AI 底层逻辑 / 经典论文
AI Enterprise Architecture：TOGAF / ArchiMate / ADM
一句话:
295 行ai-foundations/papers/91-ai-enterprise-architecture-togaf-archimate-adm.md
AI Enterprise Architecture / TOGAF / ArchiMate / ADM 解读
面向对象: Enterprise Architect / AI Solutions Architect / AI Product Architect / AI Platform PM / Senior BA。 核心问题: AI 转型如果只靠单个 reference architecture 或项目级 PRD, 很快会变成 POC 堆积、治理重复、平台能力分散、业务线各自造轮子。企业架构方法的价值是把 AI use case、能力地图、平台资产、治理门禁和路线图放进同一套可演进体系。 学习目标: 用 TOGAF ADM、ArchiMate、architecture repository、architecture governance 的思路, 把 AI 从“项目交付”提升为“企业能力演进”。
Source Anchors
| Source | Link | 用途 |
|---|---|---|
| TOGAF | https://www.opengroup.org/togaf | 参考 Architecture Development Method、architecture governance、architecture repository 和 enterprise architecture practice |
| ArchiMate | https://www.opengroup.org/archimate-forum/archimate-overview | 参考业务、应用、技术、动机、实现迁移等架构层的建模语言 |
| NIST AI RMF | https://www.nist.gov/itl/ai-risk-management-framework | 将 AI 风险管理融入 EA 的 govern、map、measure、manage 循环 |
| ISO/IEC 42001 | https://www.iso.org/standard/81230.html | 将 AI management system、责任、控制、持续改进纳入企业架构治理 |
| C4 Model | https://c4model.com/ | 用于补足系统边界、container/component 的工程表达 |
一句话:
AI Enterprise Architecture 是把 AI use case 从点状项目拉回企业能力、价值流、平台资产、控制体系和迁移路线图中管理。
1. 为什么 AI EA 不能只靠技术参考架构
技术参考架构通常能回答:
experience -> orchestration -> model gateway -> RAG -> tool gateway -> eval -> observability
但它不能单独回答:
| EA 关切 | 技术参考架构的缺口 |
|---|---|
| 哪些业务能力最值得 AI 化 | 缺少 capability / value stream 视角 |
| 哪些 POC 应该合并成平台资产 | 缺少 portfolio / reuse / product line 视角 |
| 哪些系统和团队边界会被 AI 改变 | 缺少 org / operating model / Conway 视角 |
| 哪些风险控制必须成为 enterprise standard | 缺少 governance / control library 视角 |
| 哪些数据、知识、语义资产是共享底座 | 缺少 information architecture 视角 |
| 哪些迁移步骤必须分阶段完成 | 缺少 roadmap / transition architecture |
| 哪些例外会累积成架构债 | 缺少 architecture repository 和 conformance |
AI EA 的目标不是“把所有项目审批得更慢”, 而是回答:
How do we repeatedly turn AI opportunities into safe, reusable, measurable enterprise capabilities?
2. TOGAF ADM for AI
TOGAF ADM 可以被改造成 AI capability lifecycle。
| ADM phase | AI 改造重点 | 关键产物 |
|---|---|---|
| Preliminary | 建立 AI architecture principles、risk taxonomy、平台治理边界 | AI architecture principles, AI governance charter |
| A Architecture Vision | 定义 AI 转型愿景、目标能力、价值假设、stakeholder concern | AI vision, capability heatmap, value/risk thesis |
| B Business Architecture | 建模 value stream、capability、role-task、policy decision | AI capability map, value stream map, process/decision model |
| C Information Systems Architecture | 设计应用、数据、知识、语义、RAG、tool/API 边界 | application/data/knowledge architecture |
| D Technology Architecture | 设计模型、平台、工具网关、观测、安全、部署 | AI platform reference architecture |
| E Opportunities and Solutions | 组合 use case、平台能力、vendor/build/buy、migration package | opportunity portfolio, solution building blocks |
| F Migration Planning | 排序 pilot、release、scale、platformization、debt paydown | transition architecture, roadmap, funding gate |
| G Implementation Governance | 评审 solution conformance、release gate、evidence bundle | architecture contract, compliance review |
| H Architecture Change Management | 处理模型/数据/policy drift、事故、监管变化、能力演进 | change trigger, architecture backlog |
| Requirements Management | 贯穿需求、eval、control、traceability graph | requirement-eval-control graph |
高级点在于: AI requirement management 不只是收需求, 而是持续维护需求、评测、控制和证据的关系。
3. AI Architecture Principles
AI EA 需要原则, 但原则必须能落到决策。
| Principle | Design implication |
|---|---|
| Safe value before broad automation | 高风险场景先 assist/draft/HITL, 再逐步自动化 |
| Model is replaceable, evidence is durable | 避免把供应商模型写死, 证据和接口保持可迁移 |
| Knowledge must have ownership | RAG source 必须有 owner、freshness、permission、lineage |
| Tools require contracts and controls | 工具必须通过 contract、policy、approval、audit |
| Eval is part of architecture | 无 eval 的 AI capability 不得进入 release |
| Platformize repeated controls | 模型路由、RAG、eval、tool gateway、observability、evidence 应优先平台化 |
| Risk-tiered governance | 低风险实验快速, 高风险系统强证据 |
| Human accountability remains explicit | 人机责任、override、appeal、signoff 必须建模 |
原则要进入:
- architecture review gate。
- ADR。
- platform service catalog。
- solution conformance checklist。
- evidence binder。
4. ArchiMate for AI
ArchiMate 的价值是跨层表达业务、应用、技术、动机和实现迁移。AI 系统可以这样映射:
| ArchiMate layer | AI 元素 |
|---|---|
| Strategy | AI value stream、capability、course of action、resource |
| Motivation | driver、goal、outcome、requirement、constraint、principle |
| Business | role、business process、business service、decision、policy |
| Application | AI assistant、orchestrator、RAG service、eval service、tool gateway |
| Data / Information | knowledge source、ontology、feature、evidence object、prompt/config |
| Technology | model runtime、vector DB、workflow engine、telemetry stack、deployment |
| Implementation & Migration | work package、plateau、gap、deliverable、transition architecture |
AI-specific extensions can be handled as stereotypes or tagged elements:
| Tagged concept | Use |
|---|---|
riskTier | low / medium / high / material |
modelRoute | approved model policy |
evalGate | release threshold and eval pack |
humanOversight | none / sampled / required / dual control |
evidenceObject | ADR、eval report、approval、trace sample |
dataBoundary | public/internal/confidential/restricted |
toolSideEffect | read/draft/write/high-risk |
不要为 AI 发明一套完全孤立的建模语言。最好做法是: 用 ArchiMate 表达企业层, 用 C4/sequence/detail views 补系统实现, 用 evidence graph 连接治理。
5. Architecture Repository for AI
企业 AI 架构需要 repository, 否则知识会散落在 POC 文档、Slack、PRD、供应商 deck 和代码配置里。
| Repository section | AI 内容 |
|---|---|
| Architecture principles | AI 原则、风险偏好、模型选择原则 |
| Reference architectures | RAG、Agent、Decision Service、Document AI、Customer-facing AI |
| Standards | tool contract、eval contract、telemetry schema、evidence object schema |
| Reusable building blocks | model gateway、RAG service、tool gateway、HITL、eval service |
| Capability maps | AI-enabled business capabilities and platform capabilities |
| ADR library | model, RAG, tool, privacy, release, vendor decisions |
| Control library | policy, eval, monitoring, human oversight controls |
| Evidence binder | release reports, signoffs, trace samples, incidents |
| Debt register | prompt debt、data debt、eval coverage gap、vendor lock-in |
| Roadmap | transition architecture and migration waves |
AI 架构 repository 不只是归档, 它应该驱动:
- reuse discovery。
- architecture conformance。
- audit response。
- onboarding。
- portfolio review。
- platform roadmap。
6. AI Architecture Governance
治理要按风险和成熟度分层。
| Governance object | Owner | Review cadence |
|---|---|---|
| AI architecture principles | EA + risk + platform | semiannual |
| AI use case intake | product + BA + risk | per idea |
| Architecture package | solution architect | pilot/release/scale |
| Fitness function catalog | architect + EvalOps + risk | per release / quarterly |
| Platform service catalog | platform PM | monthly |
| Evidence binder | risk/audit owner | release / audit cycle |
| Exception register | risk + product owner | monthly until closed |
| Architecture debt register | EA + engineering | quarterly |
高成熟度治理的特征:
- 门禁有自动化证据。
- 例外有过期和补偿控制。
- 架构决策有反转条件。
- 平台复用有指标。
- 业务价值和风险控制一起看。
7. Financial Retail Case: AI Architecture Roadmap
目标: 一家区域银行要在 18 个月内从分散 AI POC 走向 enterprise AI operating capability。
7.1 Capability View
| Capability | Current state | Target state |
|---|---|---|
| Customer service knowledge assistance | 多个 chatbot POC | approved policy RAG golden path |
| AML investigation support | 手工查系统和规则 | AML copilot with evidence graph |
| Credit document review | 文档人工抽取 | document AI + human validation |
| Agent workflow | 零散脚本 | tool gateway + HITL + workflow state |
| EvalOps | 项目自测 | platform eval service and release gate |
| Evidence | 审计前补材料 | evidence binder by default |
7.2 Transition Architecture
| Plateau | Description | Exit criteria |
|---|---|---|
| P0 POC inventory | 盘点所有 AI POC、数据、模型、风险、owner | inventory complete, duplicates identified |
| P1 Controlled pilot | 建立 model gateway、basic eval、RAG source registry | 3 pilots pass safe release gate |
| P2 Platform foundation | tool gateway、observability、HITL、evidence binder | common services reused by 5 use cases |
| P3 Product line scaling | customer-facing RAG、agent workflow、document AI golden paths | reuse rate and value metrics stable |
| P4 Operating capability | quarterly AI portfolio review、control library、architecture repository | audit-ready, continuous improvement |
7.3 Architecture Work Packages
| Work package | Building blocks |
|---|---|
| AI control plane | model gateway、policy engine、eval service、observability |
| Knowledge architecture | source registry、ontology、permission、freshness、citation |
| Agent action architecture | tool contract、side-effect taxonomy、approval、audit |
| Governance architecture | control library、evidence binder、fitness functions |
| Product architecture | golden paths、service catalog、adoption dashboard |
8. Templates
AI ADM Canvas
| Phase | Questions |
|---|---|
| Vision | 哪些 AI outcomes 值得投, 哪些风险不可接受 |
| Business | 哪些 value streams/capabilities/roles 会改变 |
| Information | 哪些知识、数据、语义、证据必须治理 |
| Application | 哪些 AI services、tools、workflows、systems 需要边界 |
| Technology | 哪些 model/runtime/vector/telemetry/deployment 能力是底座 |
| Opportunities | 哪些 use cases 该做、买、平台化、停止 |
| Migration | 哪些 transition plateau 和依赖 |
| Governance | 哪些 conformance checks、fitness functions、evidence gates |
| Change | 哪些 trigger 会导致架构重审 |
ArchiMate AI Mapping Table
| Element | AI example | Evidence |
|---|---|---|
| Driver | reduce AML investigation cycle time | business case |
| Goal | improve evidence completeness | metric tree |
| Requirement | brief must cite approved sources | eval contract |
| Constraint | no SAR filing decision by AI | policy control |
| Business process | alert triage | BPMN |
| Application service | AML copilot orchestration | C4 container |
| Data object | transaction evidence graph | lineage |
| Technology service | model gateway | service catalog |
| Work package | deploy tool gateway | roadmap |
9. Common Failure Modes
| Failure mode | 表现 | 修正 |
|---|---|---|
| POC sprawl | 多个团队重复做 chatbot/RAG/eval | EA inventory + platform service catalog |
| Method theater | ADM/ArchiMate 成为文档流程 | 只建模会影响决策和治理的元素 |
| Technology-only EA | 只画模型和平台组件 | 加入 business capability, information, governance, migration |
| No transition architecture | 只有目标图, 没有迁移路径 | plateau + work package + exit criteria |
| No repository | 决策和证据散落 | architecture repository and evidence binder |
| Overstandardization | 所有 use case 强行统一 | commonality/variability and risk-tiered governance |
10. 面试表达
30 秒版本:
我会把 AI 企业架构当成能力演进体系, 而不是单个技术参考架构。用 TOGAF ADM 管理从愿景、业务能力、信息系统、技术架构到迁移和治理的闭环; 用 ArchiMate 表达跨业务、应用、数据、技术和迁移层的关系; 用 architecture repository 管理可复用 building blocks、ADR、控制和证据。
2 分钟版本:
以银行 AI 转型为例, 我会先做 POC inventory 和 capability heatmap, 找到客服 RAG、AML copilot、document AI、agent workflow 等能力族。然后通过 ADM 把业务能力、数据/知识、应用服务、模型平台、工具网关、eval 和 evidence binder 放入 transition architecture。ArchiMate 用于表达 driver、goal、requirement、capability、application service、data object、technology service 和 work package 的关系。最后通过 architecture governance 和 fitness functions 让每个 release 都有证据, 每个例外都有过期, 每个可复用能力进入平台服务目录。
11. Practice Assignment
选择一个金融机构 AI portfolio, 产出:
- AI architecture principles 8 条。
- AI capability heatmap。
- ADM phase-to-artifact map。
- ArchiMate AI element mapping table。
- Transition architecture 4 个 plateau。
- Architecture repository index。
- Architecture governance cadence。
- 3 条面试叙事。
完成标准:
- 每个 use case 都连接到 capability 和 value stream。
- 每个高风险 use case 都有 eval/control/evidence。
- 每个重复能力都有 platformization 判断。
- 每个 transition plateau 有 exit criteria。