返回 Papers
AI 底层逻辑 / 经典论文

AI Exception / Risk Acceptance:例外治理架构

一句话:

228ai-foundations/papers/112-ai-exception-risk-acceptance-waiver-architecture.md

AI Exception / Risk Acceptance / Waiver Architecture 解读

面向对象: AI Product Lead / Senior BA / Product Architect / Enterprise Architect / Model Risk Partner / Operational Risk Lead / Audit Evidence Owner。 核心问题: AI 产品暂时不能满足标准控制时, 如何管理政策例外、临时 waiver、剩余风险接受、补偿控制、到期续期、证据、升级、董事会/审计可见性和硬停止条件。 学习目标: 区分 risk appetite 与 exception management: risk appetite 先定义组织愿意承担的 AI 风险边界, waiver architecture 管理的是偏离标准控制后的限时、限域、留证和退出机制。


Source Anchors

SourceLink用途
NIST AI RMFhttps://www.nist.gov/itl/ai-risk-management-framework用 Govern / Map / Measure / Manage 组织例外背景、风险测量、补偿控制和持续治理
ISO/IEC 42001https://www.iso.org/standard/42001用 AI management system 语言设计职责、运行控制、绩效评价、管理评审和持续改进
Federal Reserve SR 26-2https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm作为金融机构模型风险管理新锚点; SR 26-2 于 2026-04-17 替代 SR 11-7 和 SR 21-8
FFIEC IT Examination Handbook Management booklethttps://ithandbook.ffiec.gov/it-booklets/management.aspx用 IT governance、risk management、third-party、change、audit 和 board reporting 视角组织管理证据

一句话:

AI Exception Architecture 是把“暂时不能满足标准控制”转成有边界、有补偿、有到期、有证据、有升级、有硬停止条件的风险接受系统。


1. Thesis

Risk appetite 回答组织总体愿意承担哪些 AI 风险、哪些用途禁止、哪些用途有条件允许、不同 risk tier 的默认控制是什么。

AI exception / waiver 管理回答一个具体 use case 为什么暂时偏离标准控制、偏离哪条政策或控制、剩余风险由谁接受、用什么补偿控制限制风险、什么指标触发立即停止、到期时如何回到标准控制或停止。

所以 waiver 不是 risk appetite 的替代品。它是 risk appetite 已经定义之后, 对控制偏离的受控例外。


2. Why It Matters

AI 团队常见现实是: 产品价值明确, 但某些控制暂时不成熟。比如 golden eval set 仍在扩充, RAG citation checker 尚未覆盖全部文档格式, 第三方 vendor 证据还未自动接入 evidence binder, agent tool 已是 read-only 但 tool deny reason 尚未达到审计字段标准。

没有 waiver architecture 时, 组织会走向两个极端:

极端风险
全部阻塞低风险学习和真实反馈无法发生, 团队可能绕开治理
口头放行例外变成永久 shadow policy, 审计时无法解释谁接受了什么剩余风险

成熟机制允许有限试验, 但不允许例外失控。


3. Core Concepts

Concept含义AI 产品化要点
Exception对标准政策或控制的有限偏离绑定 control id、scope、duration、owner
Waiver正式批准的临时豁免不能无限期; 必须有到期、续期和退出路径
Risk acceptance管理层明确接受剩余风险接受的是 residual risk, 不是放弃控制
Compensating control用替代控制降低风险traffic cap、HITL、抽样、只读、披露
Residual risk补偿控制后仍存在的风险必须可描述、可监控、可升级
Hard stop触发立即暂停或回滚的条件预先写入 runbook
Shadow policy例外长期续期后变成事实政策通过 policy update、control investment 或 closure 处理

关键原则: 例外只能偏离控制, 不能绕过禁止用途; 必须限时、限量、限渠道、限用户或限数据; 审批权随 risk tier、客户影响和自动化程度上升; 例外数量、续期次数和过期未关本身就是治理信号。

GenAI / agentic AI 例外必须组合模型风险、操作风险、消费者合规、隐私、安全和第三方控制, 不能只按模型风险审批。


4. Architecture Diagram

Risk appetite and AI policy baseline
  -> standard control catalog
  -> AI use case / change request
  -> control gap
  -> exception request
  -> residual risk analysis
  -> compensating controls
  -> tiered approval
  -> restricted operation
  -> KRI dashboard and evidence binder
  -> expiry review
  -> close / renew / remediate / policy update / stop

Control plane:

Exception registry
  + policy/control IDs
  + risk tier and scope
  + residual risk owner
  + compensating controls
  + expiry and renewal rules
  + hard stop triggers
  + evidence links
  + board/audit reporting status

This should connect to release governance, incident management, model inventory, vendor risk, privacy review, security review, EvalOps and audit evidence.


5. Exception Taxonomy

TypeExampleTypical limit
Eval coverage高风险 slice 尚未完整覆盖limited pilot, sample review
Model validationindependent challenge 未完成最终报告shadow mode or capped release
Data/privacyredaction / retention 证据不足restricted data, no export
Securitygateway 缺少 tool deny attributeread-only mode, SIEM alert
Third-partyvendor SOC / SLA evidence 正在更新fallback route, traffic cap
Operationalhuman review SLA 未达标准lower traffic, queue monitoring
Evidenceevidence binder 自动化缺口manual evidence pack with owner
Policy interpretation新 agentic pattern 未被 policy 覆盖short expiry and policy update path

6. Financial Retail Case

场景: Retail bank customer-service AI copilot 从内部坐席草稿扩展到 limited customer-facing FAQ。

标准 high-tier 控制要求: source citation correctness >= 95%; complaint / fee dispute / credit commitment / account closure route to human; customer-facing answer uses approved disclosure; production trace includes model, prompt, source, policy and risk tier; write-enabled tools remain disabled。

例外请求: citation checker 对 PDF 表格型费用文档覆盖不足, 当前只能自动验证 88% 的引用; 团队希望对 5% 低风险 FAQ 流量做 30 天 pilot。

合格 waiver 决策:

AreaDecision
Scope只限低风险 FAQ, 不含投诉、费用减免、信贷承诺、账户关闭
Duration30 天, 不允许自动续期
Compensating controls每日人工抽样 100 条, RAG source allowlist, no-answer fallback, complaint hard-route
Residual risk客户可能收到不完整引用, 但高风险意图被排除并有人工抽样
Hard stopwrong fee commitment > 0; complaint escalation miss > 0; unsupported citation sample rate > 2%
Evidencewaiver memo, source registry, sample QA, trace dashboard, daily KRI, expiry decision
Exit pathPDF citation checker 修复后回到标准控制; 若未修复则停止 pilot

不合格做法: “业务压力大先上线再补”、“这个例外以后每月续一下”、“风险团队口头同意”、“客户投诉了再看”。


7. PM / BA / Architect Checklist

Product / BA:

  • 例外偏离的是哪条 policy、control 或 release gate。
  • 例外是否触碰 prohibited use; 如果触碰, 应直接 no-go。
  • business rationale 是否具体到客户价值、运营压力或监管期限。
  • scope 是否限制到用户、渠道、地区、数据、工具、模型或流量。
  • residual risk 是否说明客户、运营、合规和声誉影响。
  • compensating controls 是否能被实际执行, 而不是纸面承诺。
  • expiry date 是否明确, 且有 review forum。

Architect / risk:

  • runtime 是否能强制 scope 限制。
  • feature flag、tool gateway、policy engine 是否支持立即暂停。
  • telemetry 是否能标记 exception id、control gap、model/prompt/source/tool version。
  • evidence 是否能从系统生成, 不依赖截图。
  • approver 是否有权接受该层级 residual risk。
  • aging、repeat renewal、expired exception 是否被监控并上报。

8. Code-Lite Experiment

exception_id: AI-WVR-2026-0042
use_case: retail_service_low_risk_faq_pilot
risk_tier: high
control_gap: CTRL-CITATION-CHECKER-PDF-TABLES
scope: {channel: web_faq, traffic_cap: 5_percent, excluded_intents: [complaint, fee_waiver, credit_commitment, account_closure]}
expiry_date: 2026-07-30
residual_risk_owner: head_of_retail_service
compensating_controls: [daily_sample_review_100_cases, no_write_tools, source_allowlist_only, complaint_intent_hard_route]
hard_stop: [confirmed_wrong_fee_commitment_gt_0, complaint_escalation_miss_gt_0, unsupported_citation_sample_rate_gt_2_percent]
evidence: [waiver_memo, daily_kri_dashboard, qa_sample_log, release_trace_samples]
decision: limited_approval

Validation rule examples:

Reject if expiry_date is missing.
Reject if hard_stop is empty.
Reject if prohibited_use is true.
Reject if risk_tier is high/critical and residual_risk_owner lacks authority.
Escalate if renewal_count >= 2 or days_active > 90.

9. Interview Questions

QuestionStrong answer angle
How is waiver management different from risk appetite?Risk appetite defines baseline boundaries; waiver management controls specific deviations from standard controls after the baseline exists.
When would you approve an AI exception?When the use is not prohibited, scope is narrow, residual risk is accepted by the right owner, compensating controls are testable, expiry is fixed and hard stops are enforceable.
What makes GenAI exception handling different?It combines model risk, operational risk, consumer compliance, privacy, security, third-party, data and agent tool-control concerns.
How do you prevent exceptions becoming shadow policy?Track aging, renewals, repeat reasons, expired waivers, remediation backlog and require policy update or closure after limited renewals.
What should board/audit see?Active high/critical exceptions, residual risk accepted, breached conditions, expired items, repeat waivers, hard stops triggered and remediation progress.

30 秒版本:

Risk appetite 是基线, waiver 是偏离基线后的受控例外。我会要求每个 AI exception 都有 policy/control id、业务理由、scope、expiry、residual risk owner、compensating controls、hard stop、evidence 和 exit path。任何长期续期的例外都要升级, 因为它可能已经变成 shadow policy。


10. Pitfalls

PitfallWhy dangerousFix
无限期 waiver例外变成事实政策fixed expiry and renewal limit
只写业务理由看不到风险被谁接受residual risk memo and approver role
补偿控制不可执行审计时只剩承诺map every control to workflow, log or sample evidence
忽略第三方vendor 变化可改变数据、模型和可用性风险include third-party risk and fallback evidence
只按模型风险审批GenAI/agentic AI 还涉及工具、隐私、安全、运营和消费者合规cross-domain waiver review
到期不处理过期例外仍在生产运行automatic disable and escalation
没有 hard stop指标恶化后继续运行pre-approved pause/rollback conditions
低层级批准高风险residual risk accountability 错位tiered authority matrix

11. Practice Assignment

选择一个金融零售 AI use case, 写一份 exception architecture mini-pack, 包含 risk appetite baseline、control gap、exception taxonomy、residual risk memo、compensating controls、hard stop、expiry/renewal criteria、board/audit dashboard row 和 exit path。

完成标准: 例外没有突破 prohibited use; scope 可以被系统和运营强制执行; residual risk owner 合理; hard stop 可监控、可执行; 到期时可以续期、补齐控制、更新政策或停止; 能用 2 分钟讲清为什么这不是 shadow policy。