AI 扩展计划 / Playbooks

AI Transformation Value Office Playbook

企业 AI 转型的瓶颈通常不是缺少 use case，而是缺少把 use case 变成可投资、可治理、可停止、可规模化组合的运营机制。AI Value Office 的任务，是把分散试点、平台能力、风险证据、采用数据和收益兑现放进统一的 portfolio operating system。

379 行AI_TRANSFORMATION_VALUE_OFFICE_PLAYBOOK.md

AI Transformation Value Office / Portfolio Governance Playbook

Value Office 不是传统项目办公室的 AI 版本。它不只追踪里程碑，而是持续追问：这个 AI use case 是否还值得投，收益是否有基线和财务口径，风险是否被授权 owner 接受，用户是否真的采用，平台能力是否被复用，单位成本是否随规模下降，何时停止、继续、规模化或平台化。

1. Positioning and Adjacent Assets

连接文档	已解决的问题	本文如何承接
`docs/AI_EXECUTIVE_COMMUNICATION_MEMO_PACK.md`	单个 use case 如何向高管讲清 decision、ROI、risk、stop rule	将多个 memo 汇总为 executive dashboard、funding gate、scale/stop decision
AI 平台产品手册	model gateway、RAG、eval、cost、audit 等共享平台能力	判断哪些项目应复用平台，哪些能力值得平台化，平台成本如何分摊
`docs/AI_CAPABILITY_ASSESSMENT_RUBRIC.md`	系统分析、产品治理和架构判断能力	增加 portfolio backlog、benefits register、dependency map、decision log
`docs/AI_BOARD_AUDIT_COMMITTEE_GOVERNANCE_PACK.md`	董事会和审计委员会如何监督 material AI systems	将 oversight 落到 intake、risk gate、benefits proof、kill/scale governance

Value Office 的系统定义：

AI Value Office =
enterprise AI use case portfolio
+ funding gates
+ benefits realization
+ risk gates
+ platform reuse
+ change adoption
+ kill / scale decisions

2. Value Office 的职责边界

职责	关键问题	典型产物
Use case portfolio	哪些 AI 机会进入组合，哪些只是想法	AI portfolio kanban、intake form、portfolio backlog
Funding gate	每个阶段批准多少钱、多久、谁负责	funding memo、stage gate decision
Benefits realization	收益如何定义、测量、归因、签字	benefits register、baseline model、finance sign-off
Risk gate	哪些 use case 需要更强控制或委员会批准	risk tier、control evidence、release gate
Platform reuse	哪些能力复用，哪些经验沉淀为平台	reuse map、platform capability backlog
Change adoption	业务流程是否真的改变	adoption dashboard、manager cadence
Kill / scale decision	停止、继续、规模化、平台化的证据是什么	scale/stop memo、decision log

Value Office 不替代业务、风险或平台 owner：

不做	原因	正确边界
不替业务 owner 承诺收益	收益必须由流程 owner 承担	提供 measurement discipline 和 challenge
不替 risk owner 接受风险	风险接受必须在治理授权内完成	确保 evidence、gate、residual risk 清楚
不把所有试点推进成产品	AI 试点天然高不确定	建立停止权和资源再分配机制
不把平台当万能解法	平台化需要复用证据	先看 reuse potential、cost-to-serve、control leverage
不用模型指标替代业务指标	accuracy 不能自动变成收益	连接 workflow、adoption、quality、unit economics

3. Operating Model

3.1 Lifecycle

Stage	目标	入口条件	核心动作	出口证据	典型决策
Intake	收集 AI 机会并去重	有业务 owner、流程、问题陈述	记录 pain point、scope、data source、risk signal	完整 intake card	reject、merge、score
Prioritization	排出组合优先级	intake card 完成	scoring、capacity check、strategic fit review	ranked backlog	fund discovery、park、reject
Discovery	证明问题值得做	sponsor 承诺参与	workflow、baseline、data readiness、risk tier、no-AI option	discovery memo	fund pilot、stop、process redesign
Pilot	证明可行且可控	pilot funding + gate criteria	eval set、architecture spike、workflow trial、control design、adoption trial	pilot report	continue、release candidate、stop
Release	生产上线并受控运行	quality/risk/adoption/cost gate 通过	runbook、monitoring、incident path、training、release gate	release pack	limited release、delay、stop
Scale	扩展到更多用户或流程	limited release 证明收益和风险可控	scale economics、change management、support model、platform reuse	scale memo	scale、platformize、cap expansion
Retire	停止低价值或高风险能力	stop rule 或收益衰减	migration、data retention、vendor exit、lessons learned	retirement record	retire、replace、merge

3.2 Governance Cadence

Forum	节奏	参与者	决策材料	输出
Intake triage	weekly	portfolio lead、process analysis、platform、risk liaison	intake cards、duplicate map	reject / merge / score
Portfolio prioritization	biweekly	business sponsors、finance、risk、architecture、platform	scoring matrix、capacity view	funded discovery list
Pilot gate	4-6 weeks	business owner、portfolio lead、process analyst、architecture、data owner、risk、finance	discovery memo、baseline、risk tier、pilot budget	pilot approve / stop
Release gate	before production	AI governance、ops、platform、risk、business owner	eval report、control evidence、runbook、rollback	release / delay / reject
Monthly value review	monthly	CFO delegate、COO delegate、technology delegate、Value Office	dashboard、benefits register、cost report	scale / stop / reallocate
Quarterly portfolio review	quarterly	executive sponsors、risk committee liaison、platform lead	portfolio health、material risk、investment options	strategic rebalancing

3.3 RACI

Activity	Business owner	Portfolio lead	Process analysis	Architecture owner	Data owner	Risk	Finance	Platform	Ops
Intake problem statement	A	R	R	C	C	I	I	I	C
Portfolio scoring	A	R	R	C	C	C	C	C	I
Baseline and value hypothesis	A	R	R	I	C	I	C	I	C
Data readiness assessment	C	C	R	C	A	C	I	C	I
Architecture option and reuse map	C	C	I	A/R	C	C	I	C	C
Risk tier and controls	C	C	R	C	C	A/R	I	C	C
Funding memo	A	R	C	C	C	C	A/R	C	C
Pilot delivery	A	R	R	R	R	C	I	R	C
Eval and release evidence	A	R	R	R	C	C	I	R	C
Adoption plan	A	R	R	I	I	I	I	I	R
Benefits sign-off	A	R	C	I	I	I	A/R	I	C
Scale/stop decision	A	R	C	C	C	C	A/R	C	C
Retirement and lessons learned	A	R	R	C	C	C	C	C	R

4. Portfolio Scoring

Scoring 不是伪精确，而是让组合决策透明。每个 use case 都要把价值、可行性、数据、风险、复用、采用压力和单位经济放在同一张表里比较。

Dimension	问什么	评分证据	常见误判
Strategic fit	是否支持企业战略、监管承诺、核心能力	strategy map、OKR、board priority	把热门技术当战略
Value size	经济价值或风险价值有多大	volume、cost、loss、revenue、risk exposure	只算人工节省
Feasibility	技术和流程是否可交付	architecture spike、workflow complexity、integration map	demo 可跑就算可行
Data readiness	数据是否可用、可信、授权、可更新	source owner、quality、freshness、access	有数据仓库就算 ready
Risk tier	对客户、资金、合规、隐私影响多大	autonomy level、materiality、reversibility	human review 被当万能降风险
Reuse potential	是否能复用或沉淀平台能力	shared pattern、common data、tool/API reuse	过早抽象成平台
Time-to-value	多快能证明价值	baseline availability、pilot scope、user access	把上线时间当价值时间
Cost-to-serve	规模化单位成本和支持成本	token、license、infra、support、review、QA	只看模型调用成本
Adoption burden	需要多大行为改变	process change、training、manager incentives	完成培训就算 adoption
Regulatory sensitivity	是否触及监管、审计、模型风险要求	jurisdiction、policy mapping、customer impact	内部工具就认为无监管风险

Portfolio quadrant:

Quadrant	特征	投资策略
Scale now	高价值、高可行、证据强、风险可控	优先资金、扩大用户、沉淀 reusable pattern
Controlled bet	高价值，但风险、数据或采用约束强	小额分阶段 funding，强 gate，明确 stop rule
Platform candidate	单个收益中等，复用潜力高	与平台 backlog 绑定，用多个 use case 验证
Stop or park	价值弱、证据弱、约束强	停止、合并、回到流程优化或等待数据成熟

金融零售组合示例：

Use case	Value	Constraint	Portfolio view
AML/KYC investigator copilot	高运营价值和风险价值	高监管敏感、数据和复核复杂	Controlled bet
Customer service policy RAG	高频流程，平台复用强	中等客户沟通风险	Scale now / platform pattern
Credit policy assistant	改善材料质量和一致性	高公平借贷和决策边界	Controlled bet
Payment dispute triage	可改善处理时效	工具动作和账务影响需要控制	Scale after control proof
Wealth advisor compliance guardrail	降低话术和适当性风险	高合规敏感和采用压力	Controlled bet
AI platform capabilities	多 use case 复用	需防止大而全平台	Platform candidate

5. Benefits Realization

AI 收益必须经过完整证据链：

Baseline
-> Target
-> Pilot evidence
-> Adoption proof
-> Quality proof
-> Risk proof
-> Unit economics
-> Finance sign-off
-> Scale decision

Baseline 是起点，不是事后补表。

Baseline type	指标示例	数据来源	Owner
Volume	每月 case、alert、call、claim、transaction 数	workflow system、CRM、case management	Business owner
Cycle time	AHT、time-to-summary、time-to-decision、queue age	ops dashboard、process mining	Ops
Cost	labor cost、vendor cost、rework cost、infra cost	finance model、capacity plan	Finance
Quality	error rate、QA fail、complaint、reopen、appeal overturn	QA system、complaint system	Business / Risk
Risk	false negative、policy breach、audit finding、control gap	risk system、audit finding	Risk
Experience	customer and employee friction	survey、voice of customer、employee survey	Business owner

Target 必须同时包含 value、quality、risk、adoption、cost：

Target class	示例
Value target	可辅助 case 的平均处理时间从 13.8 分钟降到 10.5 分钟
Quality target	unsupported claim rate 低于 3%，source citation completeness 高于 98%
Risk target	high-risk topic 必须 100% supervisor approval，stale policy hit 低于 1%
Adoption target	目标用户中 70% 每周真实使用并产出工作记录
Cost target	all-in cost per assisted case 低于预算，p95 latency 在工作流 SLA 内

Net value:

net value per unit =
  gross benefit per unit
- model/token cost
- platform cost allocation
- license cost
- human review cost
- QA/control cost
- support and change cost
- incident/risk reserve where material

6. Adoption Proof and Value Leakage

Adoption proof 不是登录次数。它证明用户行为和管理节奏改变了。

Evidence	Strong example	Weak example
Usage depth	目标用户每周处理足够真实 case	完成培训
Workflow integration	AI 建议进入 case record，override reason 结构化	打开独立聊天窗口
Manager cadence	主管每周 review 低信任、override、异常案例	上线邮件告知
SOP change	新 SOP 明确哪些建议可用、哪些必须升级	用户自行判断
Behavior delta	查找知识时间下降，返工减少，工单字段更完整	用户说“感觉不错”

Value leakage 是 gross benefit 被运行现实吞掉的部分：

Leakage	Example
human review load	高风险输出需要大量复核
rework	AI 草稿错误导致二次修正
support burden	用户不断询问边界和错误处理
control overhead	审批、抽样和日志成本上升
latency	工作流等待模型或工具
customer harm adjustment	投诉、重复联系、错误承诺
adoption decay	新鲜感后使用下降

Scale 前必须证明净价值仍然成立。

7. Platform Reuse and Dependency Management

平台化不能靠愿景驱动，应由 portfolio evidence 驱动。

Reusable capability	Platform signal
model gateway	多个 use case 需要路由、quota、provider abstraction、audit
RAG ingestion	多个 use case 使用同类政策、产品、流程文档
eval harness	多个高风险 use case 需要一致 release evidence
prompt registry	prompt 变更需要审批、版本、回滚
audit and trace	高风险流程都需要 trace drilldown
cost dashboard	多业务线需要 showback or chargeback
tool gateway	agent 工具动作需要统一权限、审批和日志

Dependency map 应记录：

Dependency	Questions
Data	哪些源系统、owner、质量和权限会阻塞？
Policy	哪些政策版本和解释会影响输出？
Architecture	哪些共享服务、网关、日志和工具必须就绪？
Operations	哪些团队需要培训、支持和复核能力？
Risk	哪些控制证据或审批会阻断发布？
Finance	哪些收益口径需要签字？

8. Dashboard and Decision Log

Monthly executive dashboard 应把组合从“项目列表”升级成“决策系统”。

Section	Metrics
Portfolio flow	active use cases by stage、cycle time、blocked items
Value	forecast benefit、validated benefit、confidence
Adoption	qualified adoption、workflow penetration、manager cadence
Quality	release gate pass rate、unsupported claim、regression
Risk	high-risk systems、open control gaps、incidents、risk acceptance due
Cost	monthly run-rate、cost per unit、platform allocation、budget variance
Reuse	platform capability adoption、duplicate build avoided
Decisions	scale、stop、funding、risk acceptance、platformization requests

Decision log fields:

Field	Content
Date	决策发生日期
Forum	intake triage、pilot gate、release gate、value review、risk committee
Decision	fund、continue、release、scale、stop、retire、platformize
Evidence considered	scoring、baseline、eval、risk、cost、adoption
Decision owner	accountable person
Conditions	附带限制或下一 gate
Revisit date	何时复盘
Lessons	对 portfolio rule 或平台能力的影响

9. Thirty-Day Portfolio Operating Pack

30 天内可以完成一个受控的 AI portfolio operating pack，用于训练组合治理和价值证据。

Day range	Work	Artifact
1-5	定义企业背景、Value Office charter、intake 字段，收集并清理 20 个 use case	context、charter、clean backlog
6-10	写 problem statements，建立 scoring matrix，完成打分和四象限	scoring matrix、quadrant map
11-15	为优先 use cases 建 baseline、target、data readiness、risk tier	baseline workbook、risk register
16-20	设计 pilot gate、benefits register、quality proof、adoption proof、unit economics	pilot gate pack、benefits register
21-24	识别平台复用能力、dependency map、executive dashboard	reuse map、dependency map、dashboard
25-27	写 stop memo、scale memo、platformization memo	decision memos
28-30	映射治理框架，准备 executive、operator、technology narratives	operating pack

Pack structure:

AI Portfolio Operating Pack
1. Executive summary
2. Value Office charter
3. Portfolio backlog and scoring
4. Financial retail use case map
5. Benefits realization register
6. Risk and governance mapping
7. Reuse and dependency map
8. Monthly executive dashboard
9. Scale / stop decision memos
10. Decision log

10. Core Evidence Structures

10.1 AI Portfolio Kanban

Stage	Entry rule	Exit evidence	Example card
Intake	Business owner、workflow、pain、rough data source identified	intake card complete	customer service policy RAG
Prioritization	scoring completed	ranked and capacity checked	fraud alert prioritization
Discovery	discovery funding and owner confirmed	baseline、target、data readiness、risk tier	AML investigator copilot
Pilot	pilot gate and stop rule approved	quality、adoption、cost、risk evidence	payment dispute triage
Release	release gate passed	limited production release and monitoring	service RAG for policy topics
Scale	benefits signed off and support model ready	expanded users, process or region	inventory demand insight
Retire	stop trigger or replacement approved	retirement record and lessons learned	legacy chatbot

10.2 Benefits Register Entry

Field	Example
Benefit ID	BR-CSR-RAG-001
Use case	Customer service policy RAG
Benefit type	productivity capacity + quality improvement
Baseline	monthly policy cases, handling time, QA fail
Target	assisted case handling time, QA threshold
Leading indicators	weekly active agents、suggestion acceptance、citation click、override reason
Lagging indicators	handling time、QA fail、complaint escalation、rework hours
Guardrails	unsupported claim、stale source hit、supervisor approval
Cost model	token + RAG + platform allocation + supervisor review + support
Finance view	productivity capacity counted after staffing plan or backlog reduction is confirmed
Confidence	low / medium / high, with evidence date
Sign-off	business owner and finance reviewer

10.3 Scale / Stop Decision Memo Structure

Section	Evidence expectation
Decision requested	明确是 scale、continue pilot、restrict、stop、retire 还是 platformize
Current stage	说明处于 discovery、pilot、limited release 或 scale 哪一阶段
Evidence summary	同时呈现 baseline、target、eval result、adoption、cost、risk、incident 和 control evidence
Recommendation	给出决策、理由、范围和不建议选项
Conditions	定义 monitoring、support、source freshness、owner 和 review cadence
Stop / rollback trigger	用质量、风险、采用、成本或 incident 阈值定义停止条件
Next review	说明复盘日期、论坛和需要补齐的证据

11. Governance Anchors

Source	可借鉴点	映射
NIST AI Risk Management Framework	Govern、Map、Measure、Manage	risk gate、control evidence、portfolio dashboard、decision log
NIST AI RMF Generative AI Profile	GenAI 风险画像和治理实践	use case risk tier、eval、red-team、guardrails
ISO/IEC 42001	AI management system 的组织化和持续改进	charter、RACI、stage gate、continual improvement
OMB M-24-10 briefing	AI governance、innovation、risk management	inventory、minimum practices、senior accountability

使用框架时要避免机械套用。真正的能力体现在 artifact：intake、scoring、benefits register、risk gate、dashboard、decision log。

12. Operating Principle

Value Office 的价值，是把企业 AI 从分散试点升级成可投资、可治理、可停止、可复用、可规模化的组合能力。

Invest only where there is evidence.
Scale only where value survives adoption, quality, risk and unit economics.
Stop quickly when evidence fails.
Platformize only when reuse is proven.