AI Voice / Contact Center：坐席辅助治理架构
Contact-center AI 不是单纯的效率层。它是客户沟通控制平面：治理一通电话或聊天中什么被听见、转写、推断、建议、说出、记录、披露、升级、质检、投诉、补救和改进，并防止 AI inference 被误当成客户事实或机构承诺。
AI Voice AI / Contact Center / Agent Assist Governance Architecture 解读
配对阅读：本篇的操作手册版（模板/RACI/门禁/runbook）是
docs/AI_VOICE_AI_CONTACT_CENTER_AGENT_ASSIST_GOVERNANCE_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。
重要说明: 本文只讨论 voice bots、real-time transcription、agent assist、call summarization、next-best-action、speech analytics、QA automation、workforce coaching、disclosure、recording consent、complaints、fraud signals 和 operational telemetry 的产品与架构治理，不构成法律、监管、TCPA/TSR/recording consent 适用性、消费者保护、劳动用工、医疗/心理判断、客户通知或供应商认证结论。真实项目必须由 Legal、Compliance、Privacy、Conduct Risk、Model Risk、Contact Center Operations、Complaint Operations、Fraud/Scam Risk、Accessibility、Information Security、Data Governance、Vendor Risk、QA、Workforce Management、Internal Audit 等共同确认。
Source Anchors
| Source | Link | 用途 |
|---|---|---|
| FCC AI-generated voices robocalls declaratory ruling page | https://www.fcc.gov/document/fcc-makes-ai-generated-voices-robocalls-illegal | 作为 AI-generated voice、robocall、outbound voice automation 和 disclosure/consent risk 的监管锚点; 具体适用性需由法律/合规判断 |
| FTC Telemarketing Sales Rule, 16 CFR Part 310 | https://www.ecfr.gov/current/title-16/chapter-I/subchapter-C/part-310 | 作为 telemarketing、sales script、misrepresentation、call practices、recordkeeping 和 customer communication conduct control 的锚点; 不推导普遍适用结论 |
| CFPB Consumer Complaint Database | https://www.consumerfinance.gov/data-research/consumer-complaints/ | 用 complaints 作为 AI voice/contact center harm detection、RCA、remediation 和 control improvement 的反馈源 |
| NIST AI RMF | https://www.nist.gov/itl/ai-risk-management-framework | 用 Govern / Map / Measure / Manage 组织 voice AI risk taxonomy、control effectiveness、monitoring 和 continuous improvement |
| ISO/IEC 42001 overview | https://www.iso.org/standard/42001 | 用 AI management system、roles、operation、performance evaluation、audit 和 improvement 建立 contact-center AI operating model |
| WCAG 2.2 | https://www.w3.org/TR/WCAG22/ | 作为 digital/customer channels 的 accessibility baseline, 并扩展到 voice-adjacent UI、captions、transcripts、agent desktop、chat/voice handoff 和 customer summaries |
核心导读
Contact-center AI 不是单纯的效率层。它是客户沟通控制平面：治理一通电话或聊天中什么被听见、转写、推断、建议、说出、记录、披露、升级、质检、投诉、补救和改进，并防止 AI inference 被误当成客户事实或机构承诺。
AI 改变的是实时服务链和事后证据链。Voice bot 可能直接解释政策和执行动作，ASR 决定 agent-assist 看见什么事实，NBA 改变销售/服务路径，summarizer 把通话写入 CRM，speech analytics 触发投诉或 conduct signal，QA automation 影响员工评价和控制结论。任何一个节点把低置信转写、情绪推断或未审批话术当成事实，都会影响客户权利、员工责任和机构证据。
学习这个主题要把“客户实际听到/看到的内容”放在中心。治理边界包括 runtime disclosure/consent gate、voice accessibility、ASR confidence handling、source-grounded agent assist、sales suppression、complaint capture、summary review、QA calibration 和 vendor trace export。AI 可以辅助员工更快定位政策、提醒风险和形成摘要，但不能单独作出欺诈、投诉、催收、销售适当性或员工处分结论；最终沟通、人工判断和证据留存必须可重放。
问题定义
金融零售 contact center 是高后果沟通入口。客户在欺诈、诈骗、盗刷、账户冻结、贷款逾期、催收、保险理赔、信用拒绝、投诉、丧亲和困难时拨打电话；通话内容可能形成授权、拒绝、投诉、争议、承诺、同意、撤回或合规证据。
AI 同时进入多个节点：voice bot 直接回应客户，ASR 生成事实基础，agent-assist 实时建议话术，NBA 改变销售/服务路径，summarizer 写入 CRM，speech analytics 识别投诉/情绪，QA automation 影响员工和控制结论。系统性风险不是“语音识别准不准”，而是：
For every customer conversation,
what did the system capture,
what did AI infer,
what did AI recommend,
what did the agent actually say,
what did the customer understand,
what risk was escalated,
what evidence was preserved,
and what control proves fair treatment?
核心原理/方法
第一条原则：conversation 有 service chain 和 evidence chain 两条链。处理完客户问题不等于能证明处理公平、准确、可复核。
service chain:
customer issue -> conversation -> agent/bot response -> resolution
evidence chain:
call purpose -> disclosure/consent -> transcript -> AI recommendation
-> human decision -> final customer communication -> QA -> complaint/remediation
第二条原则：transcript 是 derived artifact，不是系统事实。ASR 会受口音、噪音、code-switching、diarization、延迟和敏感数据影响；低置信片段不能支撑高后果建议或摘要断言。
第三条原则：agent assist 必须保持 assistive role。员工最终说出的话是机构立场，不能把 AI 建议当作自动授权。
第四条原则：sentiment/emotion 是 weak signal。不得作为客户处理、欺诈结论、投诉裁决或员工纪律处分的单独依据。
第五条原则：disclosure/consent/recording/AI analytics 必须 runtime policy gate，而不是静态开场白。
系统/架构模型
参考架构：
customer voice / chat / callback
-> channel and call-purpose classifier
-> disclosure / consent / recording / AI-use policy gate
-> identity, authentication and accessibility preference layer
-> audio capture and streaming pipeline
-> ASR with confidence, diarization and redaction
-> real-time event bus
-> agent-assist guardrail service
-> approved content and policy retrieval
-> next-best-action / risk signal engine
-> fraud-social-engineering detector
-> complaint and conduct classifier
-> human decision and agent desktop
-> final-channel capture
-> call summary and case-note controls
-> evidence ledger and retention controls
-> QA automation, model monitoring and operational telemetry
-> complaints, remediation, CAPA and governance review
核心组件：
| Component | 职责 |
|---|---|
| Call-purpose classifier | 区分 servicing、collections、marketing、complaint、fraud、dispute、outbound callback |
| Consent/disclosure gate | 根据 call type、jurisdiction、relationship、automation、data use 决定 disclosure、consent、block/downgrade |
| Voice accessibility layer | relay、caption、transcript、repeat、slower speech、DTMF fallback、human handoff、language routing |
| ASR/diarization service | 转写、说话人分离、置信度、时间戳、redaction |
| Agent-assist guardrail | 显示 source、uncertainty、required disclosure、actions not allowed、escalation option |
| Approved content/RAG | 版本化管理 policy、fees、deadlines、scripts、disclosures、complaint language |
| Risk signal engine | complaint、fraud/scam、hardship、accessibility、conduct signals |
| NBA orchestration | 推荐 next step，但受 conduct、sales suppression、customer outcome 约束 |
| Final-channel capture | 保存客户实际听到/看到的内容，而不只是 AI draft |
| Evidence ledger | 连接 audio、transcript、AI runs、agent actions、summary、QA、complaint、remediation |
关键机制与取舍
Capability taxonomy：
| Capability | Customer impact | 核心控制 |
|---|---|---|
| Voice bot / IVR AI | 直接与客户交互、解释选项、执行动作 | automation disclosure、high-risk exit、human fallback |
| Real-time transcription | 成为 agent assist、QA、summary 的输入 | confidence、correction、low-confidence handling |
| Agent assist | 实时提示话术和下一步 | source grounding、prohibited actions、human accountability |
| Call summarization | 写入 CRM/case/complaint notes | facts vs inference、commitments、source links、review |
| Next-best-action | 推荐 offer、fee waiver、hardship、fraud step | multi-objective and conduct-safe objective |
| Speech analytics | 识别主题、投诉、脚本、情绪 | weak signal policy、bias review |
| QA automation | 评估脚本、投诉捕获、销售/催收行为 | calibration、human review、appeal |
| Workforce coaching | 员工培训和绩效趋势 | employee notice、evidence excerpts、challenge process |
| Fraud/scam detection | 识别 coached responses、safe-account、remote access | safe pause、specialist、customer explanation |
Agent-assist 输出应包括：
Customer-stated facts
Observable workflow facts
Relevant policy/source
Recommended response
Required disclosure or verification
Uncertainty / low-confidence transcript segments
Actions not allowed
Escalation option
Agent must confirm before saying
NBA 取舍：
| Context | 弱目标 | 强目标 |
|---|---|---|
| Collections | maximize promise-to-pay | sustainable repayment、hardship screening、conduct-safe script |
| Fraud alert | reduce fraud loss | prevent loss while preserving autonomy and review |
| Complaint call | close quickly | capture complaint、explain next steps、preserve evidence |
| Fee dispute | reduce refund | apply policy consistently、escalate edge cases |
| Cross-sell | increase conversion | suppress sales in hardship、complaint、fraud、bereavement、accessibility barrier |
证据与控制
Conversation evidence ledger：
conversation_id / call_id
call purpose and channel
customer disclosure / consent status
recording, transcription, AI analytics and training-use flags
audio pointer and retention class
transcript version with confidence
speaker diarization and timestamps
prompt_bundle_id / model_version / source manifest
AI recommendation and prohibited-action warnings
agent action and reason code
final-channel content
summary version and note classification
complaint_id / remediation_id
QA result and CAPA link
控制矩阵：
| Control objective | Control activity | Evidence |
|---|---|---|
| 分类 call purpose | runtime classifier + agent confirmation | call-purpose event、routing log |
| 管理 disclosure/consent | policy gate selects flow and blocks unsupported automation | disclosure version、timestamp、response、policy decision |
| 保持可访问服务 | relay、caption/transcript、repeat、slower speech、DTMF/human fallback | accessibility test、call samples |
| 控制 ASR error | confidence、correction workflow、language/accent eval | ASR metrics、correction log |
| Ground agent assist | approved retrieval、policy version、prohibited output guardrails | source manifest、prompt bundle、eval report |
| 防止 conduct harm | sales suppression、no unsupported promises、complaint capture | QA results、script violations |
| 治理 sentiment/emotion | weak-signal policy、prohibited use、bias testing、human review | model card、usage policy |
| 捕获最终沟通 | 保存客户实际听到/看到的内容 | audio timestamp、transcript segment、final message id |
| 控制 summaries | source-linked、reviewed、sensitive-note rules | summary version、reviewer、source links |
| 链接 complaints | complaint schema includes AI run、final content、agent action、RCA | complaint record、CAPA |
指标应同时看 AHT/containment、disclosure defects、ASR WER by language/noise、hallucinated policy rate、prohibited promise rate、agent override、complaint AI-linkage、safe-pause quality、summary factuality、accessibility completion、QA appeal、CAPA aging。
金融零售/AI产品场景
- Fraud alert inbound call：call purpose 识别为 fraud，系统播放相应 disclosure，agent assist 禁止承诺 recovery，提示 authentication/session evidence 和 safe-pause script。
- Collections call：客户说失业，NBA 从 promise-to-pay 转为 hardship path，sales suppression 生效，AI 不能生成羞辱或威胁话术。
- Fee dispute：ASR 对金额片段低置信，agent desktop 要求澄清；summary 不得断言客户同意某金额。
- Complaint capture：客户说“我要投诉/报告监管”，classifier 建议 complaint intake，员工确认后生成 complaint_id 并链接 conversation。
- Voice bot accessibility：客户使用 relay 或需要 slower speech，bot 提供 human fallback 和 transcript，不因语音障碍进入 authentication loop。
- Workforce QA：自动 QA 标记未给 disclosure，但分数需要 human calibration；ASR 错误不能直接成为员工纪律结论。
反模式
| 反模式 | 风险 | 更好的控制 |
|---|---|---|
| 所有电话一个 generic disclosure | 可能错过 call purpose、jurisdiction、AI-use、recording 差异 | runtime policy gate |
| transcript 当 truth | ASR/diarization 错误进入 case facts | audio-backed, confidence-scored, corrected transcript |
| 未复核 summary 进 CRM | hallucinated or biased notes 影响未来服务 | reviewed, source-linked summary schema |
| sentiment score 驱动动作 | emotion inference 不稳定且可能偏差 | weak-signal-only policy |
| agent assist 给最终答案 | 员工依赖 unsupported AI authority | source-grounded draft + human accountability |
| NBA 在 hardship 中优化销售 | conduct harm | sales suppression and customer-outcome objective |
| voice bot 隐藏人工路径 | 客户无法获得服务或投诉 | clear human handoff |
| complaint language 未捕获 | 补救和监管证据缺口 | classifier + agent confirmation |
| QA automation 无校准 | 错误评分影响员工和控制结论 | human calibration and appeal |
| vendor black box | 无 trace、version、evidence | contractual trace export and audit controls |
最终心智模型
Contact-center AI 的本质是 runtime communication control plane。成熟系统应能重放一通客户对话如何被分类、披露、录音、转写、辅助、总结、升级、质检、投诉、补救和改进，并始终区分客户事实、AI 推断、员工承诺和最终客户可见内容。
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」。