AI Value Stream Management / Flow Metrics Playbook

适用对象: AI PM、AI Product Operations Lead、AI Portfolio Lead、AI Product Architect、Enterprise Architect、Value Office Lead、金融零售 AI 转型负责人。 目标: 把 AI use case 从 idea 到 safe release 到 adoption / value realization 的流动管起来, 并把 DORA / SPACE、Flow Metrics、risk gates、platform runway、portfolio funding 和 business value 连接成一套可运营的价值流管理系统。 核心观点: AI Value Stream Management 不是画流程图, 而是持续管理 "价值流动、风险证据、平台能力、组织容量和收益兑现"。 边界说明: 本文是学习、架构设计和作品集材料, 不构成法律、监管、审计、模型验证、财务确认或供应商选型意见。金融零售正式项目必须由 business owner、risk、model risk、legal、compliance、privacy、security、finance、architecture、operations 和 internal audit 共同确认适用要求。

---

1. Source Anchors

这些来源作为术语和方法锚点。访问日期按 2026-06-29 记录。

Source-to-artifact mapping:

---

2. One-Sentence Positioning

AI Value Stream Management 是把 AI opportunity、delivery work、risk evidence、platform capability、adoption behavior 和 realized business value 放进同一条可度量价值流, 让组织能决定 fund、accelerate、block、scale、hold、stop 或 retire。

更短的面试版:

I manage AI use cases as value streams, not feature queues:
from idea evidence to safe release, adoption, risk-controlled operation and finance-recognized value.

管理层中文表达:

我不会用 "上线了多少 AI 功能" 或 "调用了多少次模型" 判断 AI 成功。我会看 AI 是否穿过了完整价值流: 业务问题是否真实, 数据和风险证据是否成立, 平台能力是否复用, 上线是否安全, 用户是否采用, 业务收益是否被 finance 和业务 owner 认可, 风险是否在 scale 后仍然稳定。

---

3. 目的 / 适用对象 / 核心观点

3.1 目的

本文解决五个高级问题:

1. AI use case 如何从 idea funnel 流到 production, 而不是停在 PoC 或 dashboard。
2. 如何用 Flow Metrics 找出 AI delivery 和 AI-enabled operations 中的等待、返工、风险阻塞和收益断点。
3. 如何把 DORA / SPACE 的工程生产力语言, 与 portfolio funding、platform runway、risk gates 和 business outcomes 连接。
4. 如何证明 AI 价值不是 activity、usage、demo 或模型分数, 而是受控风险下的 adoption 和 benefits realization。
5. 如何把金融零售中的 AML、客服、信贷、分行、投诉、风控等流程转成可投资、可治理、可度量的 AI value streams。

3.2 适用对象

3.3 核心观点

成熟的 AI VSM 不是 "流程可视化", 而是四个系统合一:

---

4. 为什么 AI 成功不能只看功能 / 调用量

AI 产品常见汇报方式:

• 上线了 12 个 AI features。
• LLM API calls 达到 200 万次。
• 生成了 50 万段摘要。
• 模型准确率达到 92%。
• 用户满意度问卷达到 4.6/5。
• 预计节省 30% 人工时间。

这些都是 signals, 但不是完整 value evidence。

成熟 AI 价值叙事:

Model capability
-> AI behavior evidence
-> workflow adoption
-> qualified value event
-> risk-adjusted business outcome
-> finance-recognized benefit
-> portfolio scale / stop decision

一句话:

AI value stream 的主语不是模型, 而是受 AI 改变的业务流程和决策流。

4.1 Vanity Metrics 到 Flow Metrics 的迁移

4.2 AI Value Event 的定义

AI value event 不是 "AI 被调用", 而是一个带质量、风险、采用和经济性约束的业务事件。

Qualified AI value event =
  eligible workflow item
  + AI exposure or AI-assisted action
  + user / system adoption signal
  + quality threshold pass
  + risk guardrail pass
  + auditable trace
  + unit economics within boundary

金融零售示例:

---

5. Value Stream Map: AI Delivery

AI delivery value stream 的目标是把 AI idea 从 "想法" 管到 "可控生产价值", 而不是管任务清单。

5.1 End-to-End AI Delivery Flow

Strategic theme
  -> Idea intake
  -> Discovery evidence
  -> Value stream design
  -> Data / eval readiness
  -> Build and integration
  -> Risk and release gate
  -> Limited release
  -> Adoption and operating model
  -> Benefits realization
  -> Scale / hold / stop / retire

5.2 Stage-by-Stage Map

5.3 Current-State vs Target-State AI VSM

Current-state map 要暴露真实阻塞:

Target-state map 要定义三个改进方向:

1. Reduce waiting: data access、risk review、SME review、eval queue、release approval。
2. Reduce rework: unclear problem、weak eval、ambiguous decision boundary、poor telemetry。
3. Increase value throughput: qualified value events、benefit realization、platform reuse。

---

6. Value Stream Map: AI-Enabled Operations

AI-enabled operations 是 AI 上线后的业务运营价值流。它回答: AI 是否真正改变了日常流程, 并在真实风险边界内创造价值。

6.1 Generic AI-Enabled Operations Flow

Business event / case
  -> eligibility check
  -> context and knowledge retrieval
  -> AI draft / recommendation / action proposal
  -> human judgment or policy automation
  -> workflow action
  -> QA / sampling / exception handling
  -> customer / operations outcome
  -> monitoring and incident response
  -> benefits and learning feedback

6.2 Operations Flow Metrics

6.3 Delivery Flow vs Operations Flow

高级判断:

Delivery VSM 证明 AI 能安全上线; Operations VSM 证明 AI 在真实业务流程中持续创造价值。

---

7. Flow Metrics and AI-Specific Extensions

Flow Metrics 的价值是把价值流看成一个系统, 而不是看单个团队产出。AI 场景需要扩展到 data、eval、risk、platform、adoption 和 benefits realization。

7.1 Core Flow Metrics

7.2 AI-Specific Flow Metrics

7.3 DORA Extensions for AI Flow

7.4 SPACE Signals for AI VSM

SPACE 不应该用来给个人排名。它用于解释 AI value stream 中的人和协作是否健康。

7.5 Flow Distribution for AI Portfolio

AI portfolio 如果只塞满 features, 生产会越来越慢。建议把 work item 分成六类:

---

8. Connecting Flow, DORA / SPACE, Risk Gates and Business Value

成熟 AI VSM 的 dashboard 不应该是一个单层报表, 而是一个多层 operating system。

Portfolio thesis
  -> value stream and process baseline
  -> delivery flow metrics
  -> DORA / SPACE signals
  -> risk gate evidence
  -> operational adoption metrics
  -> qualified value events
  -> benefits realization
  -> portfolio funding and scale / stop decisions

8.1 Metric Stack

8.2 Operating Review Cadence

8.3 Metric Tension Pairs

AI VSM 必须同时看正向指标和反向指标, 防止局部优化。

---

9. Portfolio-to-Platform-to-Product Trace

AI VSM 的高级能力是把一个 use case 从 portfolio thesis 一直追到 product telemetry 和 benefit ledger。

9.1 Trace Chain

Portfolio theme
  -> business capability / APQC process
  -> end-to-end value stream
  -> AI use case hypothesis
  -> platform capability dependency
  -> product workflow intervention
  -> release gate evidence
  -> adoption telemetry
  -> qualified value event
  -> benefits realization
  -> scale / stop decision

9.2 Trace Table

9.3 Platform Runway Link

AI VSM should show whether a use case creates platform leverage.

9.4 Portfolio Funding Implication

The funding question changes from:

Should we fund this AI feature?

to:

Should we fund this value stream improvement,
including product workflow, platform runway, risk assurance,
adoption support and benefits measurement?

Funding buckets:

---

10. Risk Gates and Business Value Gates

AI VSM treats gates as flow control points. A good gate can block, limit, route, accelerate or retire work based on evidence.

10.1 Gate Architecture

10.2 Gate Principles

1. Gate evidence should be created during work, not reconstructed for meetings.
2. High-risk use cases need earlier gates, not only stricter release gates.
3. Gate queue time is a flow metric. If it grows, the operating system is under-designed.
4. A gate without a stop or limit decision is just paperwork.
5. A gate should distinguish "not ready yet" from "not worth doing" from "convert this into platform or data investment"。

10.3 AI Risk Gate Checklist

---

11. Financial Retail Case: Regulated Operations Intelligence

11.1 Context

A financial retail institution wants to use AI across AML operations, customer servicing and branch knowledge. The executive goal is not "launch more AI", but:

Reduce regulated operations workload,
improve decision quality,
shorten customer and employee waiting time,
and build reusable AI platform controls
without increasing compliance, privacy, fairness or operational risk.

11.2 Portfolio Theme and Value Streams

11.3 Current-State Observations

11.4 Target Value Stream: AML Alert Triage Assistant

Alert created
  -> eligibility check
  -> retrieve case history, transaction context and policy guidance
  -> AI produces cited summary and investigation checklist
  -> analyst reviews, edits and decides next action
  -> QA samples output and decision evidence
  -> case closed or escalated
  -> telemetry updates eval set, risk controls and benefits register

11.5 Metrics for the AML Stream

11.6 90-Day Decision Example

Decision:

Hold broad scale.
Scale only to a second low-risk queue after human review workflow is improved.
Fund platform runway for reusable SME sampling, evidence binder and eval case management.
Do not expand to high-risk typologies until critical failure taxonomy and review capacity are stronger.

11.7 Portfolio Flow Snapshot

11.8 What This Case Shows in an Interview

---

12. Templates

12.1 AI Value Stream Canvas

# AI Value Stream Canvas

Value stream name: AML alert triage to investigation closure
Portfolio theme: Regulated operations intelligence
Business capability: Financial crime operations
Business owner: Head of AML Operations
Product owner: AI Operations Product Lead
Risk owner: Financial Crime Risk
Platform owner: AI Platform Lead
Review date: 2026-06-29

## 1. Business outcome
Reduce low-risk alert handling time and backlog age while maintaining investigation quality and auditability.

## 2. Current-state flow
Alert created -> analyst searches customer and transaction context -> analyst checks policy and typology notes -> analyst drafts case narrative -> QA sample -> closure or escalation.

## 3. Current bottlenecks
- Context assembly consumes analyst time.
- Policy search is inconsistent across teams.
- QA defects often relate to missing evidence or weak narrative support.
- Review capacity is limited for pilot expansion.

## 4. AI intervention
AI retrieves approved case context and policy guidance, drafts a cited summary, and recommends an investigation checklist.
AI does not close alerts, file SARs, downgrade risk, or make final compliance decisions.

## 5. Target-state flow
Alert created -> eligibility check -> approved context retrieval -> AI cited summary and checklist -> analyst review and decision -> QA sample -> telemetry feedback to eval and benefits register.

## 6. Flow metrics
| Metric | Baseline | Target / decision rule |
|---|---:|---|
| Idea-to-evidence lead time | 18 business days | <= 15 business days for similar future use cases |
| Release-to-adoption lead time | 6 weeks | repeat adoption by >= 65% eligible analysts within 6 weeks |
| Qualified value throughput | 0 | quality-approved assisted alerts per week after limited release |
| SME review queue age | 1 day | must not exceed 3 days during pilot |

## 7. Risk gates
| Gate | Evidence |
|---|---|
| Eval readiness | 500 historical case golden set, critical failure taxonomy, citation rubric |
| Release | rollback path, trace logging, human approval, QA sample plan |
| Scale | no critical evidence defect, stable QA, review capacity within threshold |

## 8. Benefits realization
Recognize benefit only for eligible low-risk alerts where AI was exposed, analyst accepted or materially used the output, QA passed, and case handling time improved versus baseline or control.

## 9. Scale / stop rule
Scale to another low-risk queue if AHT improves >= 10%, QA critical defects remain 0, citation pass >= 95%, review queue age <= 3 days, and cost per case stays within approved ceiling.
Stop expansion if any critical privacy, evidence or regulated decision boundary breach occurs.

12.2 Flow Metrics Dashboard

# AI Flow Metrics Dashboard

Audience: AI portfolio review, product operations, platform leadership, risk governance, Value Office
Cadence: weekly for flow, monthly for benefits, quarterly for portfolio funding

## Executive summary
| Decision area | Metric | Current signal | Decision |
|---|---|---|---|
| Portfolio WIP | active use cases by stage and risk tier | Pilot WIP exceeds review capacity | freeze new high-risk pilots |
| Value flow | qualified value throughput | customer service copilot improving | prepare scale memo |
| Bottleneck | SME review queue age | AML review queue aging | fund review workflow automation |
| Risk | critical guardrail breaches | none in limited release | continue release with monitoring |
| Benefits | finance-recognized benefit | partial recognition for service queue | expand causal measurement |

## Portfolio flow
| Metric | Definition | Slice |
|---|---|---|
| Flow load | active AI work items in idea, discovery, pilot, release, scale | stage, risk tier, business capability |
| Flow distribution | percent of work in feature, platform, risk, data, adoption, debt | portfolio theme |
| Evidence conversion | percent of discovery items converted to pilot, platform investment or stop decision | business unit |
| Stop decision latency | days from stop signal to decision | owner, risk tier |

## Product value stream
| Metric | Definition | Slice |
|---|---|---|
| Idea-to-evidence lead time | intake accepted to discovery evidence ready | use case, capability |
| Release-to-adoption lead time | limited release to repeat adoption threshold | cohort, role |
| Qualified value throughput | quality and risk-passed AI value events per week | workflow, risk tier |
| Benefit recognition cycle time | production release to finance-recognized benefit | business unit |

## Engineering and platform
| Metric | Definition | Slice |
|---|---|---|
| AI change lead time | spec / eval / build / release time for AI artifacts | code, prompt, model, index, policy |
| Change fail rate | AI changes requiring rollback, hotfix or control intervention | service, risk tier |
| Platform reuse lead time | request to production use of shared platform capability | capability |
| Eval queue time | eval submitted to gate decision | use case, test suite |

## Risk and operations
| Metric | Definition | Slice |
|---|---|---|
| Evidence aging | days since evidence matched current production config | evidence type |
| Critical failure rate | no-go failures per evaluated workflow item | risk category |
| Human review load | reviewer capacity used and queue age | SME, risk, QA |
| Incident recovery time | detection to fallback / rollback / stable operation | incident type |

## Adoption and SPACE
| Metric | Definition | Slice |
|---|---|---|
| Repeat adoption | eligible users using AI in target workflow over repeated periods | role, team |
| Override reason distribution | why users reject or edit AI output | workflow step |
| Trust pulse | user trust score with free-text reason coding | cohort |
| Review fatigue signal | reviewer load, after-hours review, queue stress | reviewer group |

12.3 Blocked Work Taxonomy

12.4 Benefits Realization Loop

Baseline
  -> eligibility and exposure logging
  -> adoption and action telemetry
  -> quality and risk qualification
  -> incremental effect estimate
  -> cost and risk adjustment
  -> finance / business sign-off
  -> capacity redeployment or value capture
  -> post-scale audit
  -> portfolio scale / stop decision

Template:

# AI Benefits Realization Loop

Use case: Customer service policy copilot
Business owner: Contact Center Operations
Finance owner: FP&A partner
Risk owner: Customer Conduct Risk
Review period: 2026 Q3 month 2

## Baseline
Eligible contacts: 620,000 per month
Baseline AHT: P50 7.8 minutes
Baseline reopen rate: 11.2%
Baseline complaint rate: 0.42%
Baseline QA policy defect rate: 3.1%

## Exposure and adoption
Treatment cohort: servicing queues A and B
Eligible exposed contacts: 184,000
Accepted AI answer: 118,000
Repeat adoption: 68% of eligible agents

## Quality and guardrail
Citation QA pass: 96.5%
Critical wrong policy answer: 0
PII leakage: 0
Reopen rate: no credible increase versus control
Complaint rate: stable within approved threshold

## Incremental effect
Cluster rollout estimate: AHT reduction of 0.6 minutes per accepted exposed resolved contact.
Effect recognized only for contacts passing quality and guardrail filters.

## Cost and risk adjustment
Costs included: model, retrieval, platform support, QA sampling, training, monitoring.
Risk adjustment: high-risk intents excluded from recognition until separate gate.

## Realized benefit
Recognized benefit: capacity redeployed to peak-hour backlog and training coverage.
Finance treatment: limited-scale capacity benefit accepted for Q3 operating review.

## Next decision
Scale to two additional low-risk servicing queues.
Hold complaints, hardship and vulnerable customer intents until policy and escalation controls pass separate gate.

12.5 Portfolio Evidence Pack

# AI Portfolio Evidence Pack

Portfolio theme: Regulated operations intelligence
Quarter: 2026 Q3
Decision meeting: Quarterly AI portfolio review

## 1. Portfolio thesis
Invest in AI capabilities that reduce regulated operations workload, improve quality, reuse approved platform controls, and produce measurable value evidence within one quarter.

## 2. Stage distribution
| Stage | Items | Risk profile | Decision need |
|---|---|---|---|
| Idea | collections contact strategy | high fairness and conduct risk | park until risk pathway matures |
| Discovery | credit memo assistant | high credit policy risk | continue data readiness work |
| Pilot | AML alert triage assistant | high AML evidence risk | limited scale only if review load is controlled |
| Release | customer service policy copilot | medium conduct risk | scale low-risk intents |
| Scale | branch knowledge assistant | low / medium policy risk | expand through knowledge platform |
| Retire | legacy FAQ bot | audit and content duplication risk | migrate and shut down |

## 3. Flow metrics
| Metric | Portfolio signal | Decision |
|---|---|---|
| Flow load | high-risk pilots exceed SME capacity | freeze new high-risk pilot starts |
| Gate queue time | privacy and model risk review stable, SME review aging | fund review workflow |
| Platform reuse | RAG and trace reused by 3 use cases | continue platform runway |
| Benefit recognition | 2 of 4 production use cases have finance-recognized benefit | strengthen telemetry for others |

## 4. Risk evidence
- No critical PII leakage in production AI workflows.
- Two policy citation defects added to eval regression set.
- One tool permission exception closed before release.
- SME review capacity is the primary scale constraint.

## 5. Funding recommendation
- Continue customer service copilot scale for low-risk intents.
- Hold AML broad scale until review queue and sampling workflow improve.
- Convert repeated RAG and evidence needs into platform runway funding.
- Stop independent FAQ bot maintenance and migrate content.

12.6 Scale / Stop Memo

# AI Scale / Stop Memo

Use case: AML alert triage assistant
Current stage: Limited release
Decision requested: Limited scale to second low-risk queue
Decision date: 2026-06-29

## Value evidence
- Median handling time decreased 12% in pilot queue.
- Backlog age decreased 9% for eligible low-risk alerts.
- Analyst repeat adoption reached 72%.
- Summary acceptance reached 76% after week 4.

## Quality and risk evidence
- Critical evidence defect: 0.
- Citation correctness: 96%.
- Unsupported claim rate: below approved threshold.
- AI did not close alerts, file SARs or downgrade risk.

## Flow evidence
- Eval-to-release lead time: 9 business days.
- Risk gate queue time: stable.
- SME review queue age increased to 4 days, above target.

## Unit economics
- AI cost per case inside approved limited-release ceiling.
- QA cost increased due to manual sampling.
- Platform trace and RAG components reused by customer service copilot.

## Decision
Limited scale to one additional low-risk queue.
Condition: review queue age must return to <= 3 days before further expansion.
Fund shared review workflow and evidence binder as platform runway.
Do not expand to high-risk typologies until review capacity, critical failure taxonomy and escalation controls pass scale gate.

## Stop triggers
- Any confirmed critical privacy breach.
- Any AI-attributable alert closure or SAR escalation boundary breach.
- Two consecutive monthly reviews with no handling-time or backlog benefit.
- SME review queue age above 5 business days for two consecutive review cycles.

---

13. Review Checklists

13.1 AI VSM Design Checklist

• Is the AI use case tied to a named business capability and end-to-end process.
• Is there a current-state and target-state value stream, not only a feature description.
• Are flow metrics defined from idea to release and from release to value realization.
• Are delivery flow and operations flow separated but connected.
• Are DORA metrics adapted to code, prompt, model, data, index, policy and tool changes.
• Are SPACE signals used at team / system level, not for individual ranking.
• Are WIP limits and flow load visible by risk tier and value stream stage.
• Is blocked work classified with owner and response path.
• Are risk gates integrated into flow rather than attached at the end.
• Is benefits realization designed before scale.

13.2 Flow Metrics Review

• Does flow time include waiting, review, gate queue and adoption delay.
• Is flow efficiency exposing delay and handoff, not blaming teams.
• Is flow velocity adjusted for risk, quality and qualified value events.
• Is flow distribution balanced across feature, platform, risk, data, adoption and debt.
• Is flow predictability interpreted by risk tier and service context.
• Are stop decisions tracked as healthy portfolio outcomes.
• Are metrics paired with tension indicators to prevent gaming.

13.3 Risk and Gate Review

• Does each gate have a clear go, limited go, no-go, rollback or stop decision.
• Are privacy, security, compliance, model risk and operational risk addressed before release.
• Is AI authority explicit: search, draft, recommend, triage, decision support or automated action.
• Is human accountability visible in workflow and telemetry.
• Are eval results connected to production monitoring and incident learning.
• Are critical failures added to regression, policy or runbook updates.
• Is evidence aging tracked after model, prompt, data, index or policy changes.

13.4 Benefits Realization Review

• Is there a pre-release baseline.
• Are assignment, exposure, adoption, action, outcome and guardrail logged.
• Is value counted only for quality-passed and risk-passed events.
• Are model, platform, QA, training, monitoring, governance and risk costs included.
• Is finance treatment explicit: cost reduction, capacity redeployment, loss avoided, SLA improvement or revenue protection.
• Is post-scale audit planned to detect benefit decay.
• Is the scale decision based on realized value and risk stability, not only release success.

---

14. Anti-Patterns

---

15. 30-Day Training Plan

目标: 30 天内产出一套可放入作品集的 AI Value Stream Management / Flow Metrics 证据包。

完成标准:

• One current-state and one target-state AI value stream map.
• Flow Metrics dashboard with portfolio, product, platform, risk, adoption and benefits layers.
• DORA / SPACE integration without vanity metrics.
• Risk gates and benefits gates connected to scale / stop decisions.
• At least one financial retail case with qualified value event and evidence pack.
• Clear portfolio-to-platform-to-product trace.

---

16. Interview Answers

Q1: 你如何用 VSM 管 AI use case?

30 秒版本:

我会把 AI use case 当成 value stream item, 从 idea 到 discovery、data / eval readiness、risk gate、limited release、adoption、benefits realization 和 scale / stop 全链路管理。指标上不只看功能上线和调用量, 而是看 flow time、blocked time、gate queue、qualified value events、risk guardrail 和 finance-recognized benefits。

2 分钟版本:

我会先把 use case 挂到业务 capability 和 end-to-end process, 例如 AML alert triage 或 customer servicing。然后画 current-state 和 target-state value stream, 明确 AI 插入哪个步骤、改变什么决策、谁承担 accountability。接着定义 Flow Metrics: idea-to-evidence lead time、data readiness wait time、eval design lead time、risk gate queue time、release-to-adoption lead time、qualified value throughput 和 benefit recognition cycle time。工程侧用 DORA 看 AI artifact change 的 lead time、release frequency、change fail rate 和 recovery; 团队侧用 SPACE 看 review load、SME fatigue、trust 和 collaboration。最后用 risk gate 和 benefits gate 决定 scale、hold、stop 或转成 platform investment。

Q2: 为什么不能用 AI 调用量证明 AI 价值?

30 秒版本:

调用量只能说明 AI 被使用或被系统触发, 不能说明它产生了合格业务结果。金融零售里我更看 qualified AI value event: 目标流程合格、AI 暴露、用户采纳、质量通过、风险门槛通过、可审计、单位经济成立。

2 分钟版本:

例如客服 copilot 的调用量上升可能是因为知识库难用、agent 反复问、答案不准或用户好奇。真正的价值事件应该是: eligible contact 中 AI 给出有引用的答案, agent 采纳或合理编辑, 客户问题一次解决, 7 天内无 reopen, 没有 wrong policy answer、PII leakage 或投诉上升, 并且 cost per resolved contact 在边界内。只有这样的事件才能进入 benefits realization。否则 API calls 和 tokens consumed 只是 activity, 甚至可能是成本和风险。

Q3: Flow Metrics、DORA 和 SPACE 怎么放在一起?

30 秒版本:

Flow Metrics 看价值流是否顺畅, DORA 看交付系统是否更快更稳, SPACE 看人和协作是否健康。AI 场景要再叠加 eval、risk gate、platform reuse 和 benefits realization。

2 分钟版本:

Flow Metrics 用于 portfolio 和 product value stream: flow time、velocity、efficiency、load、distribution、predictability。DORA 用于 AI SDLC: prompt、model、index、policy、tool schema 和 code 的 change lead time、deployment frequency、change fail rate、recovery 和 rework。SPACE 用于解释为什么流动变好或变差, 例如 reviewer load、SME fatigue、context switching、trust 和 collaboration。三者共同连接到 business value: 只有当交付更快更稳、团队没有被 review 和治理压垮、上线后 adoption 和 qualified value events 改善, 才能说 AI operating system 真的有效。

Q4: 高风险金融 AI 如何做 release 和 scale gate?

30 秒版本:

高风险 AI 不能把 pilot approval 当成 scale approval。我会分 intake、discovery、eval readiness、architecture、risk、release、adoption、benefits、scale gate, 每个 gate 都有证据和明确 go / limited go / no-go / rollback / stop decision。

2 分钟版本:

以 AML alert triage 为例, intake gate 要证明 business owner、process baseline 和 risk hypothesis; discovery gate 要有 workflow map、AI fit、data readiness 和 no-AI alternative; eval gate 要有 golden set、failure taxonomy、citation rubric; risk gate 要确认 AI 不自动关闭 alert、不做 SAR 决策、有人审阅、日志可追溯; release gate 要有 rollback、monitoring 和 evidence binder; adoption gate 要看 analyst acceptance、override、support load; benefits gate 要看 AHT、backlog、QA defect 和 finance treatment。scale 只有在 value、risk、review capacity、platform capacity 和 unit economics 同时成立时才放行。

Q5: 如何向 CFO 解释 AI VSM 的价值?

30 秒版本:

我会把 AI VSM 从流程管理翻译成投资组合效率: 更快发现无效 use case, 更少 late-stage rework, 更高 qualified value throughput, 更短 benefit recognition cycle, 更清楚的风险调整收益。

2 分钟版本:

CFO 不需要看 AI 调用量, 需要看资金和容量是否流向可兑现价值。我会展示 portfolio flow: 多少 idea 转成有效 pilot, 多少被及时停止, 哪些平台能力降低后续接入成本, 哪些上线用例产生 finance-recognized benefit。收益端看 capacity redeployment、loss avoided、SLA improvement、cost per value event; 成本端包括模型、平台、QA、training、monitoring、governance 和风险成本。这样 AI 预算不是一堆项目申请, 而是一套能持续学习、放大有效投资并停止低质量投资的 operating system。

Q6: 如何证明平台 runway 不只是技术成本?

30 秒版本:

平台 runway 的价值要用 flow 和 risk economics 证明: 它是否缩短 platform reuse lead time、eval design lead time、risk gate queue time、release recovery time, 并提高 evidence reuse 和 qualified value throughput。

2 分钟版本:

Model gateway、eval harness、RAG service、observability、policy-as-code、human review workflow 和 evidence binder 如果只按技术组件汇报, CFO 和业务很难感知价值。我会把它们映射到 value stream bottleneck: gateway 减少模型审批和供应商切换成本; eval harness 减少 release gate 重工; RAG service 减少重复索引和引用错误; evidence binder 降低审计重建成本; review workflow 解决 SME capacity bottleneck。平台投资的 KPI 不是接了多少模型, 而是让更多 AI use case 更快、更稳、更可审计地通过 release 和 benefits gate。

Q7: 如果 AI 用例上线了但 adoption 很低, 你怎么处理?

30 秒版本:

我不会立刻判定模型失败。我会看 release-to-adoption lead time、eligible exposure、workflow fit、trust、manager cadence、training、support load、override reason 和 incentive mismatch, 然后决定 redesign workflow、调整 change plan、限制范围或停止。

2 分钟版本:

Adoption 低通常有多种原因: AI 不在用户自然工作流里, 输出缺少 citation, 用户不信任, manager 没有把新流程纳入日常管理, 或使用 AI 增加了 QA 负担。我会把 adoption 当作 operations value stream 的一段来测: eligible users 是否暴露, 是否首次使用, 是否 repeat use, 输出是否被采纳或编辑, 采纳后 workflow outcome 是否改善。若 adoption 低但质量好, 可能需要产品和变更管理; 若 adoption 低且 override 指向质量问题, 回到 eval 和 RAG; 若 adoption 长期低且收益不成立, 应停止或缩小。

Q8: VSM 如何帮助你做 AI portfolio governance?

30 秒版本:

VSM 让 portfolio review 从状态汇报变成资金和容量决策。我们可以看到 WIP、stage aging、risk tier、blocked work、platform dependency、benefits status 和 stop signals, 然后决定 fund、scale、hold、stop、retire 或投资平台 runway。

2 分钟版本:

AI portfolio 最大问题通常不是 idea 少, 而是太多 idea 争夺产品、工程、数据、risk、SME 和平台容量。VSM 可以显示哪些 use case 卡在 data readiness, 哪些卡在 risk gate, 哪些 release 后 adoption 不动, 哪些已经有收益但 review capacity 限制 scale。这样季度 review 不再问 "每个项目进展如何", 而是问 "哪条价值流值得加速, 哪些阻塞需要平台投资, 哪些高风险 use case 应限制, 哪些低价值 work 应停止"。这就是 portfolio-to-platform-to-product 的治理闭环。

---

17. Portfolio Package

一套高级 AI VSM 作品集可以包含:

5 分钟讲述结构:

0:00-0:40  问题定义
AI success cannot be measured by features shipped or model calls.

0:40-1:30  方法
I manage AI as value streams: idea, evidence, safe release, adoption, benefits and scale / stop.

1:30-2:30  Metrics
I combine Flow Metrics, DORA, SPACE, risk gates and qualified value events.

2:30-3:30  Financial retail case
Use AML or customer service to show delivery flow, operations flow, risk controls and benefits.

3:30-4:30  Portfolio and platform
Show how repeated blockers become platform runway and how funding decisions change.

4:30-5:00  Close
The goal is not more AI activity. The goal is faster, safer, reusable and finance-recognized AI value.

Final memory card:

AI VSM = portfolio flow + product flow + platform flow + risk flow + benefit flow.

Do not manage AI by feature count or model calls.
Manage the path from idea evidence to safe release,
from release to adoption,
from adoption to qualified value events,
from value events to benefits realization,
and from benefits to portfolio scale / stop decisions.