AI Capability-Based Planning / Business Architecture / Capability Map Playbook

定位: 面向 AI BA / AI PM / Enterprise Architect / AI Transformation Lead 的高级业务架构与能力规划手册。 目标: 把企业 AI 转型从 use case list 升级为 capability portfolio, value stream, business architecture, architecture roadmap 和 funding gate。 核心观点: 企业 AI 成熟度不取决于试点数量, 而取决于哪些可复用业务能力被建设、治理、度量和持续演进。 适用范围: 金融零售企业 AI 转型, 包括 AML, 客服, 信贷, 财富/分行, 企业 AI 平台与 AI operating model。

---

Source Anchors

这些来源作为学习锚点, 不构成法律、合规、监管、采购或认证建议。正式项目必须由 legal, compliance, risk, security, privacy, architecture board, data owner 和 business owner 审查。

---

1. One-Sentence Positioning

AI capability-based planning 是把 AI 投资从“做一批场景”升级为“建设一组可复用、可治理、可度量、可融资的企业业务能力”, 并通过 value stream, capability map, maturity model, architecture runway 和 funding gate 管理从战略到落地的全过程。

---

2. 为什么 CBAP 后需要能力规划升级

你已经是 CBAP, 下一阶段的重点不是再证明会写需求、画流程、做 stakeholder analysis, 而是把 BA 能力上升到企业 AI 转型设计。

AI 项目失败很少是因为“不会写 user story”。更常见的原因是:

• 业务部门把 AI 当作 use case collection, 没有企业级 capability thesis。
• 不同团队重复做知识库、RAG、prompt、模型接入、eval 和审计日志。
• PoC 可以演示, 但无法进入受监管生产流程。
• 数据、流程、风控、架构、运营和资金节奏没有被放在同一张图上。
• AI 预算按项目切碎, 复用能力没人投资, 平台能力又被做成不落地的“大平台”。
• 用单点 ROI 评价 AI, 但忽略 capability reuse, risk reduction, process redesign 和 workforce adoption。

CBAP 之后的升级方向:

关键变化:

• 从“这个场景能不能做 AI”转成“这个企业能力是否应该被 AI 增强”。
• 从“哪个部门提出需求”转成“哪个 capability owner 对结果、风险和预算负责”。
• 从“上线一个助手”转成“沉淀可复用的 knowledge, model, eval, workflow, control 和 adoption 能力”。
• 从“一次性项目验收”转成“能力成熟度、风险证据、业务结果和架构演进的季度治理”。

---

3. 从 Use Case List 到 Capability Portfolio

3.1 典型错误路径

业务部门提 30 个 AI 场景
  -> 按热度做 PoC
  -> 采购多个工具
  -> 每个团队各建知识库和 prompt
  -> 没有统一 eval, audit, access, cost, ownership
  -> PoC 数量很多, 生产价值很少

3.2 推荐路径

Enterprise strategy
  -> AI transformation thesis
  -> Priority value streams
  -> Capability map and heatmap
  -> Capability maturity gaps
  -> Capability portfolio
  -> Architecture runway
  -> Funding gates
  -> Pilot, production, scale
  -> Quarterly capability review

3.3 五个核心问题

3.4 Capability Portfolio 的最小结构

---

4. AI Capability Taxonomy

AI capability taxonomy 不应按模型供应商或算法名称组织, 而应按企业能力、业务结果和复用边界组织。

4.1 L0 Capability Domains

4.2 L1 / L2 Capability Map

4.3 Capability Map Design Rules

• Capability 用稳定的名词短语表达, 不用项目名、产品名或 vendor 名。
• Capability 描述“组织能做什么”, 不描述“某个系统如何实现”。
• AI capability 必须绑定 business outcome, risk concern, owner, metric 和 architecture dependency。
• 区分 business capability 和 enabling platform capability, 但不要割裂两者。
• 每个 capability 至少要能回答: owner 是谁, 当前成熟度几级, 目标成熟度几级, 资金来自哪里, 复用到哪些 value streams。
• 不把“Chatbot”“RAG”“Agent”直接当作最高层能力, 它们通常是实现模式或平台能力。

---

5. Value Stream to Capability Mapping

Value stream 说明价值如何被交付, capability map 说明组织需要哪些能力才能稳定交付该价值。AI 转型要把两者连接起来, 否则会出现“流程痛点很多, 平台能力很多, 但投资无法排序”的问题。

5.1 Mapping Method

1. 选择战略级 value stream, 例如 Resolve AML alert, Originate personal loan, Serve retail banking customer。
2. 标出 value stream stages, 包括客户、员工、风险、合规和系统交互。
3. 为每个 stage 记录 pain metric, control point 和 decision point。
4. 映射所需 business capabilities, shared AI capabilities 和 control capabilities。
5. 标出 capability maturity gap, architecture dependency 和 funding gate。
6. 把 use cases 合并成 capability increments, 形成 roadmap。

5.2 Generic Matrix

5.3 AML Value Stream Example

---

6. Capability Maturity Model

能力成熟度不是“AI 模型更强”这么简单。金融零售场景需要同时看业务结果、数据、知识、架构、风险、运营和经济性。

6.1 Six-Level Model

6.2 Maturity Assessment Dimensions

6.3 Heatmap Convention

---

7. Portfolio Prioritization

AI portfolio prioritization 要避免两个极端:

• 只看业务热度, 导致高风险低准备度场景先上。
• 只看技术可行性, 导致做出没人改变工作方式的工具。

7.1 Prioritization Scorecard

Interpretation:

• 80-100: 候选为 portfolio priority, 进入 architecture and funding gate。
• 60-79: 候选为 controlled pilot, 必须补齐 red/amber dependency。
• 40-59: 适合 discovery 或 sandbox learning, 不承诺生产。
• 0-39: 暂缓, 除非监管、事故或战略压力改变优先级。

7.2 Funding Gates

7.3 Portfolio Balancing

一个成熟 AI portfolio 至少包含四类投资:

---

8. Architecture Runway

Architecture runway 是支持未来几个 capability increments 的技术、数据、治理和运营基础。它不是一次性“大平台采购”, 也不是每个项目各自搭一套。

8.1 Runway Principles

• 只建设未来 2-3 个季度明确会用到的 shared capabilities。
• 每个 runway item 必须绑定至少两个 value streams 或一个高风险强监管场景。
• Runway backlog 由 capability gaps 驱动, 不由 vendor roadmap 驱动。
• 平台能力必须有消费方、SLO、成本模型和 owner。
• 对高风险业务, runway 必须先覆盖 access, audit, eval, incident 和 rollback。

8.2 Runway Components

8.3 Example Roadmap

---

9. 金融零售案例

9.1 AML: 从 Alert Copilot 到 Financial Crime Intelligence Capability

Capability thesis:

AML AI 不是“帮 analyst 写总结”, 而是建设 financial crime operations 的证据组织、调查推理、叙事生成、QA 反馈和控制有效性能力。

Recommended roadmap:

1. Shadow mode: AI prepares evidence pack, analyst does not rely on output for final disposition。
2. Controlled pilot: AI drafts investigation summary with citations and uncertainty flags。
3. Production: AI integrated into case workflow, all outputs reviewed and logged。
4. Scale: QA defects feed eval, scenarios expand by typology and jurisdiction。

9.2 客服: 从 Chatbot 到 Omnichannel Service Intelligence

Capability thesis:

客服 AI 的核心不是“机器人回答问题”, 而是统一知识、意图、身份、上下文、服务流程和下一步动作, 降低 cost-to-serve 同时保护客户信任。

Key design choice:

• High-risk financial advice, fee dispute, hardship, fraud and complaint scenarios should route to human or constrained guidance.
• Low-risk servicing, status inquiry, document guidance and internal agent assist can scale earlier。

9.3 信贷: 从 Policy Assistant 到 AI-Assisted Credit Lifecycle

Capability thesis:

信贷 AI 不能只做“审批建议”。更稳妥的企业能力路径是从 policy reasoning, document intelligence, underwriter assist, exception routing 和 monitoring 开始, 再逐步进入决策增强。

Practical sequence:

1. Start with document intelligence and policy assistant。
2. Add underwriter evidence pack and exception checklist。
3. Add decision support with challenger signals and override reason capture。
4. Consider constrained automation only for low-risk, well-defined decisions with strong monitoring。

9.4 财富 / 分行: 从 Advisor Copilot 到 Relationship Intelligence

Capability thesis:

财富和分行 AI 的价值不只是提升销售话术, 而是增强客户理解、合规适当性、产品知识、关系经营和一线执行质量。

Design warning:

• Advisor copilot 不能变成未审查的投资建议生成器。
• 对产品推荐、收益预期、风险等级和客户适配性必须设置明确边界。

9.5 企业 AI 平台: 从工具采购到 Shared Enterprise AI Capability

Capability thesis:

AI 平台不是“买一个 LLM 网关”或“建一个统一 RAG”。平台的价值在于让业务能力更快、更安全、更可复用地进入生产。

Platform funding rule:

• 不以“统一平台愿景”拿预算。
• 以 2-3 个高价值 capability increments 的共性依赖拿预算。
• 每个 shared component 都要证明 reuse, adoption, SLO, cost model 和 retirement rule。

---

10. Templates

10.1 AI Capability Brief

10.2 Capability Heatmap

10.3 Value Stream Capability Matrix

10.4 Maturity Assessment

10.5 Portfolio Prioritization Scorecard

10.6 Architecture Runway Backlog

10.7 Funding Gate Decision Memo

# Funding Gate Decision Memo

## Decision
Approve controlled pilot for [capability name] / Do not approve production expansion for [capability name].

## Business architecture evidence
- Strategic theme:
- Value stream:
- Capability gap:
- Capability owner:
- Current maturity:
- Target maturity:

## Architecture evidence
- Chosen pattern:
- ADR summary:
- Data and knowledge sources:
- Integration boundary:
- Audit and rollback:

## Risk and control evidence
- Risk tier:
- Human oversight:
- Eval result:
- Security/privacy controls:
- Incident runbook:

## Economics
- Baseline:
- Expected benefit:
- Cost per case or user:
- Reuse potential:
- Budget cap:

## Conditions
- Production entry condition:
- Scale condition:
- Stop condition:

10.8 Capability Owner Charter

---

11. Review Checklist

Strategy and Portfolio

• Is there a clear AI transformation thesis beyond isolated use cases?
• Are priority value streams named and tied to enterprise outcomes?
• Is the candidate mapped to a capability, not only a feature or vendor tool?
• Does the portfolio balance business bets, shared runway, risk reduction and learning options?
• Are stop, scale and refresh rules defined before funding approval?

Business Architecture

• Is the value stream mapped end to end, including controls and exceptions?
• Are capability gaps visible as heatmap evidence?
• Is each capability assigned to a real owner with budget and KPI accountability?
• Are organization, role, policy and workflow changes included?
• Are customer, employee, risk and regulatory concerns represented as architecture concerns?

Data and Knowledge

• Are source of truth, owner, quality, lineage, retention and access documented?
• Are knowledge sources versioned by effective date, jurisdiction and product?
• Are retrieved documents treated as evidence, not instructions?
• Is entitlement enforced before retrieval and generation?
• Is stale or conflicting evidence handled explicitly?

AI Quality and Eval

• Is there a golden set with realistic positive, negative and edge cases?
• Does eval include domain quality, citation quality, refusal, escalation and control behavior?
• Can eval failures block release?
• Are production feedback and incidents converted into regression cases?
• Are model, prompt, retrieval index and tool versions traceable?

Architecture and Security

• Is the architecture described through stakeholder concerns and views?
• Are model gateway, retrieval, tool, audit and fallback boundaries explicit?
• Are prompt injection, data exfiltration, excessive agency and unsafe tool use addressed?
• Are human review, approval and rollback implemented in workflow, not only policy text?
• Can the enterprise reconstruct who saw what, which evidence was used, which version produced output and who approved final action?

Operating Model and Adoption

• Is there a RACI for product, process, data, knowledge, model, eval, risk, security and operations?
• Are frontline users trained on when to trust, challenge, escalate and ignore AI output?
• Are adoption metrics linked to workflow redesign instead of login counts only?
• Are supervisors and QA reviewers calibrated?
• Does the operating cadence include quality, risk, cost and benefit review?

Funding

• Does the funding request distinguish discovery, pilot, production, scale and runway?
• Is reuse value credited to shared capabilities?
• Is cost per case, user, workflow or decision measured?
• Are vendor and platform costs visible across model, storage, retrieval, observability and support?
• Is there a retirement or exit trigger?

---

12. Anti-Patterns

---

13. 30 天训练计划

目标: 在 30 天内形成一个可展示的 AI capability-based planning 作品集包, 面向金融零售 AI 转型、企业架构和高级 AI PM/BA 面试。

Weekly practice rule:

• 每周至少把一个场景从 use case 重写为 capability。
• 每周至少写一个 funding gate decision。
• 每周至少用一个风险问题挑战自己的 architecture roadmap。
• 每周至少把一个模板填成完整样例。

---

14. 面试答案

Q1: 你如何把企业 AI 转型从 use case list 升级为 capability portfolio?

30 秒版本:

我不会从“收集 AI 场景”开始排序, 而是先看企业战略和关键 value streams, 找出限制业务结果的 capability gaps。然后把零散 use cases 合并成 capability increments, 用 maturity model, architecture runway, risk controls 和 funding gates 管理投资。这样能避免 PoC 泛滥, 也能把平台能力、风险治理和业务价值放在同一张 roadmap 上。

2 分钟版本:

我的方法是五步。第一, 明确 AI transformation thesis, 例如降低金融犯罪调查成本、提升客服一次解决率或缩短信贷周期。第二, 选择优先 value streams, 画出流程阶段、决策点、控制点和 baseline。第三, 建立 capability map, 区分业务能力、共享 AI 能力、控制能力和 adoption 能力。第四, 对 capability 做成熟度和投资优先级评估, 评分维度包括战略价值、痛点强度、复用潜力、数据准备度、风险可控性、架构适配、adoption 和经济性。第五, 用 funding gates 控制 discovery, pilot, production 和 scale, 每个 gate 都要求业务、架构、eval、风险、运营和成本证据。最终交付的不是场景清单, 而是能力组合、路线图和治理节奏。

Q2: AI capability map 和传统 capability map 有什么不同?

30 秒版本:

传统 capability map 关注组织能做什么。AI capability map 还要显式表达 AI pattern、数据/知识依赖、eval、风险控制、架构 runway、adoption 和 unit economics, 因为 AI 能力的生产稳定性取决于这些运行条件。

2 分钟版本:

我会保留 capability map 的稳定性原则, 不按系统或项目命名能力。但在 AI 场景下, 每个能力必须额外连接六类信息: 第一, 它增强的是哪个 value stream stage; 第二, 它使用什么 AI pattern, 例如 RAG, decision support, document intelligence 或 agentic workflow; 第三, 它依赖哪些数据和知识源, 这些源是否有 owner, lineage, retention 和权限; 第四, 它如何 eval, 包括 golden set, rubric 和 release gate; 第五, 它有哪些风险和控制, 例如 human oversight, audit reconstruction, prompt injection defense; 第六, 它的 adoption 和成本如何度量。这样 capability map 才能从静态业务图变成 AI 投资和架构治理工具。

Q3: 如何优先排序 AML、客服、信贷、财富和 AI 平台这些 AI 投资?

30 秒版本:

我会用 portfolio scorecard, 不只看 ROI。核心维度包括战略对齐、value stream pain、复用潜力、baseline、数据/知识准备度、风险可控性、架构适配、adoption、经济性和 time-to-learning。高风险场景即使价值大, 也必须先补足 eval、审计、人审和风险控制。

2 分钟版本:

客服知识 copilot 可能较早进入 production, 因为知识边界清晰、用户量大、复用度高, 但仍需 policy freshness 和 escalation controls。AML 价值高且监管重要, 但应从 shadow mode 和 controlled pilot 开始, 重点补 evidence lineage, audit reconstruction 和 reviewer calibration。信贷需要更谨慎, 我会先做 document intelligence 和 policy assistant, 再做 underwriter assist, 不会直接进入自动拒贷或定价。财富和分行适合从 meeting summary, approved product knowledge 和 next-best-conversation 开始, 严格控制 suitability 和投资建议边界。AI 平台投资则必须绑定这些业务能力的共性依赖, 例如 model gateway, entitlement-aware retrieval, eval release gate 和 observability, 不能脱离业务消费者单独建设。

Q4: Architecture runway 在 AI 转型中怎么定义?

30 秒版本:

AI architecture runway 是未来几个 capability increments 需要提前建设的共享技术、数据、控制和运营基础, 包括 model gateway, retrieval, knowledge governance, EvalOps, tool gateway, observability, audit 和 FinOps。它必须由业务能力缺口驱动, 不能变成泛化平台建设。

2 分钟版本:

我会从 capability roadmap 反推 runway。比如未来两个季度要落地客服 copilot、AML investigation 和信贷 policy assistant, 那么共性 runway 就包括知识 owner registry、entitlement-aware retrieval、model gateway、eval golden set、audit schema、production monitoring 和 incident runbook。对 agentic workflow, 还需要 tool permission gateway, action approval, idempotency 和 kill switch。每个 runway item 都要有消费者、owner、SLO、成本模型和完成证据。这样平台投资既不会滞后业务, 也不会变成没有消费方的大平台。

Q5: 如何设计 AI funding gate?

30 秒版本:

我会把 AI funding 分成 discovery, architecture option, controlled pilot, production, scale 和 refresh gates。每个 gate 要求不同证据, 从 value stream 和 capability owner, 到 data readiness, ADR, eval, risk controls, operating model, adoption, cost per case 和 stop rules。

2 分钟版本:

Gate 0 看战略 fit, 防止只因为热度做 AI。Gate 1 看 capability discovery, 要求 value stream map, maturity gap, baseline 和 owner。Gate 2 看 architecture option, 要求 ADR, data/knowledge readiness, risk tier, build/buy/hybrid 决策和回滚思路。Gate 3 批准 controlled pilot, 要求 golden set, pilot cohort, human review 和 runbook draft。Gate 4 才允许 production, 要求 eval report, risk/security sign-off, audit reconstruction 和 operating RACI。Gate 5 管 scale, 看 adoption, benefit, incident trend 和 cost per case。Gate 6 管 refresh or retire, 看模型、vendor、风险和经济性变化。这个机制能把 AI 从一次性项目治理成长期能力。

Q6: 你如何把 NIST AI RMF 和业务架构结合?

30 秒版本:

我会把 NIST AI RMF 的 Govern, Map, Measure, Manage 嵌入 capability lifecycle。Govern 对应 owner 和 funding gate, Map 对应 value stream 和风险语境, Measure 对应 eval 和 monitoring, Manage 对应控制、incident、release 和持续改进。

2 分钟版本:

在业务架构层, 我先定义 value stream, stakeholders, concerns, capability gaps 和 risk appetite, 这对应 Map。然后把每个 capability 纳入 portfolio governance, 明确 owner, RACI, review cadence 和 funding gates, 这对应 Govern。在 solution 和 operating 层, 我设计 golden set, rubric, production sampling, drift signals, adoption metrics 和 cost metrics, 这对应 Measure。最后, 我把 human oversight, escalation, rollback, incident response, model/prompt/index versioning 和 quarterly review 放进 operating model, 这对应 Manage。这样 AI RMF 不是合规清单, 而是 capability planning 的控制系统。

Q7: 你如何向高管解释为什么不能只做 use cases?

30 秒版本:

Use case list 可以启动讨论, 但不能管理企业转型。高管真正需要的是: 哪些能力会形成可持续优势, 哪些能力能复用, 哪些风险可控, 哪些投资应该继续、停止或扩展。Capability portfolio 能把这些问题放到同一个决策框架。

2 分钟版本:

我会用一个例子解释。客服、AML、信贷和财富都可能提出“知识助手”。如果按 use case 分别做, 会产生四套知识库、四套权限、四套 eval 和四套审计。短期看每个 PoC 都快, 长期看成本、风险和治理复杂度都上升。如果把它抽象成 enterprise evidence-grounded knowledge capability, 就可以统一建设 knowledge registry, entitlement-aware retrieval, citation, eval 和 audit, 然后按不同业务角色配置输出和控制。这样既能保留业务差异, 又能形成复用经济和治理一致性。这就是从 use cases 升级到 capability portfolio 的价值。

---

15. 作品集交付物

一个高级 AI 企业架构 / 产品战略 / 能力规划作品集包应包含以下交付物:

Recommended storyline:

The problem is not lack of AI ideas.
The problem is lack of reusable, governed, measurable AI capabilities.

I start from enterprise strategy and value streams.
I identify capability gaps and maturity.
I consolidate use cases into capability increments.
I design architecture runway and control gates.
I prioritize portfolio funding using value, readiness, risk, reuse and economics.
I prove the approach through AML, service, lending, wealth/branch and AI platform cases.

---

16. Practical Operating Cadence

Quarterly review questions:

• Which capabilities moved maturity level?
• Which capabilities created reusable assets?
• Which value streams show measurable improvement?
• Which controls failed or required manual compensation?
• Which platform components are underused or over-centralized?
• Which use cases should be merged, stopped or reframed?
• Which funding gates need stronger evidence next quarter?

---

17. Final Mental Model

Use case thinking asks:

What AI thing can we build for this department?

Capability-based planning asks:

Which enterprise capabilities must become AI-enabled,
which value streams will improve,
which architecture runway is required,
which controls make it trustworthy,
which funding gates prove it deserves to scale,
and which owners will operate it after launch?

This is the shift from AI experimentation to AI enterprise transformation.