AI Personalized Pricing / Offer Decisioning / Surveillance Pricing Governance Playbook

配对阅读：本手册的原理/架构解读版是 docs/ai-foundations/papers/136-ai-personalized-pricing-offer-decisioning-governance-architecture.md。先读 paper 建立机制与取舍，再用本手册落地为模板、RACI 与门禁，两者不需要重复精读。 核心导读: personalized pricing 的难点不是“更懂客户”, 而是把 risk-based pricing、relationship value、offer relevance、experimentation、fair lending、conduct risk、privacy 和 customer harm 放进同一个受控 decisioning system。 学习主线: 从 decision taxonomy 出发, 建立 feature boundary、offer grid、policy constraints、experiment guardrails、reason handoff、complaint feedback、fairness monitoring 和 evidence replay 的完整链路。 系统边界: 本手册用于系统学习、架构训练和内部治理讨论, 不构成法律意见、监管意见、信用审批结论、定价合规结论、消费者通知建议、模型验证报告、隐私影响评估结论、conduct risk 审查结论或供应商推荐。 证据输出: 目标输出是一套可解释、可监控、可审计的 financial retail decisioning evidence pack, 覆盖 executive framing、decision taxonomy、source anchors、decision gates、required artifacts、RACI、implementation roadmap、release checklists、metrics/KRIs、anti-patterns、tabletop scenarios 和 practical templates。

核心判断:

A financial institution should not scale AI personalized pricing until it can explain the difference between customer relevance, risk-based pricing, relationship value, bounded experimentation and exploitative surveillance pricing.

---

0. Disclaimer

本文是系统学习、架构训练和内部治理讨论材料, 不构成法律意见、监管意见、信用审批结论、定价合规结论、消费者通知建议、模型验证报告、隐私影响评估结论、conduct risk 审查结论或供应商推荐。

本文不判断 ECOA、FCRA、UDAP、UDAAP、FTC Act、state pricing laws、privacy laws、fair lending rules 或其他具体法律框架是否适用于某个产品或决策。精确适用性取决于 product、decision type、customer segment、jurisdiction、channel、data source、contract terms、model use、offer presentation、customer impact 和 Legal / Compliance interpretation。

正式落地必须由 Legal、Compliance、Privacy、Fair Lending / Conduct Risk、Credit Risk、Pricing Strategy、Model Risk、Data Governance、Information Security、Customer Experience、Operations、Complaint Management、Product、Architecture、Experimentation、Vendor Management、Internal Audit 和 senior management 共同审查。

边界原则:

• AI personalization 可以提升 relevance, 但不能绕过 pricing policy。
• Risk-based pricing 需要可解释、可验证、可监控的 risk basis。
• Willingness-to-pay optimization 在金融服务中具有高 surveillance pricing and conduct risk。
• Protected attributes and proxy attributes 的具体法律处理取决于 context and Legal / Compliance interpretation, 但架构上必须进行 feature review and outcome monitoring。
• Adverse action / notice / reason handoff 是否适用由 Legal / Compliance 判断, 但 decision system 必须具备 evidence and reason capability。
• Complaints are governance data, not only operations workload。

---

1. Executive Framing

高管常见叙事:

Use AI to personalize offers.
Use real-time data to optimize prices.
Use bandits to maximize conversion.
Use next-best-action to improve customer value.

高级治理叙事应改成:

Use AI to select approved offers for eligible customers,
within explicit pricing, risk, fairness, conduct and experimentation constraints,
with explainable reasons and replayable evidence.

1.1 Executive Risk Questions

1. 哪些 decisions 会改变客户经济条件: rate、fee、limit、promotion、retention、loyalty、term?
2. 哪些差异化来自 risk, 哪些来自 relationship, 哪些来自 experiment, 哪些来自 inferred willingness-to-pay?
3. 哪些 features 被禁止用于 price-setting, 尤其是 protected/proxy、vulnerability、complaint、device、location、behavioral urgency 和 third-party surveillance signals?
4. 如果客户问“为什么我比别人贵 / 为什么我没有这个 offer”, frontline 能否给出准确解释?
5. 如果投诉、审计或监管质询发生, 是否能重放数据、模型、政策、实验、文案和人工处理?
6. Experiment 的 harm cap、stop rules、remediation path 和 fairness monitoring 是否在上线前存在?
7. Revenue uplift 是否和 complaints、fairness、customer trust、lost benefit、evidence completeness 一起汇报?

1.2 Board-Level One-Liner

The control objective is not to prevent all personalization.
The control objective is to prevent unexplainable, unfair, exploitative or unbounded economic differentiation.

---

2. Source Anchors

Source-to-control pattern:

source anchor -> risk theme -> control objective -> product gate
  -> artifact -> evidence field -> owner -> metric

---

3. Decision Taxonomy

3.1 Economic Levers

3.2 Decision Impact Classes

3.3 Differentiation Basis

---

4. Target Operating Architecture

decision inventory
  -> use case risk tier
  -> data and feature registry
  -> protected/proxy and privacy review
  -> eligibility and product policy gates
  -> approved offer grid / candidate library
  -> risk, propensity, uplift and economics models
  -> constrained optimizer
  -> fairness, conduct and trust guardrails
  -> experiment allocator
  -> explanation and adverse-action handoff service
  -> channel presentation and customer copy
  -> complaint / servicing / appeal loop
  -> monitoring dashboard
  -> evidence ledger and governance forum

Operating principles:

---

5. Decision Gates

Gate 0: Decision Inventory and Risk Tier

Gate 1: Product Economics and Policy Baseline

Gate 2: Data and Feature Boundary

Gate 3: Model and Optimizer Governance

Gate 4: Experimentation

Gate 5: Explanation, Adverse Action and Customer Copy

Gate 6: Launch Monitoring

Gate 7: Lifecycle and Change Control

---

6. Required Artifacts

6.1 Product Economics Memo Structure

6.2 Approved Offer Grid Example

The actual ranges, factors and notices must be approved by the institution's product, risk, pricing, legal and compliance teams.

---

7. RACI / Operating Model

Governance cadence:

---

8. Implementation Roadmap

Days 1-30: Baseline and Containment

Days 31-60: Controlled Build

Days 61-90: Pilot and Assurance

Days 91-180: Scale and Continuous Governance

---

9. Evidence Pack

Minimum fields for every material pricing / offer decision:

Evidence rules:

• Preserve customer-facing copy exactly as displayed.
• Record candidate offers, not only final offer.
• Store policy version, model version and experiment assignment together.
• Capture reason-code mapping before LLM or content transformation.
• Mark inferred attributes separately from verified/customer-declared data.
• Link complaints to original decision evidence.
• Treat missing evidence as a control failure.

---

10. Checklists

10.1 Release Checklist

10.2 Feature Boundary Checklist

10.3 Experiment Checklist

10.4 Explanation / Handoff Checklist

10.5 Complaint Learning Checklist

---

11. Metrics and KRIs

Balanced scorecard:

Economics: uplift is real after risk and cost.
Fairness: similarly situated customers are treated consistently.
Trust: customers understand material differences.
Privacy: sensitive and surveillance-like data is controlled.
Conduct: vulnerability and complaints are protective signals, not extraction signals.
Experimentation: exploration has harm caps and remediation.
Explainability: decisions have specific, accurate reasons where needed.
Evidence: every material term decision can be replayed.

---

12. Anti-Patterns

---

13. Tabletop Scenarios

Scenario 1: Surveillance Pricing Allegation

A media report claims the institution charges higher loan APRs to customers
who apply late at night from mobile devices and abandon comparison pages.

Expected decisions: feature registry review, prohibited feature test, customer expectation analysis, proxy monitoring, public response evidence, model retrain or suspension if needed.

Scenario 2: Bandit Learns to Withhold Better Offers

A contextual bandit discovers that a subset of customers accepts lower signup bonuses,
so the system increasingly withholds higher bonuses from them.

Expected decisions: lost-benefit analysis, fairness review, harm cap trigger, experiment pause, remediation plan, objective redesign.

Scenario 3: Credit Limit Counteroffer Without Specific Reasons

A customer requests a $20,000 credit limit. The AI decisioning system offers $5,000
but only records "model decision" as the reason.

Expected decisions: adverse-action / notice handoff review by Legal/Compliance, reason reconstruction, decision-service defect, release gate failure.

Scenario 4: Retention Offer Inconsistency

Customers who threaten to close accounts through chat receive fee waivers,
while similar customers who call politely receive no waiver.

Expected decisions: retention policy review, channel consistency, frontline scripts, complaint trend analysis, customer remediation.

Scenario 5: Vulnerability Signal Misuse

The model identifies customers under financial stress and targets them with fee-bearing
cash advance offers because response probability is high.

Expected decisions: vulnerability firewall, conduct incident review, offer suppression, financial-health alternative, governance escalation.

Scenario 6: Third-Party Segment Drives Pricing

A data broker segment improves margin prediction, but model review finds it strongly
correlates with geography, income estimate and language preference.

Expected decisions: proxy review, purpose limitation, feature exclusion or restriction, vendor due diligence, monitoring evidence.

---

14. Practical Templates

14.1 Use Case Boundary Card

Use case:
Product:
Customer segment:
Jurisdiction / policy scope:
Channel:
Decision lever:
Customer impact class:
AI role:
Business objective:
Customer benefit:
Customer harm scenario:
Decision owner:
Risk owner:
Legal/Compliance handoff owner:
Launch forum:

14.2 Pricing / Offer Policy Rule

Rule ID:
Product:
Eligible population:
Excluded population:
Allowed offers:
Rate / fee / limit range:
Risk factors allowed:
Relationship factors allowed:
Promotional factors allowed:
Features prohibited:
Protected/proxy monitoring:
Manual review triggers:
Customer reason codes:
Notice/adverse-action workflow:
Evidence fields:
Review cadence:

14.3 Feature Review Form

Feature:
Source:
Lineage:
Purpose:
Decision levers requested:
Customer expectation:
Sensitivity:
Protected/proxy concern:
Third-party/vendor involvement:
Allowed uses:
Prohibited uses:
Explanation mapping:
Monitoring metric:
Owner:
Approval decision:

14.4 Experiment Charter

Experiment name:
Decision lever:
Hypothesis:
Arms:
Eligible customers:
Excluded customers:
Randomization method:
Duration:
Harm cap:
Stop rules:
Protected/proxy monitoring:
Complaint monitoring:
Customer copy:
Reason / notice handling:
Remediation path:
Evidence fields:
Approvers:

14.5 Model / Optimizer Change Record

Change ID:
Model / optimizer:
Previous version:
New version:
Objective change:
Feature change:
Policy constraint change:
Expected economic impact:
Expected customer impact:
Fairness/proxy assessment:
Reason mapping impact:
Experiment impact:
Evidence impact:
Approval:
Rollback condition:

14.6 Complaint RCA Form

Complaint ID:
Decision ID:
Customer allegation:
Product and channel:
Offer / term shown:
Comparable offer history:
Eligibility rule:
Model version:
Experiment arm:
Reason provided:
Customer copy:
Protected/proxy flag:
Root cause category: data / feature / model / policy / experiment / copy / operations / vendor
Customer remediation:
Control remediation:
CAPA owner:
Closure evidence:

14.7 Executive Memo Skeleton

Decision:
Why now:
Customer value:
Institution economics:
Key risks:
Data boundaries:
Fairness and conduct controls:
Experiment controls:
Explanation and complaint path:
Evidence and auditability:
Go / restrict / redesign / stop recommendation:

---

15. System Evidence Pack

Operating narrative:

AI pricing and offer decisioning should be governed as a financial retail decision architecture.
The system separates eligibility, risk, economics, experimentation and customer treatment;
blocks sensitive and surveillance-like feature misuse;
uses approved offer grids and constrained optimization;
supports reason and adverse-action handoff where applicable;
monitors fairness, complaints and conduct risk;
and preserves evidence for every material customer term decision.

---

16. Decision Narratives

16.1 Opportunity Boundary: relevance is not unlimited price extraction

AI personalized pricing 的机会是更准确地匹配 risk、relationship value 和客户需求, 提升 conversion、margin、retention 和 financial health。边界是不能把 customer 360 和 willingness-to-pay prediction 直接变成价格。金融零售需要 approved offer grid、feature boundary、fairness/conduct controls、experiment harm caps、reason handoff、complaint monitoring 和 evidence replay。

16.2 Surveillance Pricing Control: feature boundary before optimization

治理 surveillance pricing 风险要先把 data source 和 feature use 做成 registry, 明确哪些数据可用于 risk, 哪些只能用于 service, 哪些禁止用于 price-setting。对 device、location、behavioral urgency、third-party segments、vulnerability and complaint signals 做 no-surprise and proxy review。再用 outcome monitoring、complaints、reason QA 和 evidence replay 验证没有把客户弱势或低议价能力转成更差条款。

16.3 Experiment Design: monetary treatment needs harm control

AI pricing experiment 影响客户真实经济条件, 所以需要 harm cap、stratified randomization、stop rules、remediation path、fairness monitoring 和 customer-copy evidence。Bandit 不能只因为某群体接受差条款就长期给他们差条款。实验成功标准不能只有 revenue uplift, 还要看 lost benefit、complaints、fairness and trust。

16.4 Reason Handoff: explain the decision, not the model in isolation

系统不先替代法律适用判断, 但架构必须准备好。每个 credit-term decision 应记录 policy rule、model factors、candidate offers、selected term、reason codes、customer message and evidence。若 Legal/Compliance 判断需要 notice or adverse-action path, 系统可以输出具体、准确、可审计原因, 而不是 generic model explanation。

16.5 Cross-Functional Architecture Value: align economics, policy and evidence

这个主题的高级价值不在于写功能需求, 而在于把 economics、policy、data、model、experiment、customer treatment、complaints and evidence 翻译成可执行 decision gates and artifacts。Pricing、Risk、Compliance、Model Risk、Privacy、CX、Ops 必须在同一张 decision map 上决策, 这是 AI 金融零售产品治理的核心能力。

---

17. Final Operating Principle

这套 playbook 的成熟度可以用一个问题检验:

When an AI-enabled financial retail product gives a customer a personalized rate,
fee, credit limit, promotion, retention offer, next-best-action, loyalty incentive or term,
can the institution prove the decision was eligible, policy-bound, data-governed,
fairness-monitored, explanation-ready, complaint-learnable, experiment-safe
and evidence-replayable?

如果答案不清楚, 不要急着扩展模型或接入更多 data broker。先把 product economics、pricing policy、feature boundary、experiment governance、explanation handoff、complaint learning 和 evidence architecture 建起来。