AI Personalized Pricing：个性化定价与 Offer 治理架构
AI personalized pricing 不是更聪明的 campaign engine。它是一套 governed economic decisioning system：决定客户看到的 rate、fee、limit、promotion、retention offer、loyalty incentive 或 servicing term，也决定机构如何解释、复核、补救和证明这些差异化待遇。
AI Personalized Pricing / Offer Decisioning / Surveillance Pricing 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_PERSONALIZED_PRICING_OFFER_DECISIONING_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 AI 驱动 rates、fees、credit limits、promotions、retention offers、next-best-actions、loyalty incentives 和 personalized terms 的产品与架构治理，不构成法律、监管、信用审批、定价合规、消费者通知、模型验证、隐私影响评估、conduct risk 审查或 vendor endorsement 结论。具体法律框架适用性取决于 product、decision type、customer segment、jurisdiction、data source、model use、offer presentation、customer impact 和 Legal/Compliance interpretation。
Source Anchors
| Source | Link | 用途 |
|---|---|---|
| FTC Surveillance Pricing feature page | https://www.ftc.gov/news-events/features/surveillance-pricing | 用作 surveillance pricing / individualized pricing concern 的官方锚点 |
| FTC 6(b) orders on surveillance pricing intermediaries | https://www.ftc.gov/news-events/news/press-releases/2024/07/ftc-issues-orders-eight-companies-seeking-information-surveillance-pricing | 用作 FTC 对 surveillance pricing products and services 信息收集关注点的锚点 |
| FTC Commercial Surveillance and Data Security rulemaking | https://www.ftc.gov/legal-library/browse/federal-register-notices/commercial-surveillance-data-security-rulemaking | 用作 commercial surveillance、data security、consumer data practices 和 dark patterns 风险讨论的锚点 |
| CFPB Circular 2022-03: adverse action notices and complex algorithms | https://www.consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms/ | 用作 complex algorithm credit decision 中 reason specificity / adverse action handoff 的锚点 |
| CFPB Consumer Complaint Database | https://www.consumerfinance.gov/data-research/consumer-complaints/ | 用作 complaints as monitoring signal / evidence loop 的锚点 |
| NIST AI RMF | https://www.nist.gov/itl/ai-risk-management-framework | 用 Govern / Map / Measure / Manage 组织 AI pricing decision governance |
| NIST Privacy Framework | https://www.nist.gov/privacy-framework | 用 privacy risk、data processing、customer trust 和 data minimization 组织 feature boundary |
| ISO/IEC 42001 overview | https://www.iso.org/standard/42001 | 用 AI management system、roles、operation、performance evaluation、audit 和 continual improvement 建立 operating model |
核心导读
AI personalized pricing 不是更聪明的 campaign engine。它是一套 governed economic decisioning system：决定客户看到的 rate、fee、limit、promotion、retention offer、loyalty incentive 或 servicing term，也决定机构如何解释、复核、补救和证明这些差异化待遇。
AI 改变的是经济优化链。传统 offer engine 可能只做 segment、规则和活动编排；AI 会把 eligibility、risk、affordability、propensity、elasticity、uplift、optimizer、experiment allocator、channel copy 和 adverse-action/review handoff 串在一起。价值来自更精细的候选生成和约束优化，风险也来自同一个位置：如果 willingness-to-pay、digital behavior、financial stress、complaint tone、location、third-party segment 或 proxy attribute 被混入价格决策，系统可能把个性化变成对低议价能力客户的提取。
学习这篇时，要把“为什么这个客户得到这个条款”作为主线。证据链应覆盖 approved feature、prohibited feature check、policy gate、candidate set、model outputs、optimizer config、experiment arm、reason mapping、customer copy、human override 和 complaint/remediation linkage。治理边界是：propensity 不能替代 eligibility，risk reason 不能掩盖 willingness-to-pay，hardship/vulnerability signal 只能用于保护或 suppression，实验必须有 harm cap 和 stop rules，解释服务必须在设计时存在，而不是投诉后再补。
问题定义
金融零售里的价格和条款通常同时表达四件事：
institution economics
+ customer eligibility and risk
+ customer treatment and trust
+ regulatory / conduct / evidence obligations
所以个性化定价不能被定义为“给每个客户最可能接受的价格”。成熟架构必须能证明：
为什么这个客户收到这个 rate、fee、limit、incentive 或 term?
哪些 policy、data 和模型允许这个决定?
还有哪些 eligible alternatives?
protected/proxy attributes 是否被控制?
实验是否有 harm cap、stop rules 和 remediation?
客户可见解释是否准确、具体且不过度承诺?
投诉或审计时能否重放 feature、model、policy、experiment、copy 和 human action?
Surveillance pricing concern 的关键不是所有差异化都不可接受。金融服务已有 risk-based pricing、relationship pricing、segment offers 和 retention concessions。高风险点在于使用 granular behavioral、device、location、browsing、psychographic、life-event、financial-stress 或 third-party surveillance data 推断 willingness-to-pay、urgency、low bargaining power 或 vulnerability，并给相似风险客户系统性更差条款。
核心原理/方法
Pricing/offer decision 应拆成多个 decision planes：
| Decision plane | 回答什么 | 不能混淆成什么 |
|---|---|---|
| Eligibility | 客户是否符合产品、渠道、风险、政策和运营要求 | 不能让 propensity 替代 eligibility |
| Risk and affordability | credit loss、fraud、capacity、servicing risk 如何影响条款 | 不能用 willingness-to-pay proxy 伪装成风险 |
| Economics | rate、fee、limit、incentive 如何影响 NPV/CLV/margin | 不能只最大化短期 revenue |
| Customer treatment | 客户是否被一致、可解释、不过度利用地对待 | 不能把“客户会接受”当成公平 |
| Experimentation | 价格/offer 如何测试，对谁测试，伤害上限是什么 | 不能把客户当成无限探索样本 |
| Explanation/evidence | 决策如何解释、复核、投诉处理和审计重放 | 不能只记录 final offer |
核心原则：
policy first,
model assisted,
optimizer constrained,
experiment bounded,
explanation ready,
complaint learnable,
evidence replayable.
系统/架构模型
参考架构：
source systems and consented data
-> data classification and feature boundary
-> protected/proxy attribute controls
-> customer eligibility and policy filters
-> risk / affordability / fraud models
-> propensity / elasticity / uplift models
-> pricing and offer candidate generator
-> constrained optimizer / rules engine
-> fairness, conduct and trust guardrails
-> experimentation allocator / holdout manager
-> decision and explanation service
-> channel orchestration and customer copy
-> adverse-action / reason / review handoff where applicable
-> complaints, servicing and customer feedback loop
-> monitoring, evidence ledger and governance review
关键层：
| Layer | 职责 |
|---|---|
| Feature registry | 标记 first-party、third-party、sensitive、protected、proxy、behavioral surveillance、consent |
| Eligibility policy | 先确定 product/channel/customer/jurisdiction/risk eligibility |
| Risk models | credit、fraud、affordability、loss、prepayment、servicing risk |
| Propensity/elasticity | 预测接受、流失、使用、响应、price sensitivity，但不得越界到 extraction |
| Candidate generator | 只生成 approved offer grid 中允许的候选 |
| Constrained optimizer | 在 policy、risk、fairness、economics 约束下选择 |
| Experiment allocator | 控制 randomization、holdout、bandit、harm cap、stop rules |
| Explanation service | 生成 internal reason、customer message、review packet |
| Evidence ledger | 保存 feature、model、policy、experiment、decision、copy、human action |
关键机制与取舍
Product economics 要显性化：
Expected Value =
expected interest income
+ expected fee income
+ interchange / partner revenue
+ deposit spread or funding value
+ loyalty / relationship lift
- expected credit loss
- funding and capital cost
- acquisition / servicing / rewards cost
- fraud / dispute / complaint / remediation cost
- cannibalization and adverse selection cost
- conduct / trust / attrition risk cost
差异化依据：
| Basis | Example | Governance view |
|---|---|---|
| Risk-based | 更高 default/fraud risk 对应更高 APR 或更低 limit | 需要 risk reason、model governance、fairness monitoring |
| Cost/value-based | 高余额关系获得 fee waiver 或 deposit tier | 需要 published or auditable criteria |
| Behavioral willingness-to-pay | 越急、越少比较、越依赖某渠道，价格越差 | surveillance pricing / conduct risk 高 |
Feature boundary：
| Data class | Examples | 定价/offer 使用边界 |
|---|---|---|
| Product/account facts | tenure、balances、payment history、relationship tier | 可用于政策和 relationship pricing，受 purpose/consent 限制 |
| Credit/risk variables | bureau、delinquency、income verification、affordability | 可用于 approved risk-based terms，需要 explainability |
| Transaction behavior | spend category、cashflow volatility、payroll pattern | 区分 risk/financial health 与 lifestyle exploitation |
| Channel/digital behavior | clickstream、device、session urgency、comparison behavior | 高 surveillance risk，通常不应直接提高 price/fee |
| Location/context | geo、branch area、travel pattern | proxy risk 高，需要审查 |
| Third-party marketing data | demographic append、propensity segment、broker score | 高 consent/proxy/explainability risk |
| Vulnerability/hardship signals | complaint tone、bereavement、financial stress | 只用于 support/protection，不用于 price extraction |
| Protected/proxy attributes | protected classes and proxies | 具体用途需法律/合规判断；通常用于 monitoring rather than setting |
Experiment guardrails：
| Guardrail | Design rule |
|---|---|
| Harm cap | 定义 customer-level incremental cost / lost benefit 上限 |
| Stratification | 按 risk、channel、region、monitoring group 分层 |
| Stop rules | complaint spike、fairness gap、loss spike、misleading copy、ops overload 触发暂停 |
| Remediation | 预定义 credit/refund/reprice path |
| Evidence | 记录 assignment probability、arm、duration、copy、rationale |
| Exploitation limit | bandit 不得长期把更差条款分配给“更容易接受坏条件”的群体 |
证据与控制
Decision evidence record：
decision_id
customer/product/channel
decision lever: rate / fee / limit / promotion / retention / servicing term
eligibility rule version
candidate offer set
approved features and lineage
prohibited features checked
risk model outputs and version
propensity/uplift/elasticity outputs and version
optimizer configuration
fairness/conduct guardrail results
experiment id / arm / assignment probability
selected offer / alternatives
reason codes and explanation mapping
customer copy version
human review / override if any
complaint/remediation linkage
控制矩阵：
| Control objective | Control activity | Evidence |
|---|---|---|
| 定义 decision scope | classify rate、fee、limit、promotion、retention、servicing、NBA | decision inventory、use case card |
| 约束 eligibility | optimization 前应用 product/channel/customer/jurisdiction/risk gates | rule log、policy version |
| 治理 data use | feature registry with sensitivity and allowed levers | Feature Use Card、lineage |
| 防 protected/proxy misuse | feature review + outcome monitoring | proxy review、fairness dashboard |
| 分离 risk 与 elasticity | 标记 risk factors 与 willingness-to-pay/propenstiy factors | model docs、feature map |
| 约束 optimizer | approved grid、min/max、guardrails、reason codes | candidate set、optimizer config |
| 治理 experiments | harm cap、stratification、stop rules、remediation | experiment charter、assignment log |
| 支持 explanation | model/policy drivers 映射 approved reasons | reason taxonomy、trace |
| handoff adverse action where applicable | relevant credit decisions route to notice/review workflow per policy | handoff record、final notice reference |
| 保护 vulnerable signals | hardship/complaint/vulnerability firewall | feature rules、audit sample |
| 监控 conduct risk | complaints、overrides、exceptions、fee disputes、retention inconsistency | complaint analysis、conduct review |
| 保存 evidence | feature、model、policy、experiment、decision、copy、human action | evidence bundle、replay test |
指标应同时看 economics、customer outcome、fairness/proxy、conduct/trust、model stability、experiment safety、explanation quality、operations 和 evidence completeness。不要只展示 revenue uplift，而不展示谁承担成本、谁失去 benefit、哪些群体条款更差、投诉是否增加、解释是否充分。
金融零售/AI产品场景
- Credit card APR/limit：risk model 和 pricing table 分离；adverse-action/reason handoff 在 decision service 中预留，不靠后补。
- Deposit relationship rate：relationship criteria 可审计，不能让 hidden propensity 决定谁看到更高 rate。
- Annual fee retention offer：retention policy 防止“会投诉的人得到更好待遇、沉默客户没有路径”的不一致。
- Balance transfer promotion：eligibility、limit、APR、fee、experiment arm、customer copy 和 conduct guardrails 全部进入 evidence record。
- Hardship-related offer：hardship/complaint/vulnerability signal 只能触发 support、sales suppression 或 review，不能触发更高 fee 或差条件。
- Personalized merchant offer：transaction behavior 可用于 relevance，但不得从敏感消费推断 protected traits 或用于 price extraction。
反模式
| 反模式 | 风险 | 更好的控制 |
|---|---|---|
| propensity model 变 pricing engine | 优化谁会接受更差条款 | propensity 与 approved pricing constraints 分离 |
| risk 和 willingness-to-pay 混合 | 无法解释高 APR 是风险还是提取 | feature labeling and model decomposition |
| customer 360 unrestricted | sensitive/proxy/surveillance data 泄漏进价格 | feature registry and allowed-use policy |
| silent non-selection at scale | 客户从不知道被排除在更好条款之外 | eligibility governance and monitoring |
| bandit 锁定更差 offer | exploration 变 unequal exploitation | stratified guardrails、regret monitoring |
| retention offer 不一致 | 投诉/威胁解锁隐藏利益 | retention policy、frontline tooling、QA |
| LLM 编造原因 | 客户收到不准确解释 | reason-code constrained generation |
| explanation late bolt-on | 缺少 reasons and evidence | explanation architecture at design time |
| fairness 只看模型层 | policy、experiment、manual override 逃逸 | end-to-end decision monitoring |
| complaint 不链接 decision trace | 无法补救系统性 conduct issue | complaint-to-evidence linkage |
最终心智模型
个性化定价治理的核心不是让推荐模型更准，而是把 pricing economics、policy decisioning、feature boundary、fairness/conduct controls、experimentation ethics、explanation handoff 和 evidence replay 组成同一套 operating system。成熟机构应能证明：每个 personalized rate、fee、limit、promotion、retention offer 或 next-best-action 都来自 approved data、approved policy 和 bounded experiment，客户待遇可解释，投诉可追溯，弱势信号不被利用，经济收益没有建立在不可辩护的 surveillance pricing 之上。
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」。