返回 Papers
AI 扩展计划 / Playbooks

AI Board / Audit Committee Governance Pack

董事会不需要知道每个 prompt 怎么写, 但必须能判断管理层是否在用可审计、可问责、可停止的方式推进 AI。强治理材料要回答四个问题:

716AI_BOARD_AUDIT_COMMITTEE_GOVERNANCE_PACK.md

AI Board and Audit Committee Governance Pack

目标: 帮助 AI PM / BA / Architect 把 AI initiative 转换成董事会、审计委员会、风险委员会和 AI governance committee 可以监督、追问和决策的材料。 定位: 本文件不是普通项目 memo, 而是 board-level oversight pack。重点是 material AI systems, portfolio risk, control effectiveness, residual risk, accountability, audit evidence, incident readiness 和 investment decision。 适用范围: 金融零售场景, 包括 credit underwriting assistant, AML copilot, customer service RAG, fraud decision support, wealth advisory guardrail, AI platform shared service。


1. Board Oversight Mindset

董事会不需要知道每个 prompt 怎么写, 但必须能判断管理层是否在用可审计、可问责、可停止的方式推进 AI。强治理材料要回答四个问题:

What material AI risk are we taking?
What controls reduce that risk and how do we know they work?
What residual risk remains and who owns it?
What investment or risk decision is needed this quarter?

PM / BA / Architect 的核心工作, 是把 technical detail 翻译成 board language:

Technical detailBoard-level translation材料表达
Model accuracy, hallucination, grounding客户伤害、错误建议、合规偏差、运营返工的概率和影响Risk metric, incident scenario, residual risk
RAG, vector index, source refresh知识来源是否被授权、是否过期、是否可追溯Data source control, freshness SLA, audit trail
Prompt injection, tool calling未授权操作、数据泄露、越权执行Threat scenario, preventive control, kill switch
Eval suite, golden set, red-team管理层如何证明上线前后控制有效Control effectiveness evidence, release gate
Human-in-the-loop哪个角色在何处承担最终责任Accountability map, approval matrix, override log
Model gateway, logging, versioning是否能重建决策、回滚、监控成本和行为变化Auditability, operational resilience, change control
Token cost, latency单位经济、SLA、客户体验和扩张成本Investment decision, scale constraint

董事会材料的质量标准:

  1. Scope is explicit. 写清 AI 是 draft, recommend, triage, approve, execute 中的哪一种角色。
  2. Materiality is stated. 说明是否影响信用、资金、客户建议、反洗钱、欺诈、隐私、模型风险或监管义务。
  3. Controls are testable. 不说 "加强监控", 要说谁监控什么、频率、阈值、证据在哪里。
  4. Residual risk is owned. 剩余风险必须有 business owner, risk owner, technology owner 和 review date。
  5. Stop rule is pre-agreed. 上线、扩张、例外和继续投资都必须有停止条件。

2. Governance Operating Model

2.1 Committee Map

Forum典型成员决策权限输入材料输出
BoardDirectors, CEO, CRO, CIO/CTO, General CounselAI strategy, material risk appetite, major investment, severe incident oversightQuarterly AI governance report, portfolio risk dashboard, incident board updateOversight challenge, risk appetite direction, funding / risk decision
Audit CommitteeIndependent directors, CFO, CAE, CISO, external audit liaisonAudit evidence, internal control, model/change governance, management attestationAudit evidence request list, control test result, attestation packAudit scope, control remediation, evidence expectation
Risk CommitteeCRO, business risk, compliance, legal, data protection, model riskHigh-risk use case approval, risk acceptance, exception, incident containmentMaterial AI system register, exception memo, risk dashboardRisk decision, residual risk acceptance, restriction
AI Governance CommitteeAI product, architecture, data, security, compliance, risk, business ownerIntake, classification, release gate, monitoring cadence, retirementUse case score, policy memo, eval report, architecture review, RACIApproved policy, release gate decision, remediation action
Management SteeringSponsor, PM, BA, Architect, Operations, FinanceExecution priority, funding allocation, adoption, operational readinessRoadmap, ROI, adoption dashboard, operating metricsDelivery decision, resource allocation

2.2 Decision Rights

DecisionOwnerRequired approversEvidence required
Classify material AI systemAI governance leadRisk owner, business owner, architecture ownerUse case intake, materiality assessment, data classification
Approve AI policyAI governance committeeCRO, CIO/CTO, Legal, CISO, CAE observerPolicy approval memo, control mapping, exceptions process
Approve high-risk pilotRisk committeeBusiness sponsor, risk owner, data owner, CISOException memo, eval result, audit log sample, stop rule
Release production AIAI governance committeeBusiness owner, technology owner, risk ownerRelease gate pack, control test, incident runbook, attestation
Accept residual material riskRisk committee or board committeeCRO, business executive, Legal where applicableResidual risk statement, control effectiveness evidence, review date
Report severe AI incidentBoard / Audit / Risk committeeCEO/CRO/CIO as applicableIncident board update, timeline, impact, containment, remediation

3. Case Coverage Matrix

Use caseMaterialityPrimary board riskKey controlsBoard metric examples
Credit underwriting assistantHigh信用决策偏差、fair lending、拒贷解释不足、模型风险Human credit officer approval, adverse action reason control, bias testing, model versioning, audit trailDisparate impact ratio, override rate, adverse action explanation completeness, approval error rate
AML copilotHigh漏报可疑活动、误报增加、监管检查证据不足Red-flag recall threshold, SAR drafting review, evidence citation, case sampling, reviewer calibrationRed-flag recall, false negative sample rate, SAR rework rate, audit completeness
Customer service RAGMedium to High错误政策承诺、投诉升级、隐私泄露、客户体验下降Approved source only, freshness check, citation requirement, PII masking, supervisor approval for high-risk topicsUnsupported claim rate, stale source hit, complaint escalation, source freshness SLA
Fraud decision supportHigh错误拦截交易、漏放欺诈、客户损害、运营拥堵Human decision for high-value cases, threshold monitoring, challenger model, appeal workflow, incident playbookFraud capture lift, false positive rate, customer appeal overturn rate, p95 decision latency
Wealth advisory guardrailHigh不当投资建议、suitability breach、客户损失、监管处罚Suitability rules, approved product universe, prohibited recommendation blocks, advisor approval, disclosure loggingSuitability breach count, prohibited advice block rate, advisor override rate, disclosure completeness
AI platform shared serviceHigh as enabling controlShadow AI、权限泄露、成本失控、不可审计的模型变更Model gateway, RBAC, policy-as-code, prompt/model registry, cost controls, centralized loggingRegistered app coverage, unapproved AI usage, audit log completeness, cost per business unit

4. Board-Level AI Oversight Questions

用途: 董事会、审计委员会和风险委员会用来挑战管理层 AI strategy, use case portfolio, controls 和 investment request。PM / BA / Architect 应提前准备事实底座和可追溯证据。

4.1 Template

FieldWhat to prepareConcrete sample
Decision topic本次需要董事会监督或批准什么"Approve risk appetite boundary for high-risk AI use cases in credit, AML, fraud and wealth advisory."
Material systems in scope哪些 AI 系统进入 material registerCredit underwriting assistant, AML copilot, fraud decision support
Business exposure受影响客户、交易、员工、资金或监管义务"Fraud decision support touches 2.4M monthly card transactions; AI recommends priority, not final block for transactions above $500."
Top risk question董事会最应该追问的一句话"Can management prove that AI-assisted credit recommendations do not create prohibited bias or unexplainable adverse decisions?"
Evidence to answer证据包清单Bias test, adverse action explanation sample, override log, policy mapping, audit trail sample
Control effectiveness metric控制是否有效的指标"100% adverse action notices include approved reason code; monthly disparate impact ratio reviewed by model risk."
Residual risk控制后仍存在的风险"Medium residual risk: borderline applications may still receive inconsistent human overrides."
Stop rule触发暂停或上报的条件"Pause expansion if adverse action explanation completeness falls below 99%, or protected-class disparity exceeds approved tolerance for two review cycles."

4.2 Board Question Bank

Oversight questionStrong management answer pattern
Which AI systems are material and why?"We classify materiality by customer impact, financial exposure, regulatory obligation, autonomy level, data sensitivity and reversibility. This quarter 6 of 18 AI systems are material."
Are any AI tools making final decisions?"No final customer-impacting credit, AML, fraud or wealth decisions are fully automated. AI drafts, recommends or prioritizes; named humans approve final actions above defined thresholds."
How do we know controls work, not just exist?"Each control has an owner, test frequency, metric and evidence artifact. For AML copilot, red-flag recall is tested monthly against 500 labelled cases and reviewer calibration is sampled weekly."
What AI risk exceeds appetite?"Customer service RAG is within appetite. Wealth advisory guardrail remains outside appetite for direct-to-customer recommendations; it is limited to advisor-facing compliance checks."
What incidents or near misses occurred?"One medium incident: customer service RAG cited a stale fee policy in 31 drafts. It was contained in 42 minutes, no final customer messages were sent without review, source refresh control was remediated."
How are we preventing shadow AI?"All production AI calls must route through model gateway. Unapproved browser AI uploads are blocked by DLP, and business units attest quarterly to no unregistered high-risk AI use."
What is the investment logic?"We fund shared platform controls because they reduce per-use-case audit, security and eval cost. We fund domain use cases only when business owner, data owner and eval path are confirmed."
What would cause management to stop?"Pre-agreed stop rules cover safety, compliance, adoption, unit economics and control failure. Stop decisions do not require waiting for annual planning."

4.3 PM / BA / Architect Preparation

RolePreparation actionBoard language
PMMap business outcome, adoption, ROI and stop threshold"This investment moves fraud loss and false positive rate, not model novelty."
BADocument AS-IS / TO-BE workflow, decision owner, exception path"AI changes triage speed; it does not change who owns customer-impacting action."
ArchitectShow data flow, model gateway, audit log, rollback, NFR"The design is reversible, permissioned, logged and testable."

5. AI Portfolio Risk Dashboard

用途: 每季度向 board / risk committee 汇报 AI portfolio 的价值、风险、控制成熟度和投资取舍。Dashboard 应该让董事看到 "哪些 AI 值得继续投, 哪些需要限制, 哪些必须停止"。

5.1 Dashboard Fields

FieldDefinitionExample
System name注册系统名称AML Copilot
Business owner业务结果负责人Head of Financial Crime Operations
AI roleDraft / recommend / triage / approve / executeDraft SAR narrative and prioritize alerts
Materiality tierCritical / High / Medium / LowHigh
Current stageDiscovery / Pilot / Limited production / Production / RetiredLimited production
Value metric与业务基线绑定的指标Analyst review time reduced from 42 min to 31 min
Risk ratingInherent and residual riskInherent High, residual Medium
Control maturity1-5 分, 基于证据4: controls tested monthly, audit log complete
Policy exceptions已批准例外数量1 exception: temporary manual upload for legacy case docs
Incidents / near misses本季度事件0 severe, 2 low near misses
Audit evidence statusComplete / partial / missingComplete for release gate and monitoring
Investment decisionScale / hold / remediate / stopHold until reviewer calibration reaches target
Stop rule statusGreen / amber / redAmber: false positive rate above target

5.2 Sample Dashboard

SystemStageValueResidual riskControl maturityIncident trendInvestment decisionStop rule
Credit underwriting assistantPilotUnderwriter cycle time -18%High30Remediate before scaleRed if disparate impact ratio outside tolerance
AML copilotLimited productionReview time -26%, SAR rework -11%Medium40 severeHold and expand only to trained analystsStop if red-flag recall below 97%
Customer service RAGProductionAHT -21%, citation completeness 98.7%Medium41 mediumScale to second product lineStop if unsupported claim rate above 3%
Fraud decision supportPilotFraud capture +7%, false positives +1.2%High30Limited scale with appeal monitoringStop if appeal overturn rate above 8%
Wealth advisory guardrailDiscoveryNo production value yetHigh20Continue discovery, no customer-facing useStop if suitability rule coverage below 95%
AI platform shared serviceProduction73% registered AI calls via gatewayMedium40Invest in audit export and cost controlsStop onboarding if log completeness below 99%

5.3 Metrics and Thresholds

CategoryMetricGreenAmberRed
Business valueValue metric vs approved baseline>= 80% of target50-79%< 50%
AdoptionWeekly active target users>= 70%40-69%< 40%
AI qualityUnsupported claim rate<= 2%> 2% and <= 3%> 3%
RiskHigh severity incident01 containedAny uncontained severe incident
AuditabilityComplete decision log>= 99%95-98.9%< 95%
CostCost per assisted case<= approved cap<= 120% cap> 120% cap
ControlControl test pass rate>= 95%85-94%< 85%

5.4 Stop Rule

Portfolio-level stop rule:

An AI initiative is paused or de-prioritized when it has no accountable business owner, no data owner, no testable control evidence, red stop-rule status for two review cycles, material incident without containment, or unit economics above approved cap with no credible remediation path.

6. Three Lines of Defense

用途: 明确 AI governance 中业务、风险合规、内审的职责边界。董事会最关心的是 "谁在第一线拥有风险, 第二线如何挑战, 第三线如何独立验证"。

6.1 Operating Model

LineOwnerResponsibilitiesEvidence
First lineBusiness owner, PM, BA, Architect, Operations, Data ownerUse case definition, process design, control operation, user training, monitoring, incident first responseAS-IS / TO-BE workflow, eval result, logs, SOP, training record, control owner sign-off
Second lineRisk, Compliance, Legal, Privacy, Model Risk, CISOPolicy, risk appetite, challenge, approval, exception review, control standards, independent monitoringRisk assessment, policy mapping, exception memo, control test review, compliance opinion
Third lineInternal AuditIndependent assurance over governance, control design, operating effectiveness, evidence integrityAudit plan, walkthrough, sample testing, findings, management action validation

6.2 Concrete RACI by Case

ActivityFirst lineSecond lineThird line
Credit underwriting assistant bias testCredit product and model team execute monthly testFair lending / model risk challenge thresholdsAudit samples test evidence and sign-off trace
AML copilot release gateFinancial crime operations owns workflow and SAR reviewCompliance approves red-flag recall and SAR quality thresholdAudit verifies case log reconstructability
Customer service RAG source refreshKnowledge owner certifies policy source weeklyCompliance reviews regulated content categoriesAudit tests source freshness evidence
Fraud decision support appeal monitoringFraud operations reviews false positives and appealsOperational risk challenges customer harm metricsAudit samples overturned decisions
Wealth advisory guardrail suitability rulesWealth product owner maintains product and client rule mappingLegal/compliance approves prohibited advice controlsAudit validates rule change approval
AI platform shared service model gatewayPlatform owner operates gateway, logging and cost guardrailsCISO/privacy define security and data controlsAudit tests access, logging and change management

6.3 Stop Rule

No high-risk AI system moves from pilot to production unless first line operates named controls for at least one full review cycle, second line records challenge and approval, and third line has no blocking evidence-integrity concern for audit-critical controls.

6.4 Metrics

MetricTargetBoard interpretation
First-line control owner coverage100% material systems每个关键控制都有业务或技术责任人, 没有 "治理团队代持风险"
Second-line challenge completion100% high-risk releases风险、合规、法律、隐私或模型风险已形成独立挑战记录
Third-line evidence reconstructability>= 95% sampled cases内审可独立重建 AI-assisted decision, 不是依赖口头解释
Open second-line conditions past due0 high severity管理层没有带着过期风险条件继续扩张
Repeat control failures0 for critical controls控制问题不是偶发缺陷, 而是需要 board 关注的治理失效信号

7. AI Policy Approval Memo

用途: AI governance committee 或 board committee 批准企业 AI policy, 包括 acceptable use, prohibited use, risk classification, approval gates, monitoring, incident reporting 和 audit evidence。

7.1 Template

SectionContent
Decision requestedApprove enterprise AI policy v1.0 effective 2026-07-15, covering employee AI use, customer-impacting AI, vendor AI, model lifecycle, data handling and incident reporting.
Policy scopeApplies to all AI systems that draft, recommend, score, summarize, triage, approve or execute work using customer, employee, transaction, product, risk or proprietary data.
Prohibited usesUnapproved customer-facing financial advice, fully automated credit denial without approved model governance, upload of restricted data to unapproved AI tools, AI execution of payments without dual control, hidden model change in production.
Risk classificationCritical, High, Medium, Low based on customer impact, financial exposure, regulatory obligation, autonomy, data sensitivity, explainability and reversibility.
Required gatesIntake, materiality classification, data review, architecture review, eval gate, security/privacy review, risk approval, release readiness, monitoring review, retirement.
Control requirementsApproved sources, RBAC, human accountability, audit log, eval thresholds, monitoring, incident runbook, vendor due diligence, model/prompt versioning.
Exception processExceptions require business sponsor, risk owner, compensating controls, expiry date, monitoring metric and board/risk committee reporting if material.
MetricsRegistered AI coverage, unapproved AI detections, high-risk use case review completion, control test pass rate, incidents, exception aging, audit findings.
Management ownerChief AI Governance Officer or equivalent; operating owners include CIO/CTO, CRO, CISO, CDO, Legal, business executives.
Board reportingQuarterly AI governance report, immediate escalation for severe AI incident or material policy breach.
Stop ruleAny AI system outside policy gates is blocked from production; repeated unapproved AI usage in a business unit triggers management attestation and remediation review.

7.2 Sample Policy Approval Summary

Recommendation:
Approve the AI Policy v1.0 with a 90-day implementation window. The policy enables controlled AI adoption while prohibiting unapproved customer-impacting decisions and restricted-data uploads.

Material changes:
1. All high-risk AI systems must be registered before pilot.
2. AI cannot make final credit, AML, fraud or wealth decisions without approved human accountability and model governance.
3. Production AI must use approved logging, versioning and incident process.
4. Exceptions expire within 90 days unless risk committee renews them with evidence.

Expected board benefit:
The policy gives the board a clear line of sight into material AI risk, control ownership and exception exposure.

7.3 PM / BA / Architect Evidence

RoleEvidence to attach
PMUse case intake form, product scope, adoption metric, investment gate
BAWorkflow decision map, control touchpoints, exception path, operating SOP
ArchitectReference architecture, data flow, access model, logging schema, rollback design

8. Material AI System Register

用途: 建立董事会和审计委员会可监督的 AI inventory。关键不是列所有 demo, 而是识别 material AI systems and material dependencies。

8.1 Register Fields

FieldDefinitionExample
System IDUnique identifierAI-AML-002
System nameBusiness-readable nameAML Copilot
Business processSupported workflowAlert investigation and SAR drafting
AI roleDraft / recommend / triage / score / executeDraft and triage
Autonomy level0 manual to 5 autonomous execution2: recommendation with mandatory review
Customer impactDirect / indirect / internal onlyIndirect customer impact
Regulatory domainCredit, AML, privacy, wealth, fraud, operationsAML / BSA
Data classificationPublic / internal / confidential / restrictedRestricted customer and transaction data
Model / vendorInternal, vendor, model familyVendor LLM through approved gateway
Key data sourcesAuthoritative dataCase notes, transaction alerts, KYC profiles, AML policy
Control ownerNamed accountable roleFinancial Crime Ops Control Lead
Risk ownerNamed second-line ownerAML Compliance Director
Audit log locationEvidence sourceCentral AI audit lake table ai_case_log
Eval cadenceTest frequencyMonthly golden set and weekly production sampling
Incident severity pathEscalation forumCRO and Risk Committee for severe event
StatusDiscovery, pilot, production, retiredLimited production
Last review dateGovernance review date2026-06-15
Next review dateScheduled review2026-09-15

8.2 Sample Register Entries

IDSystemAI roleMaterialityResidual riskStatusStop rule
AI-CRED-001Credit underwriting assistantRecommend with human approvalHighHigh during pilotPilotPause if adverse action explanation completeness < 99%
AI-AML-002AML copilotDraft SAR narrative and triage alertsHighMediumLimited productionPause if red-flag recall < 97% or audit log completeness < 99%
AI-CS-003Customer service RAGDraft customer responseMediumMediumProductionPause regulated-topic drafting if unsupported claim rate > 3%
AI-FRAUD-004Fraud decision supportScore and recommend queue priorityHighHighPilotPause if appeal overturn rate > 8% or p95 latency > SLA
AI-WEALTH-005Wealth advisory guardrailBlock prohibited recommendationHighMediumDiscoveryNo production if suitability rule coverage < 95%
AI-PLAT-006AI platform shared serviceGateway, logging, eval and policy enforcementHigh dependencyMediumProductionStop onboarding if log completeness < 99%

8.3 Stop Rule

Any AI system that meets materiality criteria but is not registered, lacks a named business owner, lacks audit log location, or lacks next review date is barred from production access and must be reported as a governance gap in the quarterly AI governance report.

8.4 Metrics

MetricTargetSample board readout
Material system registration coverage100%"All 6 material AI systems are registered with named owners."
Register freshnessReviewed within 90 days"One medium-risk system review is due in 12 days; no overdue material reviews."
Owner completeness100% business, risk, technology, data owners"AI-WEALTH-005 added named compliance owner before discovery exit."
Audit log location completeness100% material systems"All material systems point to system-of-record log locations."
Unregistered material AI detections0"Two shadow AI detections were low-risk productivity tools, not material systems."

9. High-Risk Use Case Exception Memo

用途: 当高风险用例希望在某些控制尚未完全成熟时进入 pilot 或 limited production, 必须用 exception memo 明确边界、补偿控制、期限和停止规则。

9.1 Template

SectionContent
Decision requestedApprove a 60-day limited exception for Fraud Decision Support pilot to use manual evidence upload for one legacy queue while platform connector is completed.
Use caseFraud decision support recommends case priority for debit card disputes above $100 and below $2,500. It does not auto-deny or auto-reimburse.
Why exception is neededLegacy case documents are not yet connected to the approved retrieval pipeline; without temporary upload, pilot cannot test high-volume dispute workflow.
Inherent riskHigh: customer funds access, fraud loss, false positives, operational backlog, sensitive transaction data.
Compensating controlsManual upload limited to trained fraud analysts; PII masking script runs before upload; files deleted within 24 hours; audit log captures uploader, case ID, document hash and reviewer; daily sample review by control lead.
Exposure limitMaximum 1,000 cases, one queue, 60 days, no automated customer action, no data export outside approved gateway.
MetricsFalse positive rate, fraud capture lift, appeal overturn rate, unsupported recommendation rate, PII masking failure, audit log completeness.
Residual riskMedium-high for 60 days because manual upload creates operational and privacy control dependency.
OwnerBusiness owner: VP Fraud Operations; Risk owner: Operational Risk Director; Technology owner: AI Platform Lead.
ExpiryException expires 2026-09-30 and cannot renew without connector delivery evidence or risk committee re-approval.
Stop ruleStop immediately if PII masking failure occurs, audit log completeness < 99%, appeal overturn rate > 8%, or any analyst uploads documents outside approved queue.

9.2 Board-Ready Summary

This exception does not approve autonomous fraud decisions. It approves a narrow, time-boxed data-handling exception for a pilot queue so management can measure value and risk. The residual risk is higher than normal platform operation, but bounded by volume, duration, human approval, deletion, logging and daily sampling.

9.3 PM / BA / Architect Preparation

RoleWhat to prepare
PMBusiness value if exception is approved, lost learning if denied, user cohort, adoption metric
BAExact workflow step requiring exception, manual control, reviewer role, evidence capture
ArchitectData path diagram, temporary control design, deletion proof, migration path to approved connector

10. Quarterly AI Governance Report

用途: 每季度向 board / audit committee / risk committee 汇报 AI governance 状态。重点不是罗列项目, 而是说明管理层是否知道风险在哪里、控制是否有效、投资是否集中。

10.1 Report Structure

SectionContentSample
Executive summary本季度总体结论"AI portfolio remains within approved risk appetite except credit underwriting assistant, which remains in remediation before scale."
Portfolio inventory系统数量和阶段24 registered AI initiatives: 6 material, 8 production, 5 pilots, 7 discovery, 4 retired
Material risk changes风险上升或下降"Customer service RAG moved from amber to green after source freshness control reached 99.4%."
Control effectiveness关键控制测试结果"Audit log completeness: 99.2%; high-risk release gate completion: 100%; exception aging: 1 item over 60 days."
Incidents and near misses事件、影响、整改"One medium incident; stale policy citation contained before final customer send."
Investment and value投资、ROI、扩张/停止决策"Two pilots stopped due to low adoption; platform gateway funding recommended to reduce duplicated controls."
Regulatory and audit posture审计发现、监管关注、证据质量"Internal audit found one medium issue in vendor model-change notification; remediation due 2026-08-15."
Next quarter decisions董事会需要关注的决策"Approve risk appetite for advisor-facing wealth guardrail; decide whether to scale AML copilot to second region."

10.2 Metrics

MetricCurrent quarterPrior quarterStatus
Registered AI initiatives2417Amber: inventory grew faster than review capacity
Material AI systems64Green: all have named owners
High-risk systems with completed release gate100%75%Green
Audit log completeness for material systems99.2%96.8%Green
Open high-risk exceptions21Amber
Severe AI incidents00Green
Medium AI incidents12Green trend
Unapproved AI usage detections914Green trend
AI spend against approved budget92%88%Green
Initiatives stopped or merged41Green: portfolio discipline improving

10.3 Stop Rule

The quarterly report must recommend a portfolio hold if material systems lack owners, audit evidence completeness falls below 95%, severe AI incident remediation remains open past target date, high-risk exceptions age beyond approved expiry, or unregistered AI usage increases for two consecutive quarters.

11. AI Incident Board Update

用途: 当 AI incident 可能造成客户、监管、财务、隐私或声誉影响时, 给 board / audit committee / risk committee 的紧急更新。它必须降低不确定性: 已知什么、未知什么、已控制什么、何时更新。

11.1 Template

SectionContent
Incident titleCustomer Service RAG stale policy citation incident
SeverityMedium; no confirmed customer financial loss; potential regulatory complaint exposure
Time detected2026-06-18 09:20 CT by production sampling alert
System and scopeCustomer service RAG, credit card fee dispute policy responses, 31 drafted responses across 14 agents
AI roleDraft response with mandatory agent review; no autonomous send
Confirmed impact7 drafts contained stale policy citation; 0 were sent to customers without human edit; 2 customer cases require supervisor review
Potential exposureUp to 31 cases drafted during 08:40-09:20 CT window
Containment09:35 disabled drafting for fee dispute category; read-only retrieval remains active; all affected drafts quarantined
Root cause statusPreliminary: policy source refresh job failed after document path change; monitoring detected stale source after sampling, not before retrieval
Customer/regulator actionNo regulator notification required based on current facts; Legal reassessment at 16:00 CT after case review
MetricsTime to detect 40 min; time to contain 15 min; affected drafts 31; unsupported claim rate during window 22.6%; audit log completeness 100%
Management decision neededApprove keeping category-level drafting paused until freshness control passes regression test
Next update2026-06-18 16:00 CT with final case review and restart recommendation
Stop ruleFeature remains paused until source refresh control passes 3 consecutive checks, affected cases are reviewed, and risk owner signs restart.

11.2 Incident Update Language

The incident is contained. The AI system drafted stale policy language, but the human review gate prevented automated customer communication. The management decision is to keep the affected category paused until the source freshness control and regression eval pass. Current residual risk is low for active operations because the feature is disabled for the affected category.

11.3 Board Questions to Expect

QuestionAnswer
Did customers receive incorrect advice?"Current evidence shows no unedited AI draft was sent. Two cases require supervisor review because agents partially reused draft structure."
Why did monitoring not prevent it?"The source refresh control detected failure after sampling, not before retrieval. We are moving the control to pre-retrieval freshness validation for regulated categories."
Could this happen in AML or credit?"The same source refresh mechanism is not used for AML case evidence or credit bureau data. However, the platform team is reviewing all material systems for pre-use freshness checks."
When can it restart?"Only after three checks pass, affected cases are reviewed, and risk owner signs restart. No business owner can override that restart gate."

12. Audit Evidence Request List

用途: 审计委员会和 internal audit 需要确认 AI governance controls 是否设计合理、运行有效、证据完整。PM / BA / Architect 应该按系统建立 evidence pack, 而不是在审计来时临时收集截图。

12.1 Evidence Categories

CategoryEvidence requestExample evidence
GovernanceAI policy approval, committee minutes, risk appetite, decision rightsAI Policy v1.0 approval memo, risk committee minutes, RACI
InventoryMaterial AI system register, classification rationale, owner listRegister export with 6 material systems and next review dates
Use case approvalIntake form, business case, materiality assessment, approval recordFraud decision support intake and risk approval
Data controlsData classification, source owner sign-off, lineage, retention, deletionCustomer service RAG policy source certification and refresh logs
ArchitectureData flow, access model, model gateway design, logging schema, rollbackAI platform C4 diagram, audit log schema, DR runbook
Eval and validationGolden set, test results, thresholds, red-team, regression historyAML copilot red-flag recall report and reviewer calibration
Human oversightApproval matrix, training records, override logs, reviewer QACredit underwriting assistant approval and adverse action review
MonitoringProduction dashboards, alerts, sampling evidence, incident triggerUnsupported claim monitoring for customer service RAG
Change managementPrompt/model version history, release approvals, rollback evidenceModel gateway version registry and release ticket
Vendor riskDue diligence, DPA, subprocessor list, model change notice, SLAVendor security review and contract clauses
Incident managementIncident logs, timelines, containment actions, postmortem, remediationStale policy incident board update and closure evidence
AttestationManagement certifications and exception confirmationsQuarterly management attestation pack

12.2 Sample Audit Request for AML Copilot

RequestPass evidence
Show how the system was approved for limited productionAI governance committee approval dated 2026-05-10, risk committee conditions, release checklist
Prove AI does not file SAR automaticallyWorkflow map, system permission settings, case samples showing human approval
Reconstruct 10 sampled AI-assisted casesInput, retrieved evidence, prompt version, model version, draft, reviewer, final SAR narrative, timestamp
Demonstrate red-flag recall controlMonthly golden set, 500 labelled cases, threshold 97%, actual 98.4%, exception log
Show reviewer calibrationWeekly QA sample, reviewer variance report, training completion
Show incident readinessRunbook, last tabletop exercise, escalation contacts, kill switch evidence

12.3 Stop Rule

If audit cannot reconstruct sampled material AI decisions end-to-end, or if evidence is stored only in manual screenshots without system-of-record traceability, the system cannot expand scope until audit evidence quality is remediated.

12.4 Metrics

MetricTargetSample evidence result
Evidence request fulfillment>= 95% by due date"38 of 40 evidence requests fulfilled; 2 vendor SLA artifacts due in 5 days."
End-to-end case reconstructability>= 95% sampled material cases"AML copilot reconstructed 20 of 20 sampled cases from input to final SAR narrative."
Evidence system-of-record coverage>= 90%"Most evidence comes from logs and approval records; screenshots limited to vendor portal status."
Repeat audit findings0 high / critical"Vendor model-change notification gap is a first-time medium issue."
Evidence owner completeness100%"Every audit request maps to a named business, tech, data or risk owner."

13. Management Attestation Pack

用途: 管理层每季度向 Audit Committee / Risk Committee 证明 material AI systems 的控制、例外、事件和风险声明是真实、完整、可追溯的。

13.1 Attestation Components

ComponentSignerStatement
Business owner attestationBusiness executive"The AI system is used only within approved workflow scope, and business outcomes and user adoption metrics are accurately reported."
Risk owner attestationCRO delegate / compliance owner"Residual risk is assessed, exceptions are current, and controls remain within approved risk appetite or are escalated."
Technology owner attestationCIO/CTO delegate"Logging, access control, versioning, monitoring and rollback controls operated as described for the reporting period."
Data owner attestationCDO delegate / domain data owner"Approved sources, data classification, retention and freshness controls operated as required."
Security/privacy attestationCISO / DPO delegate"No unapproved restricted-data flow or unresolved critical security/privacy issue exists for the material AI systems in scope."
Internal audit observationCAE or audit lead"Audit has reviewed evidence availability and notes open findings separately; this is not a management sign-off."

13.2 Pack Fields

FieldExample
Reporting periodQ2 2026
Systems coveredAI-CRED-001, AI-AML-002, AI-CS-003, AI-FRAUD-004, AI-WEALTH-005, AI-PLAT-006
Systems excludedNone from material register
Open exceptionsFraud manual upload exception, expires 2026-09-30; wealth advisory direct-to-customer restriction remains active
Control failuresCustomer service RAG stale policy freshness failure, remediated 2026-06-25
Incidents1 medium, 0 severe
Audit findings1 medium vendor model-change notification gap, due 2026-08-15
Residual risk changesCredit underwriting assistant remains high until bias and explanation controls pass two cycles
Management conclusionMaterial AI governance controls are operating with two amber items and no red items as of 2026-06-30

13.3 Sample Attestation Statement

For Q2 2026, management attests that all material AI systems are registered, have named business, risk, technology and data owners, and were operated within approved scope except for the disclosed fraud manual upload exception. No AI system made final automated credit, AML, fraud or wealth advisory decisions outside approved human accountability controls. One medium incident occurred and was contained; remediation evidence is attached. Management rates the overall residual AI portfolio risk as Medium with two amber items requiring Q3 follow-up.

13.4 Stop Rule

Management cannot provide a clean attestation if any material AI system lacks owner sign-off, has expired exception approval, has unresolved severe incident remediation, has audit log completeness below 95%, or operated outside approved workflow scope.

13.5 Metrics

MetricTargetSample attestation readout
Owner sign-off completion100% material systems"All 6 material systems signed by business, risk, technology and data owners."
Qualified attestations0 red, amber explained"Two amber qualifications: fraud exception and vendor notification gap."
Expired exceptions0"No exception passed expiry without risk committee review."
Severe incident remediation past due0"No severe AI incidents in Q2; one medium remediation closed on schedule."
Attestation-to-evidence traceability>= 95% claims linked to evidence"All critical statements link to register, logs, control tests or committee minutes."

14. Policy, Exception and Reporting Templates by Use Case

14.1 Credit Underwriting Assistant

Template fieldContent
Board riskFair lending, explainability, adverse action, credit policy consistency
Control evidenceBias test by protected-class proxy where legally appropriate, adverse action reason mapping, underwriter approval logs, override analysis
Business metricUnderwriter cycle time, application backlog, manual rework rate
Risk metricDisparate impact ratio, unsupported recommendation rate, adverse action explanation completeness
Investment decisionScale only after two review cycles of stable fairness and explanation metrics
Stop rulePause expansion if adverse action explanation completeness < 99%, unsupported recommendation rate > 2%, or bias metric exceeds approved tolerance

14.2 AML Copilot

Template fieldContent
Board riskMissed suspicious activity, weak SAR evidence, regulatory criticism
Control evidenceGolden set red-flag recall, SAR quality sample, reviewer calibration, case reconstruction
Business metricAlert review time, backlog age, SAR rework rate
Risk metricRed-flag recall, false negative sampling, escalated QA defects
Investment decisionHold expansion if analyst review quality declines despite time savings
Stop rulePause if red-flag recall < 97%, SAR evidence citation completeness < 98%, or reviewer calibration variance exceeds threshold

14.3 Customer Service RAG

Template fieldContent
Board riskIncorrect customer commitments, stale policy, privacy leakage
Control evidenceApproved source list, freshness log, citation completeness, PII masking test, complaint review
Business metricAverage handling time, first contact resolution, rework, complaint escalation
Risk metricUnsupported claim rate, stale source hit rate, regulated-topic error rate
Investment decisionScale by product line only after source owner and supervisor workflow are stable
Stop ruleDisable affected category if unsupported claim rate > 3%, source freshness SLA < 99%, or PII masking failure occurs

14.4 Fraud Decision Support

Template fieldContent
Board riskCustomer harm from false positives, fraud loss from false negatives, operational backlog
Control evidenceThreshold review, appeal overturn analysis, human approval for high-value cases, challenger comparison
Business metricFraud capture lift, loss avoided, case handling time
Risk metricFalse positive rate, appeal overturn rate, p95 decision latency, high-value case approval completeness
Investment decisionInvest only if fraud lift does not create unacceptable customer friction
Stop rulePause if appeal overturn rate > 8%, false positive rate increases above approved tolerance, or p95 latency breaks SLA for two days

14.5 Wealth Advisory Guardrail

Template fieldContent
Board riskUnsuitable recommendation, prohibited product suggestion, disclosure failure
Control evidenceSuitability rule coverage, approved product universe, prohibited advice block tests, advisor approval logs
Business metricAdvisor review time, compliance review throughput, prevented unsuitable drafts
Risk metricSuitability breach count, prohibited advice escape rate, disclosure completeness
Investment decisionAdvisor-facing guardrail can pilot; direct-to-customer advice remains prohibited without separate approval
Stop ruleNo production if suitability rule coverage < 95%; pause if any prohibited advice reaches customer without advisor correction

14.6 AI Platform Shared Service

Template fieldContent
Board riskShadow AI, inconsistent controls, audit gap, cost overrun, vendor dependency
Control evidenceGateway routing coverage, model/prompt registry, audit log completeness, RBAC test, cost dashboard
Business metricReuse rate, time from intake to pilot, avoided duplicate vendor spend
Risk metricUnapproved AI detections, log completeness, unauthorized access attempts, model change regression pass rate
Investment decisionFund platform controls when they reduce repeated per-use-case risk work and improve auditability
Stop ruleStop onboarding new high-risk systems if log completeness < 99%, RBAC test fails, or model change regression gate is bypassed

15. 30 / 60 / 90 Day AI Governance Cadence Plan

15.1 First 30 Days: Establish Visibility and Minimum Gates

WorkstreamDeliverableOwnerSuccess metric
InventoryMaterial AI system register v1AI governance lead + PMs100% known production and pilot AI systems registered
PolicyAI policy approval memoRisk + Legal + CIO/CTOPolicy approved with exception process and stop rules
Risk classificationMateriality rubricRisk + BA + ArchitectAll registered systems classified by impact and autonomy
Board reportingAI portfolio risk dashboard v1PMO + RiskTop 10 AI initiatives scored by value, risk, controls
EvidenceAudit evidence request listInternal Audit + ArchitectsEvidence owner assigned for each material system
IncidentAI incident severity matrix and escalation pathCISO + Risk + OperationsTabletop exercise scheduled and owners confirmed

30-day stop rule:

No new high-risk AI pilot starts until it is registered, classified, has named owners, and has an approved stop rule.

15.2 First 60 Days: Test Controls and Exception Discipline

WorkstreamDeliverableOwnerSuccess metric
Control testingControl effectiveness test for each material AI systemFirst line + second line90% critical controls tested with evidence
ExceptionsHigh-risk use case exception memo processRisk committee100% exceptions have expiry, compensating control and metric
EvalOpsGolden set and production sampling standardsAI platform + model riskEach material system has quality and risk thresholds
PlatformModel gateway and audit log coverage planArchitect + platform owner80% production AI traffic through approved gateway
TrainingBoard and management AI risk briefingAI governance leadDirectors and executives receive common terminology pack
AttestationDraft management attestation packBusiness, risk, tech, data ownersSigners confirmed and evidence gaps listed

60-day stop rule:

Any material AI system with untested critical controls, expired exception, or missing audit log evidence is frozen at current scope until remediation is complete.

15.3 First 90 Days: Board Cadence and Assurance

WorkstreamDeliverableOwnerSuccess metric
Board reportingQuarterly AI governance reportCRO / CIO / AI governance leadReport delivered with decisions, not project status only
Audit readinessEvidence pack for each material systemFirst line + Internal AuditAudit can reconstruct sampled decisions end-to-end
AssuranceInternal audit AI governance review scopeCAEAudit plan covers policy, inventory, controls and evidence quality
Investment governancePortfolio scale / hold / stop recommendationsManagement steeringAt least one low-value or uncontrolled initiative stopped or merged
Incident readinessCompleted AI incident tabletopCISO + Risk + OperationsLessons remediated and board escalation template tested
AttestationSigned management attestation packBusiness, risk, tech, data ownersClean or qualified attestation submitted with amber/red items

90-day stop rule:

If management cannot produce a quarterly report, material register, control evidence, incident path and attestation pack by day 90, board should pause expansion of high-risk AI until governance cadence is operating.

16. Common Board Questions and Answers

Board questionStrong answer
Are we using AI faster than our controls can handle?"We separated low-risk productivity AI from material AI systems. High-risk systems cannot expand unless release gates, audit logs and monitoring pass. Portfolio dashboard shows two systems held for remediation."
Which AI use case could create the largest customer harm?"Credit underwriting assistant and fraud decision support have the highest customer-impact risk. Both are limited to human-approved recommendations and have stop rules tied to fairness, explanations, appeals and false positives."
Can we prove AI did not make an unapproved final decision?"For material systems, audit logs capture user, input, source, model version, output, reviewer, decision and final action. Audit can reconstruct sampled cases end-to-end."
What happens if a vendor changes its model?"Material vendor model changes require notice, regression eval, release approval and rollback readiness. A vendor model-change notification gap is open as a medium audit issue with due date."
How are we managing regulatory expectations?"We map each material use case to domain obligations: fair lending, AML, privacy, fraud operations, suitability and model risk. Compliance and legal approve risk classification and exceptions."
Are we getting value or just running experiments?"Every funded pilot has baseline, target metric, adoption threshold, cost cap and stop rule. This quarter two pilots were stopped for low adoption and one platform investment is recommended for control reuse."
How do employees know what AI use is prohibited?"The policy blocks restricted-data upload to unapproved tools, prohibits unapproved customer-facing financial advice, and requires registration for customer-impacting AI. Training and DLP monitoring support enforcement."
What is our residual risk posture?"Portfolio residual risk is Medium overall, with credit and fraud remaining High during pilot. The risk is accepted only within limited scope and review dates, not as open-ended approval."
Do we have enough talent to govern this?"The bottleneck is not data science only; it is product, BA, architecture, risk and audit evidence capacity. The 90-day plan adds reusable platform controls and standard evidence packs to reduce manual governance load."
What should the board decide this quarter?"Approve AI policy v1.0, endorse risk appetite boundaries for high-risk use cases, fund platform audit controls, and require remediation before credit underwriting assistant scales."

17. Common Pitfalls

PitfallWhy it is dangerousBetter practice
Treating board AI update as innovation showcase董事会看不到 material risk, ownership 和 control evidenceLead with risk, controls, residual risk and decision needed
Reporting model accuracy onlyAccuracy 不能证明客户、合规、审计和运营风险可控Include business, quality, risk, adoption, cost and audit metrics
Saying "human-in-the-loop" without role detail没有说明谁在何处负责, 审计无法重建责任Define approval step, authority, override reason and evidence
Letting pilots run without stop rulePilot 会变成事实生产系统, 风险和成本失控Pre-approve stop, extend and scale criteria
Ignoring shadow AI员工可能把 restricted data 放进未批准工具Combine policy, DLP, training, model gateway and attestation
Treating vendor demo as control evidenceDemo 不能证明安全、审计、可用性或监管适配Require customer-specific eval, contract controls and audit export
Making AI governance purely technical风险真正发生在业务流程和客户结果中PM/BA own workflow and business controls; Architect owns technical enforceability
Over-centralizing all decisions中央团队成为瓶颈, 业务风险无人真正拥有Central standards, local first-line ownership, second-line challenge
Under-defining residual risk"风险可控" 无法被董事会监督State residual risk level, owner, rationale, metric and review date
No management attestation董事会无法确认报告完整性和问责链Quarterly attestation by business, risk, technology, data and security owners

18. Preparation Checklist for PM / BA / Architect

在进入 board / audit / risk committee 前, 团队应完成以下检查:

CheckPass condition
Decision clarity第一页写清 approve, hold, scale, stop, accept risk, fund, or remediate
Materiality系统已按客户影响、监管、资金、自治度、数据敏感性分类
WorkflowBA 能说明 AI 改变了哪个步骤, 未改变哪个最终责任
Business riskPM 能把每个 AI failure mode 翻译成客户、财务、合规、运营或声誉风险
Control effectiveness每个关键控制有 owner, frequency, threshold, evidence
Residual risk风险等级、接受理由、owner、review date 明确
AuditabilityArchitect 能证明一次决策可从输入重建到最终动作
Investment logic价值、成本、机会成本、平台复用和停止条件明确
Case coverage六类金融零售案例至少在 register 或 portfolio dashboard 中被评估
Stop rule每个 pilot, exception, production system 和 portfolio decision 都有可执行停止条件

掌握标准:

你能把同一个 AI 系统同时讲成业务投资、风险暴露、控制体系、审计证据、架构选择和董事会监督议题, 并且每一种讲法都指向同一组事实和同一套停止规则。