AI 扩展计划 / Playbooks

AI Executive Communication Memo Pack

企业 AI 沟通的核心不是展示 "我懂模型", 而是帮助决策者回答:

422 行AI_EXECUTIVE_COMMUNICATION_MEMO_PACK.md

AI Executive Communication Memo Pack

目标: 训练把 AI use case, architecture, risk, eval, ROI 用高管, CTO, CFO, 业务负责人, 风控合规, 数据负责人都听得懂的方式表达。定位: 本文件是 AI BA / AI PM / AI Architect / Enterprise Architect 的 executive communication artifact pack, 不替代已有学习计划、模板或 runbook。使用方式: 每个 AI case 都先写 1 页 memo, 再按受众切换表达重点。重点不是把技术讲浅, 而是把决策讲清。

1. Core Mindset

企业 AI 沟通的核心不是展示 "我懂模型", 而是帮助决策者回答:

What decision is needed?
Why now?
What are the options?
What evidence supports the recommendation?
What business metric will move?
What risk remains and who owns it?
What happens in the next 30 days?
When do we stop?

Executive communication 要把五类材料压缩成同一条逻辑链:

Layer	技术团队常说	高管需要听到
Use case	RAG, agent, workflow automation	哪个流程、哪类用户、哪个业务结果会改善
Architecture	vector DB, model gateway, tools, orchestration	为什么这个方案可控、可运维、可审计、可扩展
Risk	hallucination, prompt injection, model drift	哪些业务损害可能发生, controls 如何降低概率和影响
Eval	golden set, graders, thresholds	上线前和上线后如何证明它持续可靠
ROI	token cost, latency, pilot budget	基线、收益、成本、信心等级和停止条件

2. Communication Principles

Start with decision, not technology. 第一段先说要批准什么: discovery, pilot funding, vendor shortlist, architecture option, risk acceptance, scale, stop。
Use baseline and delta. 不说 "提升效率", 要说 "当前每月 12,000 个 case, 平均处理 18 分钟, pilot 目标把可自动辅助的 40% case 降到 12 分钟以内"。
Separate proven facts, assumptions, and bets. 高管可以接受不确定性, 但不能接受把假设包装成事实。
Always compare with no-AI and workflow-only options. AI 不是默认答案。强 memo 必须说明为什么不用 AI 不够, 为什么流程优化本身不够。
Translate AI quality into business risk. "groundedness 低" 要翻译成 "员工可能引用过期政策, 导致错误承诺、投诉升级或审计缺口"。
Make controls concrete. 不说 "有人审核", 要说 "所有高风险建议必须由授权主管审批, audit log 记录 source, prompt version, model version, reviewer, final action"。
Do not use accuracy alone. AI eval 至少包括 task success, groundedness, red-flag recall, unsafe output rate, override rate, latency, cost per case。
Explain architecture as a risk and operating decision. RAG, fine-tuning, agent, vendor product 的取舍要落到 data boundary, permissions, latency, audit, cost, rollback。
Tie ROI to adoption. 没有真实使用和流程改变, 模型质量不能自动变成业务收益。
Name the owner. 每个 memo 要明确 business owner, product owner, data owner, risk owner, architecture owner, operations owner。
Define the next gate. 高管不需要一次批准全部未来, 他们需要知道下一步决策点和证据要求。
Write the stop rule before asking for money. 停止规则越清楚, funding 请求越可信。

3. Audience Map

同一个 AI case 可以有同一份事实底座, 但不同受众的 "so what" 不同。

Audience	他们真正关心	应该强调	证据材料	避免说法
CEO / Executive Sponsor	战略价值、速度、竞争差异、组织影响	为什么现在做, 业务能力如何变化, 风险是否被控制在可接受范围	1 页决策 memo, capability impact, portfolio roadmap	"这是行业趋势, 大家都在做"
CTO / CIO	架构可行性、集成复杂度、平台复用、技术债	target architecture, build/buy/hybrid, model gateway, observability, rollback	ADR, C4, integration map, NFR, runbook	"先做个 demo, 后面再架构化"
CFO / Finance	投入产出、成本边界、机会成本、扩张经济性	baseline, unit economics, TCO, sensitivity, funding stages	business case, cost per case, budget cap, scale model	"模型很便宜, token 成本不高"
Business Owner	流程效果、用户接受度、SLA、客户体验	AS-IS / TO-BE workflow, adoption plan, operational metrics	BPMN, pilot report, adoption dashboard, SOP	"AI 会自动解决用户问题"
Risk / Compliance / Legal	风险分类、监管义务、人工监督、审计证据	decision boundary, human oversight, controls, residual risk, audit trail	control pack, risk acceptance memo, incident runbook	"我们不让 AI 做最终决策, 所以没风险"
Data Owner / CDO	数据质量、来源、权限、血缘、保留、PII	source of truth, data readiness, access model, freshness, deletion	data readiness pack, lineage, access matrix	"把文档放进向量库就可以"
CISO / Security	身份权限、数据外泄、prompt injection、工具权限	SSO, RBAC, tenant isolation, logging, red-team, kill switch	threat model, security review, vendor security evidence	"供应商说他们很安全"
COO / Operations	稳定运营、异常处理、培训、责任交接	RACI, operating cadence, support model, escalation path	operating model, runbook, training plan	"上线后用户自己会学会"
Procurement / Vendor Owner	合同、SLA、锁定、退出、供应商风险	due diligence, contractable obligations, exit plan	vendor scorecard, DPA/SLA summary, lock-in analysis	"这个 vendor demo 最好看"
Board / Audit Committee	重大风险、治理、问责、监管可解释性	portfolio governance, high-risk use case controls, incident readiness	portfolio roadmap memo, governance model, risk dashboard	"AI strategy 正在推进中"

4. Bad Expression vs Good Expression

Topic	坏表达	好表达
Use case	我们要做一个 AI chatbot。	我们请求批准一个 6 周 pilot, 用 AI-assisted copilot 帮助客服处理政策查询和投诉分流, 目标是把可辅助 case 的平均处理时间从 18 分钟降到 12 分钟以内。
Why AI	大模型现在能力很强, 可以提升效率。	当前瓶颈不是缺人回答, 而是员工查找政策、判断例外和引用来源耗时。RAG copilot 能把知识检索、引用和建议草稿前置, 人仍负责最终回复。
Architecture	我们会接入向量数据库和 LLM。	推荐 RAG + human-in-the-loop, 因为政策内容变化快, 需要可追溯引用、权限过滤和快速回滚；不推荐 fine-tuning 作为第一阶段方案。
Risk	风险可控, 因为有人审核。	高风险回复必须由主管审批；系统记录 source, retrieval result, prompt version, model version, reviewer 和 final action；unsupported claim rate 超过 3% 时暂停扩展。
Eval	准确率达到 90%。	在 300 个 golden cases 上, grounded answer rate 为 91%, red-flag recall 为 97%, unsupported claim rate 为 2.1%, 平均延迟 4.2 秒, 成本为每 case 0.18 美元。
ROI	可以节省很多人工成本。	保守模型只计算 30% 可辅助 case, 每 case 节省 4 分钟, 扣除 license、平台、review 和运营成本后, 年化净收益区间为 18 万到 42 万美元。
Vendor	Vendor A 功能最全。	Vendor A 在 workflow fit、eval export、audit log、SSO/RBAC 和退出能力上得分最高；虽然 license 成本高 18%, 但能降低集成和审计风险。
Adoption	培训完成就可以上线。	上线门槛包括 70% 目标用户每周使用 3 次以上, override reason 被结构化记录, 前线主管每周复盘低信任 case, SOP 更新完成。
Stop	如果效果不好再说。	如果第 4 周仍无法达到 red-flag recall 95%, unsupported claim rate 低于 3%, 或业务 owner 无法确认 TO-BE workflow, pilot 停止并回到 workflow-only 方案。
Incident	AI 出了一些问题, 正在修。	09:20 发现 14 个 case 可能引用过期政策, 09:35 已切换到 read-only mode, 10:10 完成影响范围抽样, 当前无资金损失证据, 下一次更新在 12:00。

5. 30 秒 / 2 分钟 / 10 分钟表达模板

5.1 30 秒表达

适用场景: 电梯沟通、会议开场、被高管临时追问。

结构:

Decision requested: 要批准的动作。
Business pain: 当前基线和痛点。
Recommendation: 推荐方案。
Evidence and risk: 已有证据和主要控制。
Next gate: 下一次决策点。

示例:

我请求批准一个 6 周、8 万美元以内的 AI pilot, 用于客服政策查询和投诉分流 copilot。当前平均处理时间 18 分钟, 主要耗时在查找政策和判断例外。推荐 RAG + human approval, 不做全自动回复。我们已有 300 个历史 case 的 eval set, 目标是 red-flag recall 高于 95%, unsupported claim rate 低于 3%。第 4 周做 go/stop review, 如果质量或 adoption 不达标就停止。

5.2 2 分钟表达

适用场景: steering committee, sponsor 预沟通, vendor/build-buy 讨论。

结构:

Decision requested: fund pilot / select vendor / approve architecture / accept residual risk / scale or stop。
Context: baseline, users, workflow bottleneck, why now。
Options: do nothing, workflow-only, AI-assisted, agentic or vendor/custom/hybrid。
Recommendation: 推荐方案和不推荐其他方案的原因。
Evidence: discovery, data readiness, eval, user feedback, architecture review。
Metrics: business, quality, risk, adoption, cost。
Risks/controls: top 3 risks and controls。
Next 30 days and Stop rule: 清楚说明下一步和停止条件。

示例:

请求批准一个受控 pilot, 而不是生产扩张。当前客服团队每月处理 12,000 个政策相关 case, 平均 18 分钟, 返工率 9%。我们比较了三种方案: 只改流程、RAG copilot、agent 自动处理。只改流程能降低部分等待, 但不能解决政策查找和引用一致性；agent 自动处理风险过高。推荐 RAG copilot, 人审核后发送。证据包括 18 个用户访谈、300 个历史 case eval、数据 owner 确认的政策源和初步架构评审。成功指标是处理时间降到 12 分钟以内, red-flag recall 高于 95%, unsupported claim rate 低于 3%, 每 case 成本低于 0.35 美元, 70% 目标用户每周使用 3 次以上。主要风险是过期政策、错误建议和过度依赖, controls 是 source freshness check、强制引用、主管审批、audit log 和 kill switch。第 4 周 review, 第 6 周决定 stop, extend, or scale to one additional team。

5.3 10 分钟表达

适用场景: executive review, investment committee, architecture review board, risk committee。

建议结构:

Minute	内容	关键句
0-1	Decision requested	"今天需要决定的是 pilot funding, 不是 full rollout。"
1-2	Business context	"当前基线是 volume, cycle time, quality, cost, risk 五组指标。"
2-3	User and workflow	"AI 改变的是这三个步骤, 不改变最终责任人。"
3-4	Options	"我们比较了 no-AI, workflow-only, AI-assisted, agentic 四个选项。"
4-5	Recommendation	"推荐 AI-assisted, 因为它在价值、可控性、上线速度之间最平衡。"
5-6	Architecture	"架构核心是 RAG, policy source of truth, role-based retrieval, audit log, eval gate, human approval。"
6-7	Eval and evidence	"上线前用 golden set, 上线后用 sampling and monitoring, 不是一次性测试。"
7-8	Risks and controls	"剩余风险主要是 stale source, unsupported claim, over-reliance, vendor outage。"
8-9	Economics	"用 conservative, expected, upside 三档模型看 unit economics 和 TCO。"
9-10	Next 30 days and stop rule	"下一步产生 5 个证据包, 第 4 周触发 go/stop gate。"

6. 1-Page Executive Memo Template

用途: 所有 AI case 的默认 1 页版本。每一行都服务于决策, 不写背景长文。

Section	写法	高管阅读测试
Decision requested	用一个动词开头: approve, fund, select, accept, scale, stop。写清金额、周期、范围、owner。	读完第一行能立刻说 yes, no, or revise。
Context	写业务基线、用户、流程痛点、为什么现在需要决策。	能看出这不是 "AI for AI's sake"。
Options	至少比较 no-AI, workflow-only, AI-assisted, agentic 或 build/buy/hybrid。	能看出推荐不是预设答案。
Recommendation	写推荐方案、适用边界、不推荐其他方案的原因。	能看出取舍, 不是功能清单。
Evidence	写 discovery, data, eval, architecture, risk, user feedback, vendor diligence 的证据。	能区分事实、假设和待验证风险。
Metrics	同时覆盖 business, quality, risk, adoption, cost。	能判断是否值得继续投钱。
Risks/controls	写 top risks, preventive/detective/corrective controls, residual risk owner。	能判断是否可接受, 谁负责。
Next 30 days	写 3-5 个具体交付物和 review cadence。	能看出资金会换来哪些证据。
Stop rule	写明确阈值、触发人、停止动作和 fallback。	能防止 pilot 无限延长。

1 页 memo 的推荐排版:

Title: Decision on AI-assisted customer service copilot pilot

Decision requested:
Approve a 6-week pilot capped at $80k for one customer service team, owned by the VP Customer Operations and AI Product Owner.

Context:
Current monthly volume is 12,000 policy-related cases. Average handling time is 18 minutes, with 9% rework. The main bottleneck is policy lookup, exception judgment, and source citation.

Options:
1. Do nothing: no implementation risk, but no reduction in backlog.
2. Workflow-only: improves routing, but leaves policy lookup effort unchanged.
3. AI-assisted RAG copilot: improves lookup and drafting while keeping human approval.
4. Agentic auto-resolution: highest automation potential, but risk is too high for first release.

Recommendation:
Choose AI-assisted RAG copilot for a controlled pilot. It gives measurable value without moving final customer communication outside human accountability.

Evidence:
18 user interviews, 300 historical cases, confirmed policy source owner, initial data classification, draft eval suite, architecture review of RAG + audit log + HITL.

Metrics:
Cycle time below 12 minutes for assisted cases, red-flag recall above 95%, unsupported claim rate below 3%, weekly active usage above 70%, cost per assisted case below $0.35.

Risks/controls:
Stale source controlled by freshness check; unsupported claims controlled by citation requirement and eval gate; over-reliance controlled by mandatory human approval and override reason capture.

Next 30 days:
Finalize workflow, build prototype, run offline eval, train 20 pilot users, complete risk and security review, prepare week-4 go/stop report.

Stop rule:
Stop if red-flag recall stays below 95%, unsupported claim rate stays above 3%, data owner cannot certify source freshness, or business owner cannot confirm measurable workflow adoption.

7. Memo Pack

7.1 AI Pilot Funding Memo

用途: 请求 discovery 或 pilot 资金。重点是证明 "值得花小钱买证据", 不是证明 "已经可以规模化"。

Section	Template
Decision requested	请求批准一个限定周期、限定范围、限定预算的 pilot。写清 sponsor, business owner, AI product owner, risk owner, target users, funding cap。推荐句式: "Approve a 6-week controlled pilot capped at $80k for one operations team, with week-4 go/stop review."
Context	写业务痛点和当前基线: volume, cycle time, backlog, error, complaint, cost per case, employee pain。说明为什么现在是合适窗口: 新政策压力、SLA 下降、成本压力、系统升级、竞争压力或监管要求。
Options	比较 4 个选项: do nothing, workflow-only, AI-assisted, agentic automation。对每个选项写 value, risk, time-to-evidence, reversibility。
Recommendation	推荐最低足够风险的 pilot。通常第一阶段推荐 AI-assisted + HITL, 不推荐直接 agentic 自动执行。说明 pilot 要验证 value, quality, risk, adoption, economics 五件事。
Evidence	包括用户访谈、流程图、历史 case 样本、数据 source owner、初步 eval set、风险初筛、成本模型。证据不足时要说 "confidence is medium because data labels are partial", 而不是假装确定。
Metrics	Business: cycle time, backlog, rework, cost per case。Quality: groundedness, red-flag recall, unsupported claim rate。Risk: policy violation, escalation, incident。Adoption: weekly active users, repeat usage, override reason。Cost: license, token, review time, support。
Risks/controls	Top risks: wrong advice, stale data, PII leakage, over-reliance, scope creep。Controls: approved sources, RBAC, citation requirement, human approval, audit log, eval gate, pilot cohort limit, kill switch。
Next 30 days	完成 AS-IS / TO-BE workflow, data readiness sign-off, golden set, prototype, offline eval, pilot SOP, training, weekly steering review。
Stop rule	停止或降级到 workflow-only 的条件: 无法建立可信基线, red-flag recall 低于 95%, unsupported claim rate 高于 3%, target users weekly usage 低于 40%, 单 case 成本超过收益上限, risk owner 不接受 residual risk。

强表达:

本次 funding 购买的是证据, 不是承诺规模化。第 6 周只有三种结论: stop, extend with named evidence gap, or scale to the next controlled cohort。

7.2 Architecture Decision Memo

用途: 在 RAG, fine-tuning, agent, workflow automation, rules engine, vendor product, custom build, hybrid 之间做架构决策。

Section	Template
Decision requested	请求批准一个架构选项和约束边界。写清本决策覆盖的 use case, 不覆盖的 future scope, 以及谁拥有 ADR。
Context	写业务 workflow、数据来源、权限边界、延迟要求、审计要求、风险等级、现有系统约束。说明架构决策如何影响 speed, control, cost, scale。
Options	至少比较 no-AI workflow, RAG, fine-tuning, agentic workflow, vendor product, hybrid。每个选项用相同维度比较: data sensitivity, accuracy need, explainability, latency, integration, operations, reversibility, cost。
Recommendation	推荐一条主路径和一条 fallback。示例: "Use RAG with role-filtered retrieval, model gateway, eval gate, HITL approval, and full audit log; defer fine-tuning until recurring failure patterns prove model adaptation is needed."
Evidence	包括 data readiness, retrieval test, source freshness, NFR, threat model, vendor/platform capability, proof of concept result, eval suite, architecture review comments。
Metrics	Technical: retrieval hit rate, groundedness, latency p95, availability, cost per task, rollback time。Business: task completion, handling time。Risk: unauthorized retrieval, unsupported claim, audit completeness, model drift。
Risks/controls	Risks: access leakage, stale knowledge, prompt injection, excessive agency, vendor outage, hidden coupling。Controls: RBAC, source allowlist, prompt/tool policy, versioned prompts, model gateway, observability, audit log, fallback to manual workflow。
Next 30 days	Finalize ADR, build thin vertical slice, run retrieval and red-team eval, validate audit log schema, review NFR with CTO/CISO, define rollback and support runbook。
Stop rule	停止该架构路径的条件: audit log 无法重建 case, RBAC 不能按角色过滤, p95 latency 超过业务 SLA, retrieval groundedness 长期低于阈值, rollback 无法在约定时间内完成, 或成本曲线无法随规模收敛。

强表达:

架构选择不是为了最先进模型, 而是为了在给定风险等级下, 用最少不可逆承诺获得可评估、可审计、可回滚的能力。

7.3 Risk Acceptance Memo

用途: 当 AI pilot 或 production release 仍有 residual risk, 需要 risk owner 和 sponsor 明确接受、限制或拒绝。

Section	Template
Decision requested	请求接受一个有边界的 residual risk, 或批准补充 controls 后进入 pilot / production。写清 risk owner, business owner, system owner, review date。
Context	写风险分类、业务场景、用户群、AI 决策边界、是否影响客户、资金、信用、合规、隐私或法律义务。明确 AI 是 recommend, draft, route, approve, execute 中的哪一种角色。
Options	比较 avoid, mitigate, transfer, accept with controls, delay release。说明每个选项对 value, timeline, risk exposure, cost 的影响。
Recommendation	推荐接受或不接受的 residual risk。强 memo 会写 "accepted only for pilot cohort and non-binding recommendations", 不会写泛泛的 "risk acceptable"。
Evidence	包括 risk assessment, control test, eval results, red-team findings, legal/compliance comments, audit log sample, user training evidence, incident drill result。
Metrics	Risk metrics: policy violation rate, unsupported claim rate, red-flag recall, unsafe output rate, PII exposure event, human override rate, incident count, audit completeness。Business metrics 用来说明风险接受换来的价值。
Risks/controls	写 preventive, detective, corrective controls。Preventive: approved sources, role permissions, refusal rules。Detective: sampling, monitoring, anomaly alerts。Corrective: kill switch, rollback, user notification, postmortem。
Next 30 days	完成 control evidence pack, reviewer calibration, incident drill, production sampling plan, risk committee review, residual risk sign-off, release readiness checklist。
Stop rule	撤销 risk acceptance 的条件: critical incident, PII exposure, repeated unsafe output, red-flag recall below agreed threshold, audit log gaps, control owner missing, or material scope expansion without new review。

强表达:

风险接受不是 "同意有风险", 而是明确谁在什么范围内接受哪一种剩余风险, 基于哪些 controls, 到哪一天重新评估。

7.4 Vendor Selection Memo

用途: 选择 AI vendor, 进入采购谈判, 或决定 build/buy/hybrid。

Section	Template
Decision requested	请求批准 vendor shortlist, preferred vendor, build/buy/hybrid decision, 或进入 contract negotiation。写清采购范围、预算上限、合同期限、退出条件。
Context	写业务 use case、企业架构约束、数据分类、集成系统、监管/审计要求、内部构建能力、上线窗口。
Options	比较 Vendor A, Vendor B, internal build, hybrid, defer。维度包括 business fit, eval quality, security, data privacy, integration, audit, SLA, cost, customization, lock-in, exit。
Recommendation	推荐 vendor 或 hybrid path, 同时写明 negotiation conditions。示例: "Select Vendor A only if audit log export, data retention, model change notification, and termination export rights are contractually committed."
Evidence	包括 demo script result, customer-specific eval, security evidence, SOC/ISO material, architecture review, data flow, DPA/SLA review, reference calls, cost model, exit plan。
Metrics	Vendor score, eval pass rate, latency, uptime/SLA, integration effort, projected cost per case, implementation timeline, support response, exit cost, risk rating。
Risks/controls	Risks: vendor lock-in, hidden subprocessors, model change, data retention, weak audit, cost overrun, custom services dependency。Controls: contract clauses, data processing terms, export rights, model version notice, budget caps, phased rollout, internal fallback。
Next 30 days	Complete security/privacy/procurement review, run customer-specific eval, negotiate contract terms, validate data flow, prepare implementation plan, define vendor governance cadence。
Stop rule	停止 vendor selection 或退回 internal/hybrid 的条件: vendor cannot pass security/data review, refuses audit export, cannot support required RBAC, cost per case exceeds threshold, SLA excludes critical dependency, or contract contradicts risk requirements。

强表达:

选 vendor 不是选最强 demo, 而是选择在我们的数据、流程、审计和运营约束下最可控的交付路径。

7.5 Scale / Stop Decision Memo

用途: pilot 结束后决定 scale, extend, pause, stop, 或转向 alternative option。

Section	Template
Decision requested	请求批准 scale to next cohort, extend pilot with named evidence gap, stop, or pivot。写清新范围、预算、owner、risk review。
Context	写 pilot 范围、目标用户、时间、原始成功标准、实际运行条件。说明哪些结果可外推, 哪些不能外推。
Options	比较 stop, extend, limited scale, broad rollout, pivot to workflow-only / vendor / architecture alternative。每个选项写 value, risk, cost, learning value。
Recommendation	推荐 scale 或 stop, 不能模糊。若 extend, 必须只因为一个明确 evidence gap, 并限制延长期。
Evidence	包括 pilot metrics, eval trend, adoption dashboard, user feedback, incident log, cost report, risk review, operations readiness, architecture performance。
Metrics	Business: cycle time, backlog, quality, customer satisfaction, cost per case。Quality: eval pass, groundedness, unsupported claim, red-flag recall。Adoption: activation, repeat usage, override reason, trust score。Risk: incidents, escalations, audit completeness。
Risks/controls	Scale risks: distribution shift, volume load, support capacity, training gaps, control fatigue, cost surprise。Controls: phased rollout, cohort gating, champion model, monitoring, rollback, budget cap, capacity planning。
Next 30 days	If scale: onboard next cohort, run regression eval, update SOP, monitor daily, review weekly, finalize production RACI。If stop: archive learnings, retire access, close vendor spend, document reusable assets。
Stop rule	停止扩张的条件: value metric falls below 70% of pilot result, unsafe output increases materially, adoption drops below agreed threshold, support backlog grows, cost per case exceeds cap, or risk owner withdraws acceptance。

强表达:

Scale decision 不能只看 pilot 平均值, 要看质量、风险、adoption 和单位经济是否在更大人群下仍成立。

7.6 Incident Update Memo

用途: AI 质量、安全、隐私、可用性、成本或 vendor incident 期间给高管同步状态。

Section	Template
Decision requested	请求批准 containment action, customer/regulator communication path, temporary shutdown, degraded mode, vendor escalation, or resource support。
Context	写 incident 类型、发现时间、影响范围、当前状态、已采取行动、下一次更新时间。区分 confirmed impact and potential exposure。
Options	比较 keep running with monitoring, degraded mode, pause affected feature, full shutdown, manual fallback, vendor escalation。每个选项写 customer impact, risk reduction, operational cost。
Recommendation	推荐一个 containment path。示例: "Pause auto-drafting for policy exception cases, keep read-only retrieval available, route high-risk cases to manual workflow."
Evidence	包括 logs, sampled cases, user reports, monitoring alerts, vendor status, audit trail, data exposure analysis, timeline。对未知事项写 "not yet confirmed" 并给出验证时间。
Metrics	Incident severity, impacted users/cases, time to detect, time to contain, error rate, unsafe output count, PII exposure count, backlog impact, recovery ETA。
Risks/controls	Risks: customer harm, regulatory exposure, repeated bad output, operational backlog, loss of trust。Controls: kill switch, manual fallback, user notification, blocked source, vendor ticket, sampling, post-incident eval。
Next 30 days	Complete root cause analysis, corrective action, regression eval, control update, user communication, risk review, postmortem, executive closure note。
Stop rule	Feature remains paused until root cause is understood, affected cases are reviewed, corrective control passes regression eval, risk owner signs restart, and monitoring is active for the affected failure mode。

强表达:

Incident update 的第一责任是降低不确定性: 已确认什么, 还不知道什么, 何时知道, 现在用什么 containment 限制损害。

7.7 Portfolio Roadmap Memo

用途: 给 ExCo, transformation office, enterprise architecture board 或 AI governance committee 决定 AI portfolio 优先级和路线图。

Section	Template
Decision requested	请求批准 portfolio focus, funding allocation, sequencing, governance model, or retirement of low-value AI initiatives。
Context	写当前 AI initiative inventory, business domains, capability gaps, spend, risk tiers, duplicated efforts, platform constraints, talent constraints。
Options	比较 scattered experimentation, platform-first, domain portfolio, risk-first sequencing, pause-and-consolidate。说明每个选项对 speed, reuse, risk, cost, accountability 的影响。
Recommendation	推荐 portfolio theme and sequence。示例: "Prioritize three domain pilots in customer operations, compliance operations, and knowledge management, supported by shared eval, model gateway, data controls, and incident process."
Evidence	包括 use case inventory, value/risk matrix, capability map, architecture dependency, vendor landscape, operating model maturity, budget analysis, talent capacity。
Metrics	Portfolio value: expected benefit, risk-adjusted ROI, time to evidence。Execution: cycle time from intake to pilot, reuse rate, platform adoption。Risk: high-risk use case coverage, control maturity, incident trend。Cost: run-rate, vendor overlap, cost per capability。
Risks/controls	Risks: pilot sprawl, duplicated vendors, shadow AI, weak ownership, over-centralization, platform bottleneck。Controls: intake gate, architecture review, shared eval standards, vendor governance, RACI, quarterly portfolio review, retirement rules。
Next 30 days	Finalize inventory, score top use cases, choose first portfolio wave, assign owners, define shared platform and governance minimums, stop or merge duplicate pilots。
Stop rule	Stop or de-prioritize initiatives with no measurable business owner, no data owner, no eval path, unacceptable risk tier, duplicate vendor capability, or no route to adoption after two review cycles。

强表达:

AI portfolio 不是越多 pilot 越好, 而是让有限的 funding、数据、架构和治理能力集中到能形成可复用 enterprise capability 的路径上。

8. Common Executive Questions And Strong Answers

Question	Strong answer pattern
Why now?	"Because baseline pain is measurable, data source is available, business owner is committed, and the pilot can produce go/stop evidence within 6 weeks."
Why AI instead of process improvement?	"Workflow-only improves routing, but the main bottleneck is knowledge lookup, source citation, and exception reasoning. AI-assisted is justified only for those steps; the rest remains workflow redesign."
Is this replacing people?	"The first release changes task composition, not accountability. AI drafts and retrieves; humans approve, handle exceptions, and own final customer action."
How do we know it is safe?	"We do not claim it is safe in general. We define a narrow scope, approved sources, role-based access, eval thresholds, human approval, audit logs, monitoring, and a kill switch."
What if it hallucinates?	"Unsupported claims are a tracked failure mode. Controls are source-grounded answers, citation requirement, offline eval, production sampling, human approval for high-risk responses, and stop threshold."
What data leaves the company?	"The data flow identifies prompt content, retrieval content, logs, telemetry, embeddings, and subprocessors. Anything outside approved boundary is blocked before pilot."
What is the ROI?	"The base case uses current volume, assisted percentage, minutes saved, quality improvement, review time, license, platform, support, and risk cost. We show conservative, expected, and upside ranges."
What if adoption is low?	"Low adoption is a stop or redesign signal, not a training-only problem. We track activation, repeat usage, override reasons, trust, and manager feedback weekly."
Why this architecture?	"It is the most reversible path that meets audit, data freshness, permission, latency, and eval needs. More autonomous patterns are deferred until evidence supports them."
Why this vendor?	"The recommendation is based on workflow fit, customer-specific eval, security/privacy evidence, audit export, integration, SLA, cost, and exit terms, not demo quality alone."
How do we avoid lock-in?	"We require export of prompts/config, logs, evals, data, embeddings where applicable, and contract terms for termination support. Architecture separates model gateway, retrieval sources, and audit evidence."
Who owns this after launch?	"Ownership is split but explicit: business owner owns workflow value, product owner owns roadmap, data owner owns sources, platform owner owns reliability, EvalOps owns quality gates, risk owner owns oversight."
Can auditors reconstruct decisions?	"Audit evidence includes user, role, input, source documents, retrieval results, model version, prompt version, output, reviewer, approval, and final action."
What happens when the model changes?	"Model changes go through version notice, regression eval, release gate, rollback plan, and risk review for material behavior changes."
What is the worst credible failure?	"For this use case, the worst credible failure is a customer-facing answer based on stale policy that triggers financial or compliance impact. Controls focus on freshness, citations, approval, monitoring, and fast containment."
When do we stop?	"We stop when quality, risk, adoption, or unit economics miss the pre-agreed thresholds, or when an owner required for safe operation is absent."
How does this fit the enterprise AI roadmap?	"This use case reuses shared capabilities: approved sources, model gateway, eval suite, audit log, risk controls, and operating cadence. It is not a one-off chatbot."

9. Executive Language Cheatsheet

Instead of saying	Say
LLM	AI capability, model component, probabilistic reasoning component
RAG	Answers grounded in approved enterprise sources with citation and freshness controls
Agent	AI workflow that can plan and call tools within a controlled permission boundary
Prompt engineering	Governed instruction, policy and response design with versioning and regression tests
Hallucination	Unsupported claim that can create operational, customer, compliance or audit risk
Eval	Evidence that the system meets business, quality, risk and cost thresholds before and after release
Guardrail	Preventive, detective and corrective control
Fine-tuning	Model adaptation that may help recurring pattern quality but increases lifecycle and regression obligations
Vector database	Search index for approved knowledge, requiring access control, freshness, retention and deletion rules
Human-in-the-loop	Named human accountability at specific decision points, with audit trail and override reason
Model drift	Change in model or data behavior that can reduce quality, safety or business value over time
Token cost	One component of cost per task, not the full cost of operating the capability

10. Memo Quality Checklist

Before sending any executive memo, check these criteria:

Check	Pass condition
Decision clarity	The first section states one specific decision and a yes/no path.
Scope clarity	The memo defines included workflow, excluded workflow, user group and time horizon.
Option discipline	No-AI and workflow-only options are compared, not ignored.
Evidence discipline	Claims are backed by named evidence: interviews, data, eval, architecture review, risk review, vendor evidence.
Metric balance	Metrics include business outcome, AI quality, risk, adoption and cost.
Risk realism	Top risks are concrete and mapped to controls and owners.
Architecture relevance	Architecture is explained through control, operation, scale and rollback impact.
Financial credibility	ROI uses baseline, sensitivity and cost categories beyond token cost.
Operating ownership	Memo names business, product, architecture, data, risk, operations and vendor owners where relevant.
Stop rule	Stop condition is measurable, time-bound and tied to an action.
Executive compression	A sponsor can retell the recommendation in 30 seconds without technical translation.

11. Practice Routine

For each AI case in the learning portfolio:

Write the 1-page memo first.
Rewrite the same memo for CTO, CFO, business owner, risk/compliance and data owner.
Record a 30-second version and remove every sentence that does not support a decision.
Practice the 2-minute version with one hostile question: "Why should we believe this?"
Practice the 10-minute version with three interruptions: risk, cost, and architecture.
Convert the memo into portfolio evidence by linking it to architecture diagram, eval matrix, risk/control pack and business case.

The standard of mastery:

You can explain the same AI initiative as business value, architecture decision, risk decision, eval design, funding request and portfolio roadmap without changing the underlying facts.