AI 扩展计划 / Playbooks
AI Voice AI / Contact Center / Agent Assist Governance Playbook
核心判断:
674 行AI_VOICE_AI_CONTACT_CENTER_AGENT_ASSIST_GOVERNANCE_PLAYBOOK.md
AI Voice AI / Contact Center / Agent Assist Governance Architecture Playbook
面向对象: Advanced AI PM / Senior BA / AI Product Architect / Contact Center Transformation Lead / Enterprise Architect / Conduct Risk / Compliance / Model Risk / Complaint Operations / Fraud-Scam Risk / Accessibility Lead / QA Lead / Workforce Operations / Vendor Risk / Financial Retail Business Owner。 核心问题: 如何把 voice bots、real-time transcription、agent assist、call summarization、next-best-action、speech analytics、QA automation、workforce coaching、disclosures、recording consent、accessibility、fraud/social engineering signals、complaints、model risk 和 operational telemetry 落成一套可上线、可审计、可运营的金融零售 contact-center AI 控制体系? 学习目标: 形成 action-oriented playbook, 能支持 executive framing、taxonomy、decision gates、artifacts、RACI、operating model、implementation roadmap、evidence pack、release checklists、metrics、anti-patterns、tabletop 和 portfolio deliverables。
核心判断:
Voice AI in financial retail should be governed as regulated communication infrastructure, not as a call deflection feature.
0. Disclaimer
本文是学习、架构训练和作品集材料, 不构成法律意见、监管意见、合规结论、TCPA/TSR/recording consent 适用性结论、消费者保护意见、劳动用工建议、隐私影响评估结论、模型验证结论、供应商合规认证或客户通知建议。
AI-generated voice、outbound calling、telemarketing、robocall、prerecorded/artificial voice、call recording、real-time transcription、sentiment analytics、employee monitoring、customer disclosure 和 consent 的具体适用性取决于 call type、channel、jurisdiction、customer relationship、客户角色、员工角色、产品类型、通话目的、数据用途、保留策略、供应商能力、机构政策以及 counsel/compliance interpretation。
正式项目必须由 Legal、Compliance、Privacy、Conduct Risk、Model Risk、Operational Risk、Contact Center Operations、Complaint Operations、Fraud/Scam Risk、Accessibility、Information Security、Data Governance、Vendor Risk、Workforce Management、Product Owner、Internal Audit 和必要的外部顾问共同评审。
1. Executive Framing
高管通常会用这些目标启动 contact-center AI:
- 降低 Average Handle Time。
- 提升 containment 和 self-service。
- 减少 after-call work。
- 自动发现 QA 问题。
- 加速员工培训。
- 从通话中挖掘客户洞察。
这些目标都合理, 但在金融零售中不充分。更高阶的 framing 是:
Every AI-assisted conversation is a customer-impacting control event.
It may create commitments, disclosures, complaints, consent evidence,
fraud interventions, sales conduct risk, accessibility obligations,
employee coaching evidence and regulatory records.
高管要批准的不是“AI 提效项目”, 而是一个 communication governance operating system。
Executive decision questions:
| Question | Why it matters |
|---|---|
| Which calls are eligible for AI voice automation? | 高风险场景需要人工、专员或额外 disclosure/consent gate |
| What must customers be told? | AI-generated voice、recording、transcription、analytics 和 outbound use 可能有不同 disclosure/consent boundary |
| What may AI recommend to employees? | Agent assist 影响客户最终听到的机构立场 |
| What is the system of record? | Audio、transcript、summary、CRM note 和 complaint file 的证据地位不同 |
| How are complaints linked to AI traces? | 没有 linkage 就没有 RCA、remediation 和 model improvement |
| How do we measure safe success? | AHT/containment 必须被 access、conduct、complaint、evidence 和 customer outcome 平衡 |
Executive one-liner:
We will not scale contact-center AI until we can prove what AI heard,
inferred, suggested, changed, disclosed, recorded, escalated and remediated.
2. Source Anchors
| Anchor | Official link | Playbook 使用方式 |
|---|---|---|
| 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 automation 和 disclosure/consent risk 的监管锚点; 不作普遍适用结论 |
| 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、calling practice、recordkeeping 和 customer communication control 的锚点 |
| CFPB Consumer Complaint Database | https://www.consumerfinance.gov/data-research/consumer-complaints/ | 用 complaint themes、consumer harm language 和 complaint linkage 设计 RCA、remediation 和 monitoring |
| NIST AI RMF | https://www.nist.gov/itl/ai-risk-management-framework | 用 Govern / Map / Measure / Manage 组织 AI risk lifecycle、control monitoring 和 improvement |
| ISO/IEC 42001 overview | https://www.iso.org/standard/42001 | 用 AI management system、role accountability、operation planning、performance evaluation、audit 和 improvement 搭建 operating model |
| WCAG 2.2 | https://www.w3.org/TR/WCAG22/ | 作为 accessible customer channels baseline, 扩展到 voice-adjacent UI、transcripts、captions、agent desktop 和 channel handoff |
Source nuance:
- FCC/FTC anchors 是 risk lens, 不是本 playbook 对 TCPA/TSR/telemarketing/robocall/recording consent 的法律结论。
- WCAG 2.2 不能替代电话/语音渠道的全部无障碍判断, 但可作为 digital transcript、chat handoff、agent desktop、document 和 customer portal 的 baseline。
- NIST AI RMF 和 ISO/IEC 42001 用来组织治理系统, 不是单一模型评估表。
3. Scope and Taxonomy
3.1 Capability Taxonomy
| Capability | Included use cases | Primary risks |
|---|---|---|
| Voice bot | inbound self-service, authentication, FAQ, servicing, payment arrangement, fraud triage | disclosure, wrong advice, inaccessible flow, failed handoff |
| AI-generated outbound voice | callback, fraud alert, collections reminder, marketing-like outreach, servicing notice | call-purpose classification, consent/disclosure, impersonation concern |
| Real-time transcription | ASR, diarization, confidence, redaction | transcription error, sensitive data exposure, evidence misuse |
| Agent assist | scripts, knowledge retrieval, response draft, compliance reminder, escalation prompt | unsupported promise, policy hallucination, overreliance |
| Call summarization | after-call notes, complaint summaries, fraud/dispute notes, QA summaries | hallucinated facts, omitted commitments, sensitive-note leakage |
| Next-best-action | offer, route, hardship path, fraud pause, fee review, complaint step | conduct risk, sales pressure, objective misalignment |
| Speech analytics | topic, script adherence, silence, interruption, complaint phrases | over-surveillance, weak signal misuse |
| Sentiment/emotion signal | anger/frustration/stress/empathy scoring | bias, false inference, unfair customer or employee treatment |
| QA automation | script compliance, complaint capture, quality scoring | false QA findings, lack of calibration, employee impact |
| Workforce coaching | training, scorecards, call review, manager prompts | employee notice, appeal, fairness, model drift |
| Fraud/social engineering detection | scam script, coached answers, voice anomaly, remote access hints | false positive, opaque blocking, privacy |
3.2 Call-Purpose Taxonomy
| Call purpose | Examples | Default governance posture |
|---|---|---|
| Servicing | balance, card replacement, address change, fee question | AI eligible with disclosure and source-grounded responses |
| Complaint | unfair treatment, dispute escalation, regulator mention | complaint capture and evidence preservation required |
| Fraud/scam | unauthorized transaction, suspicious transfer, identity concern | specialist route, safe pause policy, final-channel capture |
| Collections / hardship | delinquency, payment plan, job loss | conduct controls, hardship screening, sales suppression |
| Marketing / sales | new product, retention, upgrade | heightened script, consent and suitability/sales practice controls |
| Credit / insurance / investment support | adverse action, claim, suitability, coverage | approved content, human review, no unsupported advice |
| Bereavement / vulnerable situation | death notice, accessibility barrier, language need | dignified handoff, sensitive note control, alternative channel |
| Authentication / security | identity proof, account access, reset | fallback, fraud controls, accessibility and privacy review |
4. Target Operating Model
4.1 Control Plane
conversation intake
-> call-purpose classification
-> disclosure / consent / recording / AI-use gate
-> accessibility and language preference
-> ASR / diarization / redaction
-> real-time agent assist and source retrieval
-> risk signal detection: complaint, conduct, fraud, hardship, accessibility
-> next-best-action with policy constraints
-> human decision and final-channel capture
-> call summary and case note review
-> evidence ledger
-> QA automation and model monitoring
-> complaint / remediation / CAPA
4.2 Operating Principles
| Principle | Practical meaning |
|---|---|
| Call purpose before model | 先判断通话目的, 再决定 AI 能做什么 |
| Disclosure by runtime policy | disclosure/consent 由 policy engine 根据情境决定, 不是写死脚本 |
| Audio-backed evidence | transcript 和 summary 是 derived artifacts, 需要 source links |
| Assist not authority | Agent assist 只辅助员工, 不替代机构判断 |
| Weak signals stay weak | sentiment/emotion 不作为单独决策依据 |
| Customer outcome over containment | containment 不能牺牲投诉、无障碍、欺诈保护和公平处理 |
| Final-channel capture | 保存客户实际听到/看到的内容 |
| Complaint learning loop | 投诉进入 model/product/process/vendor 改进 |
| Proportional retention | 音频、转写、摘要、QA 和 training data 各自有 purpose 和 retention |
5. Decision Gates
Gate 0: Use Case Eligibility
| Question | Pass condition |
|---|---|
| Does AI directly speak to customer or influence what an employee says? | Customer-impacting use case documented |
| Does the call involve sales, collections, fraud, complaint, credit, insurance, investment, hardship or legal-sensitive content? | High-risk workflow review triggered |
| Can customers reach a human or accessible alternative? | Handoff and accessibility path tested |
| Is the vendor/system able to export trace evidence? | Evidence export validated |
Gate 1: Call-Purpose, Disclosure and Consent
| Question | Pass condition |
|---|---|
| How is call purpose classified? | classifier plus agent/customer correction path |
| Which disclosure script applies? | versioned script selected by runtime policy |
| Which processing purposes apply? | recording, transcription, AI assist, QA, coaching, training separated |
| What if customer objects or policy cannot be satisfied? | fallback route and evidence |
| Is outbound AI-generated voice allowed for this scenario? | legal/compliance policy decision recorded |
Gate 2: Data and Retention
| Data asset | Required decision |
|---|---|
| Audio recording | source of evidence, retention class, access roles |
| Transcript | confidence, correction, redaction, versioning |
| AI recommendation | prompt/model/source version, employee action |
| Summary | review state, source links, sensitive-note classification |
| Sentiment/emotion signal | permitted use, prohibited use, retention |
| QA score | calibration, appeal, employee visibility |
| Training data | de-identification, consent/purpose, vendor restrictions |
Gate 3: Model Behavior and Guardrails
| Check | Pass condition |
|---|---|
| Agent assist is grounded in approved sources | source manifest and citation visible to agent |
| Model avoids unsupported promises | eval pass for refund, fee waiver, fraud recovery, credit, hardship |
| Model avoids diagnosis and character judgment | red-team pass |
| Model handles low ASR confidence | asks for clarification or escalates |
| Model detects complaint language without deflection | complaint scenario eval pass |
| Model does not change regulated meaning in simplification | content QA pass |
Gate 4: Human Handoff and Escalation
| Situation | Required behavior |
|---|---|
| Customer asks for human | route without penalty and preserve context |
| Complaint language appears | capture complaint, explain next steps, link evidence |
| Fraud/scam risk appears | safe pause or specialist route per policy |
| Accessibility barrier appears | alternative route, deadline protection, defect capture |
| Hardship/bereavement appears | specialist handoff and sales suppression |
| AI uncertainty is high | agent warning and no high-impact action without review |
Gate 5: Production Monitoring and Change Control
| Question | Pass condition |
|---|---|
| Are control metrics live? | dashboard covers access, disclosure, conduct, complaint, evidence, model |
| Are AI-involved complaints reviewed? | weekly/monthly governance rhythm |
| Can prompt/RAG/model changes be traced? | change record and regression eval |
| Are vendor changes controlled? | notification, test, rollback and audit clauses |
| Are CAPA items funded? | owner, due date and closure evidence |
6. Required Artifacts
| Artifact | What it proves |
|---|---|
| Contact-Center AI Use Case Inventory | AI capabilities are classified by customer impact and risk |
| Call-Purpose Taxonomy | disclosure, consent, scripts, routing and retention depend on call context |
| Disclosure and Consent Decision Matrix | legal/compliance policy is operationalized into runtime gates |
| Data Purpose and Retention Matrix | audio/transcript/summary/sentiment/QA/training data have purpose boundaries |
| Agent-Assist Guardrail Pack | employees receive grounded, limited, reviewable recommendations |
| Approved Content and Script Registry | AI outputs are tied to approved policy, fee, disclosure and complaint language |
| ASR and Speech Analytics Eval Pack | transcription and speech signals are validated across language/noise/channel |
| Sentiment/Emotion Use Policy | weak signals are not misused for high-impact decisions |
| NBA Objective and Constraint Spec | recommendations optimize customer outcome under conduct constraints |
| Call Summary Schema | notes separate facts, commitments, unresolved items and AI uncertainty |
| Evidence Ledger Schema | every AI-involved conversation is replayable |
| QA Automation Calibration Plan | automated QA is reviewed, explainable and appealable |
| Complaint Linkage Schema | complaints connect to conversation, AI run, final content, remediation and RCA |
| Operational Telemetry Dashboard | production controls are monitored beyond efficiency |
| Vendor Control Addendum | trace export, retention, training use, accessibility, incident and change controls |
6.1 Product / Architecture Decision Matrix
| Decision | Senior PM / Architect question | Decision record evidence |
|---|---|---|
| AI voice eligibility | Which call types can be automated without harming access, consent or customer outcome? | use-case risk tiering |
| Human fallback | When must customer reach human or specialist? | handoff policy and SLA |
| ASR confidence handling | How do low-confidence segments affect AI suggestions and summaries? | ASR control spec |
| Summary authority | Can summary update CRM automatically? | summary review and source-link policy |
| NBA objective | What is optimized and what is prohibited? | objective/constraint memo |
| Sentiment use | What decisions cannot use emotion score? | prohibited-use policy |
| Complaint capture | What language triggers complaint route? | complaint classifier and agent confirmation |
| Recording/AI analytics | What is disclosed, recorded, analyzed and retained? | decision matrix and evidence ledger |
7. RACI and Operating Model
| Capability / Control | Accountable | Responsible | Consulted | Informed |
|---|---|---|---|---|
| Use case risk tiering | AI Governance / Business Owner | Product Owner | Legal, Compliance, Model Risk, Operations | Senior Management |
| Call-purpose taxonomy | Conduct Risk | Product / Contact Center Ops | Legal, Complaint Ops, Fraud | AI Governance |
| Disclosure and consent policy | Legal / Compliance | Product / Platform | Privacy, Operations, Vendor Risk | QA, Audit |
| Voice accessibility | Accessibility Lead | Design / Engineering / QA | Legal, CX, Operations | Business Owner |
| ASR and speech analytics validation | Model Risk | Data Science / AI Platform | QA, Operations, Fairness/Privacy | Governance Forum |
| Agent-assist guardrails | Product Owner | AI Platform / Ops Enablement | Compliance, Legal, Frontline QA | Contact Center Agents |
| NBA controls | Business Owner / Conduct Risk | Product / Decisioning Team | Model Risk, Complaints, Fraud | Risk Committee |
| Call summary controls | Operations / Complaint Ops | Case Management / AI Platform | Legal, Privacy, QA | Audit |
| Complaint linkage | Complaint Operations | Case Platform / Analytics | Compliance, Product, Model Risk | Senior Management |
| QA automation | QA Lead | QA Analytics / Ops Managers | Workforce, Legal, Employee Relations | Agents |
| Vendor controls | Vendor Risk | Procurement / Technology | Legal, Security, Privacy, Model Risk | Business Owner |
| CAPA closure | Business Owner | Product / Ops / Platform | Compliance, Audit | AI Governance |
Governance cadence:
| Cadence | Forum | Output |
|---|---|---|
| Daily during pilot | Production control standup | incidents, latency, control breaches, customer-impact triage |
| Weekly | AI-involved call QA review | call samples, complaint misses, summary defects, agent-assist errors |
| Monthly | Contact-center AI risk committee | metrics, KRIs, complaint themes, fraud outcomes, accessibility defects |
| Quarterly | Model and conduct risk review | threshold tuning, eval results, fairness, CAPA aging, vendor changes |
| Semiannual | Tabletop exercise | scam, complaint, consent defect, outage, bad summary, accessibility failure |
| Annual | AI management system review | policy effectiveness, audit results, roadmap and funding decisions |
8. Implementation Roadmap
Days 1-30: Foundation and Boundary Setting
| Day range | Work | Deliverable |
|---|---|---|
| 1-5 | Inventory current voice/contact-center AI capabilities and vendors | Use Case Inventory |
| 6-10 | Define call-purpose taxonomy and high-risk call types | Call-Purpose Taxonomy |
| 11-15 | Build disclosure/consent/recording/AI-use decision matrix with Legal/Compliance | Runtime Policy Matrix |
| 16-20 | Map audio, transcript, summary, QA, sentiment and training data flows | Data Purpose and Retention Matrix |
| 21-25 | Define evidence ledger schema and final-channel capture requirement | Evidence Schema |
| 26-30 | Select pilot scope with clear exclusions | Pilot Boundary Decision |
Days 31-60: Controlled Build
| Day range | Work | Deliverable |
|---|---|---|
| 31-35 | Implement ASR confidence, redaction, diarization and correction workflow | ASR Control Spec |
| 36-40 | Build agent-assist guardrail pack and approved content registry | Guardrail Pack |
| 41-45 | Configure complaint, fraud/scam, hardship and accessibility signal routes | Escalation Workflow |
| 46-50 | Design call summary schema and review workflow | Summary Control |
| 51-55 | Build QA/eval scenario suite | Eval Pack |
| 56-60 | Prepare dashboards for access, disclosure, conduct, complaint, evidence and model risk | Telemetry Dashboard |
Days 61-90: Pilot and Assurance
| Day range | Work | Deliverable |
|---|---|---|
| 61-65 | Run pilot with limited queues and manual review | Pilot Launch Record |
| 66-70 | Sample calls for ASR, agent-assist, summary, complaint and disclosure defects | QA Calibration Report |
| 71-75 | Test accessibility paths and human fallback | Accessibility Evidence |
| 76-80 | Review AI-involved complaints and remediation | Complaint Learning Report |
| 81-85 | Tune thresholds and update guardrails through change control | Threshold Review Memo |
| 86-90 | Decide scale, restrict, redesign or retire | Go/No-Go Decision Record |
Days 91-180: Scale with Control Maturity
| Workstream | Scale activity | Evidence |
|---|---|---|
| Governance | Extend risk committee and CAPA funding | committee minutes, CAPA closure |
| Model risk | Add regression eval for every model/prompt/RAG/script change | validation pack |
| Operations | Train supervisors and agents on AI assist limits | training records |
| Complaints | Integrate AI trace into complaint case management | complaint schema |
| Vendor | Contract for trace export, retention controls and incident SLAs | vendor addendum |
| Accessibility | Add periodic assistive technology testing | test report |
| Internal audit | Review evidence completeness and control effectiveness | audit report |
9. Evidence Pack
Minimum evidence fields:
| Field | Purpose |
|---|---|
| conversation_id | common call reference |
| customer_id_hash / account_ref | controlled linkage without unnecessary exposure |
| call_purpose | servicing, complaint, fraud, collections, sales, dispute |
| channel | inbound, outbound, callback, voice bot, human agent, relay |
| disclosure_script_id | versioned customer disclosure |
| consent_or_policy_state | customer response or policy basis |
| recording_flag | audio capture status and purpose |
| transcription_flag | real-time/post-call transcription purpose |
| ai_assist_flag | whether AI influenced employee |
| analytics_flags | sentiment, QA, coaching, fraud, training use |
| audio_pointer | source evidence location and retention class |
| transcript_id | version, confidence, correction state |
| model_route | model/provider/version/endpoint |
| prompt_bundle_id | system/developer/policy prompts |
| source_manifest | approved content/RAG source versions |
| recommendation_id | AI recommendation and guardrail warnings |
| agent_decision | accepted, modified, rejected, escalated |
| agent_reason_code | why action was taken |
| final_channel_event_id | what customer heard/saw |
| summary_id | source-linked summary version |
| risk_signal_ids | complaint, fraud, hardship, accessibility, conduct |
| complaint_id | linked complaint if any |
| remediation_id | customer remediation |
| QA_result | automated and human review |
| CAPA_id | improvement action |
Evidence rules:
- Store raw audio, transcript, summary and QA outputs as separate artifact classes。
- Keep generated summaries source-linked and reviewable。
- Capture final customer communication, not only AI draft。
- Preserve policy/script/source versions used at the time of call。
- Record customer objections or withdrawal according to policy。
- Treat evidence gaps as control defects, not merely logging bugs。
10. Checklists
10.1 Pre-Launch Checklist
| Check | Passing evidence |
|---|---|
| Use case and call-purpose risk tier assigned | risk assessment |
| Disclosure/consent decision matrix approved | Legal/Compliance signoff record |
| Recording/transcription/analytics purpose separated | data purpose matrix |
| Human fallback and specialist routing tested | handoff test evidence |
| Voice and digital accessibility paths tested | accessibility report |
| ASR eval includes language, accent, noise and channel conditions | ASR eval pack |
| Agent-assist guardrails pass scenario eval | eval report |
| Approved content registry versioned | source manifest |
| Summary schema and review workflow implemented | summary QA samples |
| Complaint linkage tested end to end | complaint test case |
| Operational telemetry dashboard live | dashboard screenshot/export |
| Incident and rollback process ready | runbook |
10.2 Runtime Call Checklist
| Check | Evidence |
|---|---|
| Call purpose identified | call-purpose event |
| Required disclosure presented | script ID and timestamp |
| Consent/policy state recorded where needed | consent event |
| Accessibility or language need honored | preference event |
| Low-confidence transcript flagged | ASR confidence |
| AI suggestion grounded in approved source | recommendation source |
| Agent decision recorded | reason code |
| Complaint/fraud/hardship signal routed | escalation event |
| Final customer message captured | final-channel event |
| Summary reviewed if stored to CRM/case | summary review state |
10.3 Agent-Assist Review Checklist
| Check | Passing evidence |
|---|---|
| No unsupported refund, fee waiver, fraud recovery or approval promise | transcript/QA |
| No legal conclusion or regulatory conclusion | QA sample |
| No diagnosis of customer capacity, mental state or honesty | QA sample |
| Required disclosure or complaint route shown | agent desktop evidence |
| Source link visible to agent | source manifest |
| Uncertainty shown when ASR/RAG confidence low | UI evidence |
| Employee did not rely on sentiment alone | reason code |
| Human escalation used for high-risk scenario | escalation event |
10.4 Call Summary Checklist
| Check | Passing evidence |
|---|---|
| Customer-stated issue preserved | summary field |
| Agent commitments separated | summary field |
| Amounts, dates and deadlines accurate | source comparison |
| Complaint/dispute language not omitted | complaint classifier |
| AI uncertainty documented | uncertainty field |
| Sensitive details excluded from general notes | note classification |
| Source transcript/audio linked | evidence links |
| Reviewer captured for high-risk calls | reviewer field |
10.5 Vendor Checklist
| Check | Passing evidence |
|---|---|
| Trace export includes prompts, model versions, outputs and timestamps | export test |
| Data retention and deletion are configurable | admin evidence |
| Customer data not used for vendor training unless approved | contract/DPA setting |
| Accessibility support is documented | VPAT/SLA/test evidence |
| Model or feature changes require notification | change clause |
| Incident support and rollback SLA defined | runbook |
| Subprocessor and data location transparent | vendor risk file |
| QA and workforce analytics explainable | score explanation |
11. Metrics and KRIs
| Metric | Why it matters |
|---|---|
| AI-involved call volume by purpose | confirms exposure and risk mix |
| Missing disclosure defect rate | detects runtime policy failures |
| Consent/policy state mismatch | detects evidence and routing gaps |
| Voice bot containment with safe-exit rate | balances automation and access |
| Human fallback success rate | protects customer access |
| Accessibility-related call abandonment | detects exclusion |
| ASR low-confidence rate by segment | identifies language/noise/channel risk |
| Agent-assist grounded answer rate | measures source discipline |
| Prohibited-output rate | measures guardrail health |
| Agent override/modify rate | reveals AI usefulness and risk |
| Complaint capture defect rate | conduct risk KRI |
| AI-involved complaint uphold rate | customer harm signal |
| Safe-pause false positive/negative | fraud and autonomy balance |
| Summary factuality defect rate | evidence quality |
| Omitted commitment rate | dispute and remediation risk |
| Sentiment-use exception count | prohibited-use monitoring |
| QA appeal overturn rate | workforce fairness and model quality |
| Final-channel capture rate | evidence completeness |
| CAPA aging | governance execution |
| Model/prompt/RAG change regression pass rate | change control health |
Balanced scorecard:
Efficiency: handle time, after-call work, containment.
Access: human fallback, accessibility, language support.
Conduct: disclosure, complaint capture, sales/collections controls.
Protection: fraud/scam, hardship, vulnerable situation escalation.
Evidence: final-channel capture, summary quality, AI trace completeness.
Workforce: QA calibration, coaching fairness, appeal outcomes.
Learning: complaint RCA, incident remediation, CAPA closure.
12. Anti-Patterns
| Anti-pattern | Why it fails | Better pattern |
|---|---|---|
| “AI may be used” as one blanket disclosure | Ignores call purpose, channel, jurisdiction, processing purpose | runtime disclosure/consent policy gate |
| Voice bot as universal front door | Blocks high-risk, accessibility or complaint paths | risk-based eligibility and human fallback |
| Transcript as system of record | ASR errors become institutional facts | audio-backed transcript with confidence and correction |
| Unreviewed summary in CRM | Hallucinations affect future customer treatment | reviewed, source-linked, classified summary |
| Sentiment score used for action | Emotion inference is weak and biased | QA attention only unless reviewed and approved |
| Agent assist without sources | Policy hallucination and unsupported promises | source-grounded recommendations |
| NBA optimizes AHT only | Creates conduct and complaint risk | customer-outcome constrained optimization |
| Complaint route hidden in bot | Deflects regulated customer concerns | complaint capture and clear escalation |
| QA automation as discipline engine | False positives harm employees and trust | calibrated QA, evidence, appeal |
| Vendor black box | No audit, no trace, no RCA | trace export and contractual controls |
| Training data purpose creep | Service calls become model data without governance | purpose matrix, de-identification and approval |
13. Tabletop Scenarios
Scenario 1: AI-Generated Fraud Alert Call
The bank wants to use an AI-generated voice to call customers after high-risk card activity.
The call may ask customers to confirm activity and route them to fraud support.
Decision points: call purpose, AI-generated voice eligibility, disclosure, authentication, customer callback safety, evidence, spoofing risk, human fallback, vendor controls。
Scenario 2: Missing Complaint Capture
A customer says three times that the bank treated them unfairly and that they will contact a regulator.
Agent assist continues to provide fee explanations but never prompts complaint capture.
Decision points: complaint trigger, agent training, AI eval defect, remediation, complaint linkage, QA sample expansion。
Scenario 3: Bad Call Summary
A collections call summary says the customer promised to pay by Friday.
Audio review shows the customer said they could not pay and asked about hardship options.
Decision points: source of truth, summary review, customer remediation, employee coaching, model incident, CRM correction。
Scenario 4: Sentiment Misuse
Speech analytics flags a customer as angry and high-risk due to loud background noise and accent.
The queue system routes future calls to a retention script and flags the employee for poor empathy.
Decision points: prohibited use, bias testing, employee appeal, customer routing, telemetry correction。
Scenario 5: Accessibility Failure in Voice Bot
A customer using relay service cannot pass the voice bot authentication and cannot reach a human before a fraud dispute deadline.
Decision points: accessible fallback, deadline protection, defect severity, complaint/remediation, release gate update。
Scenario 6: Hardship Sales Pressure
During a delinquency call, next-best-action recommends a refinance product after the customer says they lost their job.
Decision points: sales suppression, hardship route, conduct QA, NBA objective correction, customer outcome monitoring。
14. Practical Templates
14.1 Call-Purpose Decision Record
Use case:
Call purpose:
Inbound/outbound:
Customer relationship:
Customer role:
Automation type:
AI-generated voice involved:
Recording/transcription involved:
Analytics/coaching/training use:
Disclosure script:
Consent/policy state:
Human fallback:
Jurisdiction policy reference:
Residual risk:
Approver:
14.2 Disclosure and Consent Matrix
| Scenario | Disclosure | Consent/policy handling | Fallback |
|---|---|---|---|
| Inbound servicing with human agent and recording | recording disclosure per policy | record state | continue or non-recorded route if supported |
| Human call with real-time agent assist | AI-use disclosure if policy requires | record policy state | agent-only route |
| Voice bot servicing | automated system/AI voice disclosure per policy | record response if required | human agent |
| Outbound AI-generated voice | scenario-specific legal/compliance decision | block if policy not satisfied | human outbound or secure message |
| Post-call QA analytics | purpose and employee/customer notice per policy | purpose-bound retention | exclude from analytics if required |
14.3 Agent-Assist Prompt Contract
You are assisting a trained employee.
Use only approved sources.
Separate customer-stated facts from inference.
Show uncertainty and low-confidence transcript segments.
Do not make legal conclusions.
Do not promise refunds, fee waivers, fraud recovery, approvals or outcomes.
Do not diagnose customer capacity, emotion, disability or intent.
Flag complaint, fraud, hardship, accessibility and vulnerable-situation signals.
Recommend escalation when policy requires.
The employee must decide what to say.
14.4 Evidence Ledger Event
event_type:
conversation_id:
timestamp:
actor: customer / agent / voice_bot / ai_assist / supervisor
source_artifact:
model_or_script_version:
content:
confidence:
policy_basis:
agent_decision:
final_channel_flag:
linked_case:
retention_class:
14.5 Complaint RCA Template
Complaint ID:
Conversation ID:
AI capability involved:
Customer-stated issue:
Alleged harm:
Final customer communication:
AI recommendation:
Agent decision:
Source transcript/audio:
Root cause category:
Customer remediation:
Control change:
Model/prompt/script/vendor change:
CAPA owner:
Closure evidence:
14.6 Go / No-Go Review Memo
Pilot scope:
Excluded call types:
Disclosure/consent readiness:
Accessibility readiness:
Evidence completeness:
Agent-assist eval result:
ASR eval result:
Complaint linkage test:
Operational telemetry:
Top residual risks:
Decision: scale / restrict / redesign / retire
Conditions:
Approvers:
15. Portfolio Deliverables
| Deliverable | What it demonstrates |
|---|---|
| Contact-Center AI Reference Architecture | 你能把 voice AI、agent assist、evidence、complaints 和 controls 连接起来 |
| Call-Purpose Taxonomy | 你能把不同通话目的转成不同治理路径 |
| Runtime Disclosure/Consent Gate | 你不会把法律/合规问题硬编码成一句脚本 |
| Agent-Assist Guardrail Pack | 你能控制员工 AI 辅助的 customer-impact risk |
| Call Summary and Evidence Schema | 你理解 transcript/summary/CRM/complaint 的证据差异 |
| Sentiment/Emotion Signal Policy | 你能限制弱信号和高风险误用 |
| NBA Conduct Constraint Spec | 你能把 customer outcome 和 conduct risk 放进优化目标 |
| QA Automation Operating Model | 你能平衡自动化质检、员工公平和控制有效性 |
| Complaint Learning Loop | 你能把 complaints 转化为 model/product/process improvement |
| Executive Dashboard | 你能用 access、conduct、evidence、protection 和 efficiency 平衡汇报 |
Portfolio storyline:
I designed a contact-center AI governance architecture for financial retail.
The system classifies call purpose, manages disclosure and consent as runtime policy,
controls transcription and agent assist, captures final customer communications,
links complaints to AI evidence, and measures whether automation improves service
without weakening fair treatment, accessibility or operational controls.
16. Interview Answers
Q1: 如何设计金融零售 voice AI 的治理架构?
30 秒:
我会把它设计成 regulated communication control plane。先按 call purpose 分类, 再由 runtime gate 决定 disclosure、consent、recording、AI assist 和 routing。通话中保留 ASR confidence、AI recommendation、agent decision 和 final-channel capture。投诉、QA、模型监控和 CAPA 必须接入同一条 evidence chain。
Q2: Agent assist 和普通知识库有什么本质不同?
30 秒:
Agent assist 会实时影响员工对客户说什么, 所以它是 customer-impacting decision support。必须 source-grounded、显示 uncertainty、禁止 unsupported promise、禁止 diagnosis、记录员工采纳或修改原因, 并保留客户最终听到的内容。
Q3: Call summary 可以自动写进 CRM 吗?
30 秒:
高风险场景不应无审查写入。Summary 是 derived artifact, 不是事实本身。它要链接 transcript/audio, 区分客户陈述、员工承诺、系统动作、未解决事项和 AI 不确定性, 并对敏感信息做 note classification。
Q4: 怎么看 sentiment/emotion analytics?
30 秒:
我会把它当 weak signal, 可用于 QA attention 和趋势分析, 但不能单独决定客户处理、投诉裁决、欺诈结论、销售路径或员工惩戒。需要 bias testing、human review、prohibited-use policy 和申诉机制。
Q5: 如何平衡 AHT 降低和 conduct risk?
30 秒:
指标不能只看效率。我会同时看 disclosure defects、complaint capture、AI-involved complaint uphold rate、summary factuality、sales suppression、accessibility fallback、safe-pause quality、final-channel capture 和 CAPA aging。效率必须在控制有效的前提下优化。
17. Final Operating Principle
这套 playbook 的成熟度可以用一个问题检验:
When AI participates in a customer call,
can the institution prove the call was properly classified,
disclosed, recorded, transcribed, assisted, summarized, escalated,
quality-checked, complained about, remediated and improved?
如果答案不清楚, 不是缺一个更好的 voice bot 或 copilot。问题是 contact-center AI 还没有成为可治理的 customer communication operating system。