AI Management Information:董事会报告架构
AI Management Information / Board Reporting Architecture 不是把 AI 项目周报换成董事会格式,也不是堆更多图表。它要把 AI portfolio 的事实、价值、风险、控制有效性、客户伤害、事件、供应商集中度、采用质量和整改动作做成可追溯的信息产品,让管理层、董事会和审计委员会能监督、挑战、投资、暂停、扩张或要求整改。
AI Management Information / Board Reporting Architecture 解读
配对阅读:本篇的操作手册版(模板/RACI/门禁/runbook)是
docs/AI_MANAGEMENT_INFORMATION_BOARD_REPORTING_PLAYBOOK.md。第一遍读本篇建立原理与架构判断;第二遍做案例时再用 playbook 查表落地,两者不需要重复精读。
Source Anchors
| Source | Link | 用途 |
|---|---|---|
| NIST AI RMF | https://www.nist.gov/itl/ai-risk-management-framework | 组织 AI risk lifecycle、impact、control 和治理报告。 |
| NIST AIRC AI RMF functions | https://airc.nist.gov/airmf-resources/airmf/ | 用 Govern / Map / Measure / Manage 设计 MI taxonomy、metric owner 和 action loop。 |
| ISO/IEC 42001 | https://www.iso.org/standard/42001 | 用 AI management system、绩效评价、管理评审和持续改进连接 MI 到 AIMS。 |
| Federal Reserve SR 26-2 | https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm | 用 2026 revised model-risk guidance 的 risk-based、materiality、inventory、monitoring 和 governance 思维校准金融机构 MI。 |
| OCC Model Risk Management Handbook legacy link | https://www.occ.gov/publications-and-resources/publications/comptrollers-handbook/files/model-risk-management/index-model-risk-management.html | 作为 legacy context; 当前 OCC 页面已重定向, 当前模型风险锚点应转向 SR 26-2 / OCC 2026 guidance。 |
核心导读
AI Management Information / Board Reporting Architecture 不是把 AI 项目周报换成董事会格式,也不是堆更多图表。它要把 AI portfolio 的事实、价值、风险、控制有效性、客户伤害、事件、供应商集中度、采用质量和整改动作做成可追溯的信息产品,让管理层、董事会和审计委员会能监督、挑战、投资、暂停、扩张或要求整改。
成熟的 MI architecture 关注“董事会依赖的事实如何生成、校验、追溯和触发行动”。没有这层架构,board pack 很容易退化为 innovation showcase、项目状态报告或人工拼接的证据 collage:看起来完整,但无法回答数字从哪里来、定义是否稳定、阈值是否合理、谁负责行动、行动是否关闭。
可以把它理解为:
AI Board MI =
metric contract
+ telemetry/evidence lineage
+ risk appetite threshold
+ management action
+ board decision use
1. 问题定义
董事会和审计委员会通常不缺 AI 叙事,缺的是可决策事实。关键问题包括:
What facts feed board oversight?
How is each metric defined?
Where did the number come from?
Which threshold makes it reportable?
Who owns remediation?
What decision can be made from it?
金融零售 AI 的管理信息难点不在“缺少数据”,而在“数据不能形成可决策事实”。模型网关有 token、latency、usage;客服系统有投诉和质检;风险系统有事件和控制测试;财务系统有成本和收益。但如果这些系统没有共同的定义、lineage、阈值和 action loop,董事会只能看到活动量,不能判断 value、risk、control effectiveness 和 customer harm 是否在风险偏好内。
| 断点 | 表现 | 后果 |
|---|---|---|
| Telemetry disconnected from value | 有 token、latency、usage,没有 customer outcome 和 process baseline | AI 成功被 activity 指标替代 |
| Control effectiveness not measured | 控制存在于流程图,但没有 pass rate、sample result、exception aging | 无法判断控制是否有效 |
| Incident taxonomy inconsistent | 安全、模型、客服、投诉各自记录事件 | customer harm 和 root cause 无法汇总 |
| Vendor concentration invisible | 每个 use case 独立审批,portfolio 共享同一 model/vendor/evidence stack | 单点变化引发组合风险 |
| Adoption overstated | 登录或调用被当成采用 | 不能判断用户是否产生合格价值事件 |
| Manual board reporting | 报告靠访谈、截图和手工汇总 | 内审、审计委员会和监管问询时无法重建事实 |
治理材料回答“谁监督、监督什么、在哪个委员会决策”。MI architecture 回答“监督依赖的事实如何产生、质量如何保证、异常如何行动、报告如何被审计”。
2. 架构模型
AI MI 要以指标契约为中心,而不是以 PPT 页面为中心。每个 board-level metric 都必须有定义、来源、口径、阈值、负责人、更新频率、数据质量规则和行动含义。
AI systems and workflows
-> model gateway / agent gateway telemetry
-> RAG and knowledge source logs
-> tool action and workflow event logs
-> control tests, evals, red-team, QA samples
-> incidents, complaints, customer harm, appeals
-> cost, adoption, value, finance baseline
-> vendor, model, dependency and inventory registry
-> MI data product layer
- metric contracts
- lineage graph
- quality rules
- threshold and risk appetite rules
-> management dashboards
-> board / audit committee pack
-> decision, action, attestation and evidence closure
| 架构原则 | 设计含义 |
|---|---|
| Report from systems of record | Board number 不应首次在 slide deck 中被创造 |
| Separate metric logic from presentation | Dashboard、board pack、audit extract 复用同一 metric contract |
| Lineage before polish | 漂亮的红黄绿图表如果没有 lineage,就是弱 MI |
| Thresholds are controls | 阈值需要 owner、rationale、review cadence 和 exception process |
| Action log is part of MI | 报告红灯但没有 owner、due date 和 closure evidence,是不完整控制 |
| Combination view matters | 单个 use case 合格不代表组合风险、集中度和依赖关系合格 |
Metric contract
| Field | Example |
|---|---|
| Metric name | Unsupported claim rate |
| Decision purpose | 是否允许 customer service RAG 扩大到新产品线 |
| Numerator | AI-assisted responses sampled as unsupported by approved source |
| Denominator | sampled AI-assisted responses for regulated topics |
| Source systems | model gateway trace, RAG citation log, QA review system |
| Grain | response_id, use_case_id, period |
| Owner | Customer operations QA owner + AI product owner |
| Threshold | Green <= 2%, Amber > 2% and <= 3%, Red > 3% |
| Escalation | Red triggers product-line freeze and risk committee action |
MI lineage
business event
-> AI trace
-> control/eval result
-> data quality rule
-> metric calculation
-> dashboard tile
-> board statement
-> management action
如果 board pack 写着“AI customer harm incidents decreased 20%”,机构应能展示 incident taxonomy version、source systems、de-duplication logic、severity rule、customer impact classification、query version、owner sign-off 和 action status。
SR 26-2 语境
SR 26-2 的价值在于把金融机构 MI 拉回 risk-based、materiality-driven、inventory、monitoring 和 governance 的思维,而不是把所有 GenAI/agentic AI 机械塞进同一套模型风险表。AI board MI 应分层呈现:
- 传统模型风险范围内的复杂定量模型和关键估计。
- 非生成式 AI 或自动化决策中可沿用模型风险原则的部分。
- GenAI、RAG 和 agentic AI 的 broader AI controls,包括 prompt、corpus、tool authority、human oversight、eval、incident、change 和 evidence。
董事会需要 consolidated AI value/risk view,同时也需要看清不同治理域的边界,避免 scope confusion。
3. 关键机制/生命周期
AI MI 的生命周期从 use case 入库开始,到指标退役或控制关闭结束。
define oversight decision
-> create metric contract
-> map source systems and lineage
-> define data quality rules
-> set thresholds and action triggers
-> publish management dashboard
-> produce board view
-> challenge / decide / fund / hold / remediate
-> track action closure
-> attest and archive evidence
-> review metric relevance
关键机制包括:
- Decision-first metric design:每个指标都要指向 scale、hold、stop、fund、remediate、accept residual risk 之一;否则只是观察性数据。
- Metric contract versioning:口径、分母、抽样方法、阈值、source system 改变时,要能解释趋势变化来自业务现实、定义变化还是数据质量。
- Risk appetite threshold:红黄绿不是美化,而是控制;阈值要连接已批准的 risk appetite、stop rule、escalation 和 exception。
- Action log closure:每个 amber/red signal 都必须连接 owner、due date、status、closure evidence 和 residual risk。
- Attestation and challenge:报告前需要 owner signoff、数据质量检查和可挑战记录;报告后需要记录董事会或管理层的 challenge 与决定。
一个轻量 lineage prototype 可以用结构化数据验证 MI 可行性:
Input entities:
ai_trace(response_id, use_case_id, model_id, timestamp, source_doc_ids, ai_exposed)
qa_review(response_id, supported_by_source, regulated_topic, reviewer_id, review_date)
complaints(case_id, response_id, severity, ai_attributable, remediation_required)
metric_contract(metric_id, numerator_rule, denominator_rule, threshold_green, threshold_red, owner)
Metric:
unsupported_claim_rate =
count(response_id where regulated_topic = true and supported_by_source = false)
/ count(response_id where regulated_topic = true and qa_review exists)
Lineage output:
metric_id, reporting_period, source_tables, query_version,
denominator_count, numerator_count, threshold_status, action_required
学习价值不在于 SQL 本身,而在于能解释董事会数字为何变化:真实业务变化、指标定义变化、数据质量变化、source-system 延迟,还是阈值更新。
4. 证据与控制
MI control 的核心是让报告数字可被重建、挑战和行动。它要求数据产品、治理流程和报告材料共享同一套事实基础。
| 控制层 | 控制问题 | 证据 |
|---|---|---|
| Metric design control | 指标是否服务具体决策 | metric contract、decision purpose、approval |
| Source control | 数字是否来自系统记录而非人工拼接 | source-system mapping、query version、data extract log |
| Lineage control | board statement 是否可追溯 | lineage graph、transformation logic、quality checks |
| Threshold control | 红黄绿是否连接风险偏好 | threshold rationale、review cadence、exception log |
| Data quality control | 指标是否完整、及时、可比较 | completeness, freshness, reconciliation, anomaly checks |
| Action control | 异常是否触发 owner 和整改 | action log、due date、closure evidence |
| Access and retention control | 报告证据是否受控保存 | access log、retention policy、audit export |
| Attestation control | 报告前是否确认准确性 | owner signoff、challenge record、versioned pack |
Board-level MI 应覆盖 value、risk、control、harm、adoption、cost 和 concentration,避免单一维度叙事。
| MI domain | 示例指标 | 决策用途 |
|---|---|---|
| Value | qualified value events, AHT reduction, backlog age | scale/fund/stop |
| Control effectiveness | unsupported claim rate, citation completeness, HITL bypass | hold/remediate |
| Customer harm | AI-attributable complaint rate, appeal overturn rate, remediation count | incident response/risk appetite |
| Model and vendor concentration | risk-weighted exposure by vendor/model/source | diversification/resilience |
| Adoption quality | eligible workflow adoption, override reason mix, review burden | training/process redesign |
| Cost and resilience | cost per resolved case, latency SLO, outage impact | investment/vendor strategy |
| Change and release | high-risk release failures, rollback count, evidence completeness | governance improvement |
5. 金融零售/AI产品场景
Scenario: A retail bank runs six AI systems.
| System | AI role | Main board concern |
|---|---|---|
| Customer service RAG | draft grounded answer for agent | wrong policy commitment, complaint, stale source |
| Credit memo assistant | summarize documents for underwriter | fair lending, explanation, unsupported recommendation |
| AML copilot | draft case summary and evidence narrative | missed suspicious activity, weak SAR evidence |
| Fraud triage assistant | prioritize cases for analyst | customer friction, false positives, fraud loss |
| Branch knowledge assistant | answer staff policy questions | inconsistent advice, outdated policy |
| AI platform gateway | shared control plane | shadow AI, auditability, vendor concentration |
董事会监督议题应转成可追溯 MI:
| 监督议题 | MI metric | Source / lineage |
|---|---|---|
| Are customers being harmed? | AI-attributable complaint rate, appeal overturn rate, remediation count | complaint system + case tags + AI exposure trace |
| Are controls working? | citation completeness, unsupported claim rate, HITL bypass count, control test pass rate | RAG logs + QA samples + workflow approval events |
| Are we getting value? | qualified value events, AHT reduction, backlog age, cost per resolved case | workflow events + finance baseline + AI cost ledger |
| Are we concentrated? | risk-weighted exposure by model/vendor/knowledge source | AI inventory + dependency graph + gateway routing |
| Are we adopting safely? | eligible workflow adoption, override reason mix, review burden | user telemetry + workflow eligibility + review queue |
一个 decision-useful board slice 应该类似这样:
Decision requested:
approve limited scale of customer service RAG to two additional product lines.
Evidence:
qualified value events 74%;
unsupported claim rate 1.6%;
source freshness SLA 99.4%;
AI-attributable complaints flat to baseline;
vendor concentration amber.
Management action:
scale low-risk intents only;
run cross-use-case regression before credit-card dispute policies;
reduce model concentration before direct customer response.
这里的关键不是格式,而是每个数字都有 metric contract 和 lineage,每个异常都有 action owner,每个扩张决定都有边界条件。
6. 反模式
| 反模式 | 危险 | 更好的做法 |
|---|---|---|
| Board pack as slide assembly | 数字不能追溯或审计 | 建 MI data product、metric contract 和 lineage |
| Usage as value | 高使用可能代表返工、强制流程或糟糕 UX | 使用 qualified value events 和 outcome evidence |
| Controls without effectiveness metrics | “控制存在”不等于控制有效 | 定义 pass rate、sample result、exception aging |
| Thresholds without risk appetite | 红黄绿变成任意颜色 | 阈值连接 approved appetite、stop rule 和 exception |
| Incident counts without harm taxonomy | 客户影响被隐藏 | 记录 severity、harm type、remediation 和 AI attribution |
| Mixing GenAI with model-risk scope carelessly | 范围混乱,保证薄弱 | 分清治理域,同时统一 inventory 和 portfolio view |
| No action log | 报告不驱动控制改进 | 每个 amber/red signal 有 owner、due date、closure evidence |
| Manual evidence collage | 季度报告无法复现 | 报告从系统记录和版本化 query 自动生成 |
7. 最终心智模型
AI board reporting 的成熟度不在于页面多漂亮,而在于董事会看到的每个关键数字是否能被追溯、解释、挑战和行动。MI 是一个信息架构问题:把生产事实、控制测试、客户影响、价值结果、供应商依赖和整改动作转成 decision-useful evidence。
最终要形成一条稳定链路:
business event
-> AI trace
-> metric contract
-> lineage and quality rule
-> risk appetite threshold
-> management dashboard
-> board decision
-> action closure evidence
董事会不需要更多 AI activity reporting。它需要证据证明 value、risk、control effectiveness、customer harm、concentration 和 adoption 在风险偏好内,或者已经被明确 owner 和行动计划接住。
SOTA 状态标注 (2026-07-01)
本篇属于第二、三遍深读池(参考架构/深读笔记),未列入 12 周主线必读。时效基线为写作时点;引用前请按 CLAUDE.md 全局时效性硬规则复查最新进展。模块级 SOTA 对照见 docs/AI_SYSTEMATIC_LEARNING_ROADMAP_2026.md 各周「2026 SOTA 对照」行与文末「SOTA 检查」。