 分钟讲清战略、架构和投资路线 |
16. 面试答案
Q1:为什么 AI PM / 架构师要用 Wardley Map,而不是只做 roadmap?
30 秒回答: Roadmap 告诉团队什么时候做什么,但不解释为什么这个顺序是对的。Wardley Map 先把 user need、value chain、组件演进位置和外部趋势画出来,再推导 build / buy / partner、平台边界和架构投资。对 AI 来说,这能避免被模型 hype、供应商 demo 或孤立 PoC 牵着走。
2 分钟回答: AI 产品不是单一功能,它依赖 workflow、data、knowledge、model、eval、security、audit 和组织责任。传统 roadmap 很容易把这些依赖压扁成日期和功能列表。Wardley Map 的优势是展示地形:上层哪些体验和决策直接影响用户价值,下层哪些组件已经商品化,哪些仍是定制瓶颈,哪些趋势会不可逆地改变市场。比如 foundation model API 正在商品化,所以我不会把单一模型当战略护城河;但 AML evidence graph、credit reason code、wealth suitability boundary 和 eval dataset 更接近差异化和控制点,应该 build 或 hybrid。地图让路线、平台投资和供应商决策可辩护。
Q2:如何用地图做 AI build / buy / partner 决策?
30 秒回答: 我先看组件在 value chain 的可见性和 evolution stage,再看差异化、风险、控制权、规模经济和内部运行能力。高可见、高差异、高责任的组件倾向 build / hybrid;低可见、成熟、非差异的组件倾向 buy;需要领域迁移或实施加速时 partner,但核心 policy、eval、audit 和 decision boundary 不外包。
2 分钟回答: 以客服知识平台为例,contact center AI、语音转写、基础检索都比较产品化,可以 buy;但金融产品政策、费用豁免边界、客户资格、过期文档处理、引用规则和升级路径是本地化控制,需要内部 owner。以 AML 为例,LLM summarization 可以通过 gateway 消费,但 red flag taxonomy、case narrative boundary、SAR approval 和历史 case eval set 是核心资产。Partner 可以帮助 taxonomy 迁移和流程落地,但必须有 knowledge transfer 和内部接管条件。这个决策不是采购偏好,而是地图上的位置推导。
Q3:AI 平台和业务 AI 产品的边界怎么划?
30 秒回答: 平台做低可见、重复、控制强、可服务多个 use case 的能力,例如 model gateway、EvalOps、tool gateway、knowledge ingestion baseline、audit、cost governance。业务产品保留 user need、workflow、domain policy、domain eval、adoption 和业务结果。
2 分钟回答: 我会用 Team Topologies 的视角看:platform team 应该提供 compelling internal product,降低 stream-aligned teams 的认知负荷,而不是接管业务流程。比如平台提供 model gateway、prompt registry、retrieval eval、tool gateway 和 observability;AML 团队负责 red flags、case workflow、SAR boundary;信贷团队负责 policy exception、reason codes、fairness slices;财富团队负责 suitability 和 disclosure。interaction mode 也要设计:早期通过 collaboration 共同探索,稳定后变成 X-as-a-Service,能力缺口通过 enabling team facilitation 补齐。
Q4:AI 竞争优势来自哪里?
30 秒回答: 长期优势通常不来自单个 foundation model,而来自高价值 workflow、受控数据和知识、领域 eval、用户信任、分发入口、合规控制和组织执行速度。地图可以显示哪些节点会商品化,哪些节点值得投资。
2 分钟回答: 在 AI 领域,很多底层能力会快速产品化或商品化:模型 API、基础 RAG、observability、OCR、语音识别。真正难复制的是企业自己的交易历史、case labels、知识治理、审批流程、监管解释、客户关系和 adoption 机制。例如 AML Copilot 的优势不是“会总结文本”,而是能把交易、实体、历史案例、red flags 和 SAR narrative 证据链串起来;财富顾问的优势不是“能聊天”,而是能在 suitability、disclosure 和 advisor workflow 中形成可信建议准备材料。地图帮助我把资源投到不会马上被市场商品化的节点。
Q5:如何避免 AI 平台过早平台化?
30 秒回答: 我要求至少有多个高价值 use case 证明同一能力重复出现,并且平台能显著降低认知负荷、风险和交付时间。第一个 PoC 不建大平台,先抽象最小 golden path。
2 分钟回答: 过早平台化的典型表现是先建统一 agent platform、统一 UI builder、统一 ontology,但业务场景还没有验证。我的做法是从 2-3 个真实场景中提取重复组件,比如 AML、客服和信贷都需要 model gateway、eval、audit、knowledge ingestion 和成本归因,那么这些适合平台化。相反,AML red flags、信贷 exception policy、财富 suitability 不应由平台统一。平台 MVP 应该包括 gateway、prompt/config registry、EvalOps、knowledge baseline、observability 和 cost dashboard,而不是承诺无代码生成所有 AI 应用。
Q6:AI gateway / EvalOps 为什么是架构投资,不只是工具采购?
30 秒回答: AI gateway 是模型接入、路由、成本、日志和政策控制面;EvalOps 是质量、风险和发布门禁控制面。工具可以买,但 control policy、eval contract、golden set、risk thresholds 和 evidence binder 必须由企业掌握。
2 分钟回答: 当企业有多个 AI use case,模型调用、prompt 变更、RAG index、tool call 和输出质量都会持续变化。没有 gateway,团队会分散接模型,成本和审计不可控;没有 EvalOps,上线只能靠主观体验。工具采购只能解决部分执行问题,真正的架构投资是建立统一的控制面:请求日志、模型版本、fallback、quota、dataset registry、failure taxonomy、release gate、production sampling、incident loop。这些能力让 AI 从 PoC 进入可运行、可审计、可复盘的生产系统。
Q7:金融零售高风险 AI decisioning 怎么设计边界?
30 秒回答: 先明确 AI 是 inform、draft、recommend、act with approval 还是 autonomous act。信贷、财富、AML 等高风险场景一般不让 AI 直接做最终决定,而是让 AI 生成证据、摘要、建议和草稿,由授权人员或规则系统承担决策责任。
2 分钟回答: 以信贷为例,AI 可以提取材料、引用政策、发现缺失信息、草拟贷审 memo、提示例外原因,但自动批准、自动拒绝、自动生成客户 adverse action reason 都需要严格控制。以财富顾问为例,AI 可以做 meeting prep、产品比较、风险提示草稿,但 suitability 判断和客户沟通必须经过顾问和监督流程。地图上这些 decision boundary 是高可见、高风险、高差异节点,因此不能交给黑箱供应商。必须配套 human oversight、audit trail、fairness slice eval、policy-as-code 和 rollback。
Q8:如何把 Wardley Map 转成 12 个月 AI roadmap?
30 秒回答: 先识别 value chain bottleneck 和可复用低层能力,再按 evolution 和依赖安排顺序:先补 doctrine 和控制面,再做高价值场景 pilot,最后把重复能力平台化并规模化。
2 分钟回答: 我会把 roadmap 分成三条线。第一条是场景线:AML、客服、信贷、财富按价值和风险推进 pilot。第二条是平台线:model gateway、EvalOps、knowledge governance、tool gateway、observability 按重复痛点建设。第三条是治理线:risk tiering、release gate、evidence binder、incident process。顺序不是任意的:如果没有 knowledge governance,RAG 场景会失败;如果没有 EvalOps,高风险场景不能 scale;如果没有 gateway,供应商和成本会失控。地图让 roadmap 顺序由依赖和演进推导,而不是由组织政治推导。
17. 作品集交付物
一个高级 AI PM / 架构师作品集不应只展示 PRD,而要展示战略判断和架构决策证据。
| Deliverable | 内容 | 证明的能力 |
|---|---|---|
| AI Wardley Map Pack | AML、客服、信贷、财富、AI gateway / EvalOps 五张地图 | 能从 user need 到 value chain 和 evolution 做战略分析 |
| Evolution Heatmap | 20 个 AI 组件的 Genesis / Custom / Product / Commodity 判断 | 能识别商品化趋势和投资窗口 |
| Build-Buy-Partner Memo | 每个案例的 sourcing decision 和边界 | 能做 vendor / build / partner 取舍 |
| Platform Boundary ADR | 平台团队与业务团队的 owner、interaction mode、metrics | 能设计 AI 平台边界和组织协作 |
| Competitive Landscape Map | hyperscaler、foundation model provider、SaaS、domain vendor、SI、open source、internal platform 对比 | 能看竞争地形,不停留在 feature list |
| Architecture Investment Thesis | gateway、EvalOps、knowledge governance、tool gateway 等投资论证 | 能把架构投资转成业务、风险和速度语言 |
| Financial Retail Case Portfolio | AML Copilot、客服知识平台、信贷 decisioning、财富顾问、AI gateway / EvalOps | 能把金融零售业务背景转成高级 AI 策略 |
| ARB Review Pack | map、ADR、risk tier、eval gate、control pack、exit plan | 能进入架构评审和高管决策场景 |
| Interview Answer Bank | 10 道 build / buy / platform / governance / AI strategy 答案 | 能在面试中讲清高级战略判断 |
| 30-Day Training Evidence | 每日输出、复盘问题、最终 synthesis | 能证明学习方法和可持续迭代能力 |
17.1 Portfolio narrative
我用 Wardley Mapping 处理 AI 产品战略,不从模型或供应商开始,而从 user need 和 value chain 开始。
对每个金融零售场景,我先识别用户价值、决策边界、风险和依赖,再把组件放到 evolution axis。
成熟和非差异组件通过 buy / hybrid 加速,核心 workflow、policy、eval dataset、audit 和 decision boundary 保留内部控制。
当多个场景共享低可见但高复用能力时,我把它们沉淀为 AI platform capabilities,并用 Team Topologies 定义平台团队与业务团队的交互方式。
最终输出不是一张静态图,而是一套可评审、可投资、可上线、可复盘的 AI strategy operating system。
18. Executive Summary 模板
# Executive Summary: AI Wardley Mapping Strategy
## Strategic Position
本阶段不把 AI 战略定义为“采购某个模型”或“上线若干 chatbot”,而是围绕金融零售高价值工作流建立 AI 产品与平台组合。
## Key Map Insights
- Foundation model access 正在商品化,单模型不是长期差异化。
- Knowledge governance、domain eval dataset、decision boundary 和 audit evidence 是金融零售 AI 的关键控制资产。
- 多个 use case 共享 model gateway、EvalOps、tool gateway、observability 和 cost governance,适合平台化。
- AML、信贷、财富等高风险场景必须保留人工责任、政策解释和监管证据链。
## Recommended Plays
- Buy / hybrid 成熟底层能力:model access、basic RAG tooling、observability、OCR、contact center AI。
- Build 差异化和控制节点:domain workflow、policy boundary、eval dataset、audit evidence、decision rights。
- Partner 加速 taxonomy、implementation 和 change management,但用明确 handover 防止能力外包。
- Platformize 低可见、重复、控制强能力,保持业务团队对 user need 和 outcome 负责。
## Architecture Investments
- AI gateway
- EvalOps
- Knowledge governance
- Tool gateway
- Policy-as-code
- AI observability and incident loop
- Cost governance
## Governance
所有高风险 AI use case 必须通过 risk tiering、eval gate、human oversight、audit trail、incident runbook 和 quarterly map review。
19. 最终判断框架
当一个 AI 产品或平台决策进入评审时,用下面五句话结束讨论:
- 这个 user need 是否足够重要,且不是技术驱动的伪需求。
- 这条 value chain 中哪些节点真正决定差异化、控制权和风险。
- 每个关键组件在 evolution axis 上的位置是否支持当前 build / buy / partner 决策。
- 平台边界是否降低业务团队认知负荷,而不是夺走业务责任。
- 架构投资是否服务于地图上的 bottleneck、climatic pattern 和 strategic gameplay,而不是服务于技术偏好。
能回答这五个问题,AI PM / 架构师就能把 AI 战略从“功能路线”升级为“态势感知 + 架构投资 + 组织执行”的组合能力。