AI Collections / Hardship / Delinquency Treatment Architecture Playbook

面向对象: Advanced AI PM / Senior BA / AI Product Architect / Enterprise Architect / Credit Risk / Collections Strategy / Hardship Operations / Conduct Risk / Compliance / Complaint Operations / Model Risk / Frontline Operations / Vendor Management / Customer Experience。 核心问题: 如何把 AI early delinquency detection、contact strategy、hardship options、repayment treatment、vulnerable-customer signals、agent assist、complaint linkage、model risk 和 operational controls 设计成可落地、可审计、可度量、可持续改进的金融零售 operating system? 学习目标: 形成一套 action-oriented playbook, 支持高管 framing、taxonomy、decision gates、required artifacts、RACI、implementation roadmap、evidence pack、QA/eval、metrics、checklists、anti-patterns、tabletop 和 portfolio deliverables。 定位: 本文不是基础催收流程介绍, 而是训练你把 collections、hardship、customer treatment、conduct risk、AI governance 和 financial retail architecture 合成一套 senior PM/architect 决策框架。

核心判断:

Collections AI should not make institutions better at pressuring customers. It should make them better at detecting difficulty early, offering sustainable treatment, controlling conduct risk, and proving fair outcomes.

---

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、FDCPA/Reg F 适用性判断、消费者保护意见、信贷建议、债务咨询、催收话术审批、客户通知建议、模型验证意见或监管报告建议。

正式项目必须由 Legal、Compliance、Conduct Risk、Credit Risk、Collections Operations、Hardship Operations、Complaint Operations、Privacy、Model Risk、Operational Risk、Information Security、Data Governance、Customer Experience、Frontline Operations、Vendor Management、Internal Audit 和必要的外部顾问共同判断。

特别注意: FDCPA、Regulation F、州法、产品规则、servicing obligations、fair lending / fair treatment controls 和机构内部政策的 exact applicability 取决于 creditor/debt collector role、产品、jurisdiction、servicing context、客户关系、沟通渠道、第三方 vendor arrangement、账户状态和 counsel/compliance interpretation。本文只提供 architecture and governance thinking, 不给出“适用/不适用”“合规/不合规”的法律结论。

---

1. Executive Framing

高管不应把 AI collections 立项描述为“自动化催收”或“降低人工成本”。更成熟的 framing 是:

Build an AI-enabled treatment architecture that reduces credit loss,
detects hardship earlier, prevents harmful contact,
supports dignified repayment options,
and creates evidence of fair customer outcomes.

高管要回答的不是“AI 能提升多少回收率”, 而是:

• 是否能在客户刚进入风险时提供支持, 而不是等到 late-stage pressure?
• 是否能避免多产品、多 vendor、多渠道 over-contact?
• 是否能把 hardship、dispute、fraud、complaint、accessibility 和 vulnerable-customer signals 纳入同一 treatment view?
• 是否能证明 repayment plan 是客户可理解、可负担、可复核的?
• 是否能约束 agent-assist, 防止 AI 生成误导、羞辱、恐吓或 unsupported promise?
• 是否能把投诉、补救、QA 和模型改进闭环?
• 是否能在监管、审计、诉讼或 board review 中重放关键决策证据?

Executive one-liner:

AI collections transformation is a customer-treatment and evidence-control program, not a dialer optimization project.

---

2. Source Anchors

Source nuance:

• CFPB/FTC anchors 是 regulatory language anchors, 不是本文的法律结论。
• NIST AI RMF and ISO/IEC 42001 提供 AI governance scaffolding, 不是单一模型打分表。
• WCAG 2.2 不只属于前端团队, 它影响 hardship self-service、payment plan enrollment、complaint intake、document upload 和 AI-generated notices 的 release gate。

---

3. Operating Principles

---

4. Treatment Taxonomy

Taxonomy 目标是 routing, control and evidence, not moral classification。

Controlled vocabulary:

---

5. Reference Architecture

account and payment systems
  -> delinquency event hub
  -> customer contact profile
  -> complaint / dispute / fraud / hardship services
  -> accessibility and language preference service
  -> AI signal detection and risk tiering
  -> treatment policy and option engine
  -> contact orchestration
  -> customer self-service and agent-assist
  -> specialist handoff
  -> final-channel capture
  -> evidence ledger
  -> QA, model risk, conduct and accessibility monitoring
  -> RCA, remediation and CAPA

Architecture capabilities:

---

6. Decision Gates

Gate 0: Use Case Boundary

Gate 1: Legal/Compliance Applicability Review

Gate 2: Data Purpose and Minimization

Gate 3: Contact Strategy

Gate 4: Treatment Option and Hardship

Gate 5: Agent-Assist

Gate 6: Complaint, Remediation and Learning

---

7. Required Artifacts

---

8. RACI / Operating Model

Governance cadence:

---

9. Implementation Roadmap

Days 1-30: Foundation

Days 31-60: Controlled Design and Pilot Build

Days 61-90: Scale, Assurance and Governance

---

10. Evidence Pack

Minimum evidence fields:

Evidence rules:

• Preserve customer-stated facts without spreading unnecessary sensitive details。
• Store raw trace and curated evidence summary separately。
• Capture what the customer actually saw or heard, not only AI draft。
• Record human decisions and reasons, especially when overriding AI。
• Treat missing evidence as a control defect。
• Vendor events must reconcile to enterprise evidence, not remain in vendor-only files。

---

11. QA / Eval / Model Risk

Scenario eval suite:

Model risk review questions:

• Is AI recommending, drafting or deciding customer-impacting actions?
• Could output affect fees, contact intensity, hardship options, plan terms, disputes, complaints or credit outcomes?
• Which high-harm scenarios are in validation data?
• How are false negatives measured when hardship, complaint or accessibility signals are missed?
• How are false positives measured when customers are over-suppressed, over-escalated or blocked from normal service?
• What drift monitoring detects harsher tone, hallucinated policy, plan unsustainability or vendor mismatch?
• What change control applies to prompt, model, policy content, RAG source, channel UI, vendor workflow or threshold?

Accessibility QA:

• Automated accessibility scans are required but insufficient。
• Manual keyboard and screen reader tests must cover end-to-end payment, hardship, complaint and document upload journeys。
• AI-generated content and dynamic chat states must be tested。
• Documents should have accessible formats or equivalent alternatives。
• Session timeout and save-and-resume must reflect high-stress customer journeys。

---

12. Checklists

12.1 Release Checklist

12.2 Contact Strategy Checklist

12.3 Hardship Treatment Checklist

12.4 Agent-Assist Checklist

12.5 Vendor Checklist

---

13. Metrics and KRIs

Balanced scorecard:

Risk: delinquency and loss are reduced.
Treatment: customers receive sustainable options.
Conduct: pressure, misleading content and over-contact are controlled.
Access: customers can complete high-stakes journeys.
Autonomy: customers have choice, review and complaint routes.
Evidence: decisions and communications are replayable.
Learning: complaints and QA become funded improvements.

---

14. Anti-Patterns

---

15. Tabletop Scenarios

Scenario 1: Early Delinquency and Job Loss

A customer misses the first payment and tells the chatbot they lost their job.
The delinquency model recommends urgent contact because the account is high risk.

Expected decisions: early support vs pressure, hardship route, affordability screen, message content, evidence capture, human handoff and follow-up cadence。

Scenario 2: Over-Contact Across Products

The same customer has a delinquent credit card, personal loan and overdraft.
Three internal teams and one vendor are scheduled to contact them within 48 hours.

Expected decisions: enterprise frequency cap, contact consolidation, product priority, vendor suppression, customer explanation and governance metric。

Scenario 3: Inaccessible Hardship Form

A screen reader user cannot complete the hardship upload flow before a deadline.
The AI chatbot says the online form is required and provides no alternative.

Expected decisions: equivalent access, deadline protection, channel switch, defect severity, remediation, accessibility release gate and vendor escalation。

Scenario 4: Unsupported Agent Promise

An agent-assist draft tells the customer that a fee will be waived if they make
a payment today, but the approved policy only allows review after documentation.

Expected decisions: prompt/content guardrail, employee coaching, customer correction, QA sampling, model eval update and remediation review。

Scenario 5: Complaint During Automated Campaign

A customer files a complaint alleging harassment. The next day, automated SMS
and vendor calls continue because complaint status did not sync to collections.

Expected decisions: complaint hold integration, contact suppression, apology/remediation, RCA, CAPA owner, vendor reconciliation and control testing。

Scenario 6: Fraud Dispute and Delinquent Balance

A customer says the balance became delinquent because of an unresolved fraud
claim. The payment recommender continues to offer a standard plan for the full amount.

Expected decisions: dispute linkage, treatment review, communication wording, plan suppression or adjustment under policy, complaint route and evidence preservation。

---

16. Practical Templates

16.1 Use Case Boundary Card

Use case name:
Customer journey:
Products/accounts covered:
Entity roles and vendors involved:
Channels:
AI functions:
Customer-impacting actions:
Out-of-scope decisions:
Legal/Compliance review owners:
Model risk tier:
Required evidence:
Go/No-Go approvers:

16.2 Treatment Decision Record

Case ID:
Customer-stated situation:
Account status:
Support need:
Contact constraints:
Complaint/dispute/fraud/hardship flags:
Eligible option set:
AI recommendation:
Model uncertainty:
Human decision:
Customer explanation:
Final-channel content ID:
Follow-up date:
QA sample required:

16.3 Data Purpose Matrix

16.4 Approved Hardship Explanation Pattern

What we understand:
Account and balance:
Options available:
What changes if you choose this option:
Fees, dates and limitations:
What we cannot promise here:
What information may be needed:
Accessible format or language support:
Human help and complaint route:

16.5 Complaint RCA Template

Complaint ID:
Customer-stated issue:
AI run ID:
Contact attempt IDs:
Final-channel content IDs:
Treatment decision:
Alleged harm:
Root cause category:
Remediation:
Control gap:
CAPA owner:
Closure evidence:
Recurrence metric:

16.6 Executive One-Pager

Program objective:
Customer journeys covered:
Top customer harms reduced:
AI capabilities:
Key guardrails:
Human review model:
Complaint and remediation linkage:
Evidence readiness:
Top metrics:
Residual risks:
Decisions needed:

---

17. Portfolio Deliverables

Portfolio storyline:

I designed an AI-enabled collections treatment architecture for financial retail.
It detects delinquency and hardship signals early, respects contact preferences,
offers sustainable repayment and hardship options, protects vulnerable situations,
guards agent assistance, links complaints to evidence, and measures fair customer outcomes.

---

18. Interview Answers

Q1: AI collections 的高级架构目标是什么?

30 秒:

目标不是更高压地催收, 而是 treatment orchestration。AI 应帮助早期识别 delinquency and hardship, 控制 contact strategy, 展示 approved options, 支持 agent-assist, 连接 complaint/remediation, 并通过 evidence ledger 证明客户被公平、有尊严、可访问地处理。

Q2: 如何避免 AI repayment recommender 造成 conduct risk?

30 秒:

我会把 recommender 放在 policy-approved option universe 内, 加 affordability guardrail、plain-language explanation、unsupported-promise block、human review、final-channel capture 和 plan sustainability metrics。模型不能单独决定高影响 treatment, 更不能为了短期回款推荐不可承受方案。

Q3: Contact strategy 该怎么治理?

30 秒:

Contact strategy 需要 enterprise contact ledger, 跨产品和 vendor 限频去重, 并检查 preference、language、accessibility、complaint/dispute/hardship holds。AI 可以优化时机和渠道, 但必须先通过 conduct and customer-context controls。

Q4: 投诉如何接入 AI collections architecture?

30 秒:

投诉记录要链接 AI run、contact attempt、final-channel content、treatment decision、human reason、remediation 和 RCA。投诉主题要反馈到 eval suite、content library、contact controls、vendor CAPA 和 model monitoring, 否则系统只是在事后灭火。

Q5: 面试中如何体现高级 PM/架构师判断?

30 秒:

我会强调 balanced scorecard: roll rate 和 loss 只是其中一部分, 还要看 plan durability、complaints、over-contact、accessibility completion、false negative hardship signals、unsupported script defects、final-channel capture 和 CAPA aging。高级架构不是单点模型, 而是 customer treatment operating system。

---

19. Final Operating Principle

这套 playbook 的成熟度可以用一个问题检验:

When a customer becomes delinquent or asks for help,
does the AI system reduce risk while preserving dignity, access,
choice, affordability, complaint rights, human judgment and evidence?

如果答案不清楚, 企业不是缺一个更聪明的催收模型。问题是 collections strategy、hardship operations、conduct risk、accessibility、model risk、frontline operations、vendor oversight 和 complaint governance 还没有成为同一套 treatment operating system。