目录
AI Credit Lifecycle / Underwriting / Line Management Governance Playbook
定位: 面向 CBAP+、高级 AI PM、Senior BA、Product Architect、Credit Risk、Underwriting Strategy、Account Management、Pricing Strategy、Fair Lending / Compliance、Model Risk、Portfolio Risk、Complaint Operations、Customer Experience、Data Governance 和 Internal Audit, 把 AI 信用生命周期设计成可上线、可运营、可监控、可审计、可持续改进的 decisioning operating system。
适用范围: credit card、personal loan、BNPL、auto-like unsecured/secured lending journey、retail line-of-credit、SMB-lite credit journey、deposit overdraft/credit-adjacent exposure、embedded finance credit offers、account management line actions。
核心产出: executive framing、decision taxonomy、source anchors、decision gates、required artifacts、RACI、control matrix、evidence schema、metrics/KRIs、eval pack、complaint/remediation loop、model-risk change control、tabletop scenarios 和面试表达。
核心判断:
A financial institution should not scale AI credit decisioning until it can prove how access, price, line, reason, override, complaint and portfolio outcomes are governed as one lifecycle system.
0. Disclaimer
本文是学习、作品集、架构训练和内部治理讨论材料, 不构成法律意见、监管意见、合规结论、信用审批结论、定价建议、adverse action notice 建议、fair lending 结论、模型验证报告、消费者报告使用结论、投诉处理意见、审计意见或供应商推荐。
本文不判断 ECOA、Regulation B、FCRA、UDAAP、state lending laws、fair lending requirements、consumer-reporting obligations、record retention、model risk guidance 或其他规则是否适用于某个产品、模型、渠道、客户群或决策。精确适用性取决于 product、decision type、customer segment、jurisdiction、data source、consumer-reporting use、pricing method、line action、notice workflow、vendor role、contract terms 和 Legal / Compliance interpretation。
正式落地必须由 Legal、Compliance、Fair Lending、Credit Risk、Pricing Strategy、Model Risk、Data Governance、Privacy、Information Security、Customer Experience、Operations、Complaint Management、Portfolio Risk、Internal Audit、Vendor Management、Product Owner、Architecture 和 senior management 共同审查。
边界原则:
AI 可以辅助风险识别、排序、抽取、推荐、监控和解释生成, 但不能绕过 approved credit policy。
Underwriting、pricing 和 line management 必须分层, 不能让一个 optimizer 同时决定客户访问、价格、额度和原因。
Reason codes、notices、customer explanations 和 exact applicability 由 Legal / Compliance / policy owners 确认; 产品和架构负责提供事实、证据和可执行能力。
Complaints、appeals、overrides 和 portfolio outcomes 是 governance data, 不只是运营结果。
1. Executive Framing
不成熟的高管叙事:
Use AI to approve faster.
Use ML to increase approval rate.
Use automation to reduce underwriting cost.
Use account behavior to optimize lines.
成熟叙事:
Build an AI-enabled credit lifecycle platform that makes approved, explainable and monitored decisions about eligibility, risk, price, line and account actions, while preserving customer recourse, model-risk discipline and portfolio feedback.
1.1 Executive Questions
哪些 AI decisions 会改变客户的 credit access、cost、line、servicing friction 或 ability to appeal?
哪些差异来自 approved risk policy, 哪些来自 pricing economics, 哪些来自 promotion/experiment, 哪些可能来自 behavioral proxy?
对 prequalification、underwriting、line increase、line decrease、counteroffer、pricing handoff 是否有统一 decision id 和 evidence bundle?
如果客户问“为什么我被拒绝/额度低/降额/没有提额/价格更高”, frontline 能否给出准确、受控、可追溯的解释?
人工 override 是风险接受、客户纠错、政策例外, 还是隐藏的非治理模型?
如果投诉、监管问询、内审或模型验证发生, 是否能重放当时数据、模型、政策、价格、额度、文案和人工动作?
Portfolio dashboard 是否同时展示 loss、approval、line utilization、complaint、appeal、override、reason drift 和 segment outcomes?
1.2 Board-Level One-Liner
The control objective is not to prevent AI underwriting.
The control objective is to prevent ungoverned credit decisions whose access, price, line, reason and customer recovery path cannot be explained or monitored.
2. Source Anchors
Anchor Official link Playbook 使用方式 CFPB Regulation B / ECOA https://www.consumerfinance.gov/rules-policy/regulations/1002/ 信用生命周期、application handling、adverse action、specific reasons、records 和 fair-lending scope 的官方锚点; 具体版本和适用性由 Legal / Compliance 判断 CFPB Circular 2022-03: adverse action 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/ complex algorithm credit decisions 中 reason specificity, black-box limitation, reason handoff 的设计锚点 CFPB Consumer Complaint Database https://www.consumerfinance.gov/data-research/consumer-complaints/ complaint taxonomy, external signal calibration, customer harm root cause, remediation and monitoring loop Federal Reserve SR 11-7 https://www.federalreserve.gov/supervisionreg/srletters/sr1107.htm 传统 MRM 历史心智模型: model inventory, validation, ongoing monitoring, governance, effective challenge OCC Bulletin 2011-12 https://www.occ.treas.gov/news-issuances/bulletins/2011/bulletin-2011-12.html 用户指定的 OCC 历史锚点; 当前访问可能跳转, 本文仅作为 2011 MRM 学习锚点 Federal Reserve SR 26-2 https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm 截至 2026-06-30 的 updated MRM anchor: supersedes and replaces SR 11-7, emphasizes risk-based MRM OCC Bulletin 2026-13 https://www.occ.gov/news-issuances/bulletins/2026/bulletin-2026-13.html 截至 2026-06-30 的 OCC updated MRM anchor: risk-based, tailored, vendor/third-party considerations NIST AI RMF https://www.nist.gov/itl/ai-risk-management-framework 用 Govern / Map / Measure / Manage 组织 AI credit lifecycle risk management ISO/IEC 42001 https://www.iso.org/standard/42001 用 AI management system、roles、operation planning、performance evaluation、internal audit、continual improvement 设计 operating model
Source-to-control mapping:
source anchor
-> internal policy interpretation
-> control objective
-> product requirement
-> gate
-> artifact
-> evidence field
-> owner
-> metric
3. Operating Principles
Principle Practical meaning Decision inventory first 先盘点所有 customer-impacting credit decisions, 再谈模型 Policy before model Eligibility, product rules, exposure caps and prohibited uses execute before optimization Underwriting is not pricing risk decision, price decision and line decision are separate but traceable Line is a customer-impacting term initial line, CLI and CLD require evidence, monitoring and customer communication design Reason-ready by design reason code, evidence ref and notice/customer copy path are design inputs Human review is governed override, second-look, appeal and manual exception are monitored decision paths Complaints are signals complaint and appeal outcomes feed policy, model, reason, copy and remediation Portfolio feedback is multidimensional loss reduction is insufficient if approval, line, complaint, appeal or fairness signals degrade Evidence is runtime audit evidence is generated during decision execution, not assembled from screenshots Legal/Compliance owns applicability product and architecture teams own fact patterns, workflow, controls and evidence
4. Decision Inventory and Taxonomy
4.1 Lifecycle Decision Map
Stage Decisions AI role examples Governance tier Prospect / prequal audience, suppression, prequal score, channel ranking, offer copy targeting model, eligibility screen, propensity model Medium to high Application intake completeness, document extraction, income verification, fraud/identity routing OCR, document AI, anomaly model, agent assist High Underwriting approve, decline, counteroffer, manual review, policy knockout credit score model, rules, ensemble, workflow router High Pricing handoff risk tier, APR band, fee, term, promotion eligibility pricing-risk model, grid mapping, constrained optimizer High Initial line starting line, cash line, exposure cap, requested amount handling line model, utilization forecast, capacity cap High Account management proactive CLI, requested CLI, CLD, temporary line, authorization overlay behavior model, trigger engine, line optimizer High Servicing / recourse reason explanation, reconsideration, data correction, appeal, complaint routing RAG, LLM summary, complaint classifier High if customer-impacting Portfolio governance vintage monitoring, challenger, policy review, complaint RCA anomaly detection, drift monitoring, portfolio model Medium to high
4.2 Decision Impact Classes
Class Examples Required governance Low impact assistance internal summary for analyst, non-decision QA usage policy, logging, spot QA Medium exposure/relevance prequal ranking, offer ordering, document triage feature review, copy control, monitoring High credit access approval, decline, counteroffer, manual review routing policy gate, MRM, reason, evidence, monitoring High economic term APR, fee, line, term, secured amount pricing/line policy, reason alignment, exception governance High customer recovery appeal, complaint, correction, line reinstatement case workflow, evidence bundle, SLA, CAPA High systemic risk batch CLD, model replacement, partner channel expansion governance committee, challenger, staged rollout, executive reporting
4.3 Controlled Vocabulary
Use Avoid decision_idscattered application/account ids only decision_contextgeneric “credit model result” risk_tierundefined score bucket line_action_typefree-text “account update” reason_codeLLM-generated explanation override_reason“business decision” evidence_refscreenshot-only proof complaint_root_causecustomer sentiment only customer_recovery_actionvague service note
5. Target Operating Architecture
decision inventory and risk tier
-> legal/compliance/policy interpretation
-> data and feature registry
-> protected/proxy and purpose-use controls
-> product eligibility and exposure policy
-> application intake and verification services
-> fraud / identity / credit / capacity / line models
-> underwriting orchestration
-> risk-based pricing handoff
-> line assignment / CLI / CLD service
-> reason attribution and notice/customer-copy service
-> human review, override, appeal and complaint workflow
-> evidence ledger
-> model, fairness, portfolio, complaint and customer-harm monitoring
-> challenger and change control
-> governance forums and management reporting
5.1 Architecture Capabilities
Capability What it must do Decision inventory Register all AI-influenced credit decisions, owners, risk tier, customer impact Feature registry Track source, lineage, purpose, allowed levers, protected/proxy risk, retention Policy engine Execute eligibility, product rules, exposure caps, line grids, exception constraints Verification layer Convert documents, income, identity and bureau/cashflow data into versioned decision facts Model layer Separate credit risk, fraud, capacity, line, utilization, attrition and challenger models Orchestration Combine rules, models, pricing, line, reason and human review into deterministic workflow Pricing handoff Pass only approved risk/pricing facts to pricing grid or engine Line service Manage initial line, CLI, CLD, temporary lines, exposure caps and reinstatement Reason service Generate ranked approved reasons and bind evidence refs Human review workbench Present evidence, support overrides, capture decisions and calibration Complaint/appeal case Link customer challenge to decision evidence and recovery action Evidence ledger Preserve data, rules, model, line, price, reason, copy, human action and monitoring tags Governance cockpit Report model, portfolio, line, pricing, complaint, override and fairness outcomes
6. Decision Gates
Gate 0: Use Case Boundary and Risk Tier
Question Pass condition Which customer-impacting decision can AI influence? decision family and lifecycle stage documented Does it affect access, cost, line, explanation, review or customer recovery? impact class assigned Is AI final, recommendatory, assistive, monitoring-only or copy-generating? AI role documented Which customers, products, channels and jurisdictions are in scope? scope matrix approved by owners Which exact applicability questions need Legal/Compliance? named handoff owner and artifact
Gate 1: Legal / Compliance / Policy Interpretation
Question Pass condition Which regulations, policies and internal standards may be relevant? policy mapping memo exists Does the decision create adverse action, notice, record, retention or customer communication needs? workflow decision documented by Legal/Compliance How are prequal, prescreen-like, application, CLI, CLD and counteroffer journeys distinguished? decision-context taxonomy approved Which customer-facing claims are allowed? approved copy library and prohibited language list Which records must be retained and for how long? retention requirements translated into evidence design
Gate 2: Data and Feature Boundary
Question Pass condition Are all data sources listed with lineage and owner? feature registry complete Are consumer-reporting, bureau, income, cashflow, transaction, device, channel and third-party sources tagged? source classification complete Are protected/proxy/sensitive/vulnerability signals identified for review? proxy-risk register populated Are allowed and prohibited uses defined by decision lever? purpose-use matrix approved Are prohibited uses technically blocked? feature access tests and decision-service tests pass Is missingness analyzed by channel, segment and document type? missingness and friction review complete
Gate 3: Application Intake and Verification
Question Pass condition Can document AI/OCR errors affect decline, line or pricing? verification confidence thresholds and human review triggers defined Are incomplete application and missing-document states distinct from decline? status taxonomy and customer copy approved Are alternate documentation paths available where policy allows? operational workflow and evidence fields ready Are identity/fraud holds separated from credit decline reasons? fraud-credit boundary test complete Are accessibility/language/channel friction metrics monitored? journey monitoring dashboard live
Gate 4: Underwriting Policy and Model Governance
Question Pass condition Are rules, cutoffs and models tied to policy versions? policy/model map complete Are target, label, population, reject handling and limitations documented? model card and validation package approved Can model outputs map to approved reason families? reason mapping QA passed Are manual review bands and second-look criteria explicit? review policy and queue rules approved Are champion/challenger comparisons reviewed beyond AUC/KS? performance, reason, fairness and portfolio comparison complete
Gate 5: Line Assignment and Account Management
Question Pass condition Is line decision separated from approval decision? line decision contract implemented What are min usable line, max exposure, capacity cap and portfolio cap? line grid and exposure policy approved Are CLI, requested CLI, CLD and temporary line separate workflows? line action taxonomy implemented Does CLD have customer harm, notice/servicing and appeal design? CLD impact review and case workflow ready Are multi-product exposure and partner/channel exposures aggregated? exposure service tested
Gate 6: Risk-Based Pricing Handoff
Question Pass condition Which variables can pricing use? allowed pricing feature list approved Is risk tier immutable after underwriting handoff? pricing request contract enforces it Are APR/fee/term grids versioned? pricing grid id captured in decision evidence Are promotion and experiment effects labeled separately from risk pricing? differentiation basis stored Can price/counteroffer reasons align to approved reason families where applicable? reason alignment test passed
Gate 7: Reason, Notice and Explanation
Question Pass condition Is reason catalog context-specific for application, counteroffer, CLI, CLD? catalog approved Does each reason bind to evidence refs? evidence mapping test passed Are reasons ranked as principal drivers? principal reason logic validated Is LLM prevented from inventing reasons? prompt/tool constraints and output tests pass Are channel and language explanations consistent? consistency eval complete Are notice workflow, timing and customer copy approved where applicable? Legal/Compliance sign-off recorded
Gate 8: Human Review, Override and Appeal
Question Pass condition Which cases require human review? trigger rules and queue design approved What can reviewers override? authority matrix and dual-control rules defined Are override reasons structured? override taxonomy implemented Are reviewers calibrated? training, QA and calibration plan complete Are appeal and reconsideration linked to decision evidence? case workflow tested Are override and appeal outcomes monitored by segment, channel and reviewer? dashboard ready
Gate 9: Launch, Monitoring and Kill Switch
Question Pass condition Are model, decision, line, pricing, reason, complaint and portfolio metrics live? monitoring dashboard validated Are thresholds and owners defined? alert matrix approved Can the system fall back to prior policy/model/grid? rollback and fallback test complete Are customer impact and remediation triggers defined? customer harm playbook linked Are governance forums scheduled? launch review and recurring cadence active
Gate 10: Lifecycle Review and Change Control
Question Pass condition How are model, feature, policy, line grid, pricing grid, reason catalog and copy changes approved? change workflow active How are complaints and appeals turned into CAPA? RCA/CAPA process active How are challenger results promoted or rejected? challenger governance decision recorded How are vendors monitored? change notice, evidence, SLA and audit rights tracked How are stale rules and models retired? review calendar and sunset criteria maintained
7. Required Artifacts
Artifact What it proves Decision Inventory 所有 AI-influenced credit decisions 被识别、分级、归属 Use Case Boundary Card 产品、客户、渠道、AI role、impact、out-of-scope 清楚 Legal/Compliance Applicability Memo exact applicability 由合规/法务解释, 架构已承接流程和证据 Credit Policy Map eligibility、knockout、manual review、exposure、line、exception 有版本 Feature Registry 数据来源、lineage、purpose、allowed/prohibited levers、proxy risk 可追踪 Verification Design OCR/document/income/identity errors 有 confidence gate 和人审路径 Model Cards target、data、population、validation、limitations、reason mapping 明确 Line Management Policy initial line、CLI、CLD、temporary line、reinstatement and exposure caps Pricing Handoff Contract risk tier、pricing grid、allowed variables、promotion/experiment labels Reason-Code Catalog context-specific reason, principal logic, evidence refs, templates Human Review SOP review triggers、authority、override taxonomy、QA/calibration Challenger Charter champion/challenger hypothesis, metrics, guardrails, promotion criteria Monitoring Specification metrics, thresholds, owners, alert routing, governance cadence Complaint / Appeal Playbook decision linkage, evidence retrieval, harm severity, remediation, CAPA Evidence Ledger Schema runtime evidence fields sufficient for replay and audit Vendor Evidence Pack model documentation, data use, change notice, SLA, audit and exit rights
8. RACI
Activity Product Credit Risk Compliance / Legal Fair Lending Model Risk Data / ML Engineering Ops / Underwriting Complaint Ops Internal Audit Decision inventory A C C C C C C C C I Policy and line grid C A C C C C C C I I Applicability and notice interpretation C C A C I I I C C I Feature boundary A C C C C C R I I I Model development C A I C C R C C I I Independent validation I C I C A/R C C C I I Reason catalog A C A/C C C C R C C I Human review workflow C A C C C I R A/R C I Monitoring A A C C C R R C C I Complaint root cause C C C C C C C C A/R I Audit evidence review I I I I C I I I I A/R
Legend: A accountable, R responsible, C consulted, I informed。机构内部三道防线和审批权会不同, 此表用于作品集和方案讨论, 不替代内部治理章程。
9. Control Matrix by Lifecycle Stage
Stage Primary controls Evidence Key KRIs Prequal audience review, copy approval, policy screen, proxy review audience version, copy id, prequal id prequal exposure, prequal-to-approval gap, complaints Intake document confidence, missing-info state, alternate path, accessibility OCR output, confidence, request logs missingness by segment/channel, resubmission rate Verification income/identity/bureau source validation, stale data control verification result, timestamp, source id verification fail rate, false fail appeal rate Underwriting policy rules, model validation, manual review bands rule hits, score, model version, decision contract approval/decline/counteroffer, calibration, reason drift Pricing risk tier handoff, allowed pricing features, grid version pricing request/response, grid id APR/fee distribution, exception rate, price complaints Initial line line grid, capacity cap, exposure cap, min usable line line basis, line model, exposure snapshot line distribution, activation, utilization, vintage loss CLI trigger governance, benefit/risk screen, suppression trigger id, eligibility, customer response CLI rate, post-CLI loss, CLI complaints CLD trigger evidence, harm review, communication, appeal trigger event, line action, notice/copy, case path CLD complaints, reversal, attrition, failed transactions Override authority matrix, reason taxonomy, QA reviewer action, reason, approver, evidence override rate, loss by override, segment skew Complaint/appeal decision linkage, evidence retrieval, RCA/CAPA case id, decision id, outcome, recovery upheld rate, SLA, recurring root cause Portfolio review vintage, reason, line, complaint, model and fairness monitoring dashboard snapshots, governance minutes loss, roll rates, reason drift, evidence completeness
10. Evidence Ledger Schema
Minimum fields for every high-impact credit lifecycle decision:
Field Purpose decision_id全链路 join key application_id / account_id连接核心业务记录 customer_context_idchannel, product, relationship, journey state decision_contextprequal, underwriting, initial_line, requested_CLI, proactive_CLI, CLD, pricing, appeal request_timestamp / effective_timestamp决策和生效时间 data_snapshot_id决策使用的数据版本 feature_snapshot_id特征值、missingness、excluded features policy_versioneligibility, underwriting, line, exposure, pricing policy rule_resultsdeterministic rules and knockout/pass results model_versionscredit, fraud, capacity, line, propensity, reason, LLM assistant model_outputsscores, bands, confidence, limitations flags decision_outcomeapproved, declined, counteroffer, review, incomplete, line_increase, line_decrease line_outputrequested amount, approved line, cap, line action type pricing_outputrisk tier, grid id, APR/fee/term where applicable reason_codesranked approved reasons evidence_refspolicy hits, feature refs, bureau group, income verification, human notes human_reviewreviewer, queue, action, override reason, approver customer_copy_idnotice/template/channel/language/version experiment_or_challenger_tagchampion/challenger/holdout/experiment monitoring_cohortvintage, segment, model band, channel, product complaint_or_appeal_idlinkage to recovery and root cause retention_classrecord retention and access control class
Evidence quality rules:
If a field changes a customer outcome, it needs versioning。
If a human changes a model/policy outcome, it needs structured reason and authority。
If a customer can complain about it, it needs a retrieval path for servicing and complaint ops。
If a model uses it, it needs lineage, monitoring and change control。
11. Metrics and KRIs
11.1 Executive Dashboard
Metric group Measures Access application volume, approval, decline, counteroffer, manual review, incomplete Line initial line distribution, average line by risk band, CLI, CLD, reversal Price APR/fee/term distribution, pricing exceptions, promotion/risk split Portfolio activation, spend, utilization, payment rate, roll rate, delinquency, charge-off, risk-adjusted margin Model calibration, drift, stability, challenger delta, missingness, reason distribution Customer harm complaints, appeal rate, upheld rate, wrong reason, wrong line, wrong price, time to recover Fairness/proxy segment outcomes across approval, price, line, review SLA, reason, appeal Operations review queue aging, override rate, QA defects, evidence completeness
11.2 Alert Triggers
Trigger Escalation reason distribution shifts materially after model/policy change Model Risk + Compliance + Credit Risk CLD complaints or reversals spike Account Management + Complaint Ops + Compliance prequal-to-decline gap rises Product + Credit Risk + Marketing Compliance manual review SLA disparity by channel/language Operations + Product + Fair Lending override rate spikes by reviewer/team Underwriting Ops + Credit Risk + Model Risk line distribution drops without approved policy change Credit Risk + Portfolio Risk + Product pricing exception rate rises Pricing Strategy + Compliance evidence completeness below threshold Engineering + RiskOps + Internal Control challenger improves loss but worsens complaint/appeal/reason metrics Governance committee review before promotion
12. Eval and Test Pack
Eval Test method Pass signal Policy execution replay approved scenarios through rules engine deterministic outcomes match policy truth table Feature boundary attempt to use prohibited feature in decision service request blocked and logged Reason fidelity compare decision facts to reason codes reasons are supported, ranked and context-appropriate Notice/channel consistency same decision across email/PDF/web/call center/chat customer-visible core reason is consistent LLM guardrail prompt LLM to invent or modify adverse reason unsupported reason blocked Document extraction noisy income/document samples by channel/language confidence gate routes uncertain cases to review Override calibration reviewers decide benchmark cases decisions and reasons align with policy within tolerance Line action simulation CLI/CLD scenarios across risk bands and account states correct line action, reason, copy and case path Pricing handoff pricing receives manipulated variables only approved fields accepted; grid id captured Portfolio backtest prior vintages replayed with new model/policy loss, approval, line, reason and segment impacts reviewed Complaint replay complaint case retrieves historical decision bundle evidence complete and servicing packet coherent Rollback disable new model/grid/reason catalog system falls back without evidence loss
13. Complaint, Appeal and Customer Harm Loop
13.1 Complaint Taxonomy
Complaint theme Possible root cause “I was preapproved but denied” prequal copy, eligibility screen, underwriting mismatch, changed data “Reason is wrong” reason attribution defect, template issue, LLM misuse, data correction “Income documents ignored” document AI failure, missingness workflow, reviewer queue issue “Credit line too low” line grid, capacity cap, min usable line policy, exposure aggregation “Credit line decreased without clear reason” CLD trigger, notice/copy issue, stale data, batch policy defect “Different price than expected” pricing grid, promotion eligibility, risk tier mapping, copy issue “Could not appeal or reach a person” recourse design, servicing knowledge, SLA, channel accessibility
13.2 Complaint Workflow
complaint intake
-> classify theme and harm severity
-> link decision_id / account_id / application_id
-> retrieve evidence bundle
-> determine customer recovery action
-> root cause classification
-> individual resolution
-> cohort impact assessment
-> CAPA and governance reporting
13.3 Customer Recovery Actions
Root cause Recovery action pattern data error correct data, rerun decision where policy allows, update reason, preserve audit trail document extraction error human review, accept corrected evidence, retrain/eval document pipeline wrong reason issue corrected communication where approved, fix reason mapping, audit affected cohort wrong line line review or reinstatement where approved, update line rule, monitor affected cohort pricing/grid error reprice/refund/credit process where approved, freeze grid, run cohort impact reviewer error supervisor review, calibration, QA issue, reviewer training model drift pause/rollback, challenger review, validation issue, governance decision
14. Model Risk and Change Control
14.1 Model/System Inventory Scope
Object Register because credit risk model approval, decline, counteroffer and reason fraud/identity model routing, hold, denial-like experience income/document AI converts evidence into decision facts cashflow/capacity model line, approval and affordability-like constraints line assignment model exposure and customer access pricing-risk model APR/fee/term handoff propensity/utilization model line and offer economics reason attribution logic customer explanation and notice evidence LLM/RAG assistant reviewer summary, customer explanation, complaint triage monitoring/challenger model production alert and policy change influence
14.2 Change Classes
Change class Required review feature source change Data Governance, Credit Risk, Model Risk, Fair Lending review policy cutoff change Credit Risk, Product, Compliance, Model Risk, Portfolio review model replacement validation, challenger comparison, reason/fairness/portfolio review line grid change exposure impact, customer harm, complaint forecast, governance approval pricing grid change pricing policy, risk-tier mapping, customer communication controls reason catalog change Compliance/Legal, Credit Policy, Model Risk, CX approval LLM prompt/tool change eval regression, reason-invention tests, logging review vendor model change due diligence, validation evidence, change notice, fallback plan complaint taxonomy change Complaint Ops, Compliance, Product, monitoring update
Dimension Promotion question Risk performance Does challenger improve risk separation and calibration for intended population? Approval / access Does it change approval, counteroffer or manual review mix within appetite? Line impact Does it shift line distribution, utilization and exposure concentration acceptably? Pricing handoff Does risk-tier migration create price or term changes requiring review? Reason fidelity Are reason codes stable, specific and supported by evidence? Fairness/proxy Are segment outcomes acceptable under internal review? Complaints/customer harm Is complaint, appeal or wrong-reason risk expected to remain controlled? Operations Can underwriters and servicing teams execute new evidence and reason paths? Evidence Can all decisions be replayed under the new configuration?
15. Tabletop Scenarios
Scenario 1: Prequalified Customer Gets Declined
Drill question Expected evidence What did the customer see? prequal copy id, channel, audience rule What changed at full application? application data, bureau/income verification, policy/model facts Was the reason specific and supported? reason codes, evidence refs, notice template Is this an isolated case or systemic mismatch? prequal-to-decline trend, complaints, segments What changes? copy update, prequal model/policy adjustment, monitoring threshold
Scenario 2: Batch CLD Causes Complaints
Drill question Expected evidence Which trigger caused line decrease? trigger id, model/rule version, account facts Were affected customers concentrated by segment/channel/geography? impact analysis Did customer communication explain the action and path? copy id, servicing script, appeal route Were any decisions based on stale or erroneous data? data freshness and correction review What recovery is needed? reinstatement criteria, cohort review, CAPA
Scenario 3: New Model Improves Loss but Shifts Reasons
Drill question Expected evidence Which reason families shifted? before/after reason distribution Are new reasons supported by actual decision facts? reason fidelity eval Did approval, line or price shift by segment? segment matrix Did manual review/override behavior change? override monitoring Promote, hold or revise? governance decision memo
Scenario 4: Reviewer Override Pattern Spikes
Drill question Expected evidence Which reviewer/team/channel changed? override dashboard Are overrides improving false declines or bypassing policy? vintage performance and QA sample Are certain groups more likely to receive favorable overrides? segment and channel analysis Is training or authority unclear? calibration results What control response? retraining, authority adjustment, sampling increase
Scenario 5: Customer Says Reason Is False
Drill question Expected evidence Was the reason generated from actual facts? decision snapshot and reason evidence Did data correction change the decision? correction case and rerun result where allowed Did LLM or channel copy alter the approved reason? output trace and template version Are similar cases affected? reason defect cohort query What customer recovery is appropriate? corrected communication and case outcome
16. 30 / 60 / 90 Day Implementation Roadmap
First 30 Days: Inventory and Risk Boundaries
Workstream Output Decision inventory lifecycle map covering prequal, underwriting, pricing, line, complaint Source/policy mapping official anchor to internal policy/control map Feature registry starter data source, owner, allowed/prohibited levers, proxy-risk flag Existing monitoring review gaps across approval, line, reason, complaint, override Evidence gap assessment can current systems replay a decision?
Days 31-60: Control Design and Pilot
Workstream Output Decision contract structured schema and service integration plan Reason catalog sample application decline, counteroffer, requested CLI, CLD Line governance line action taxonomy, grid review, CLD harm controls Pricing handoff risk tier and pricing grid contract Human review override taxonomy, authority matrix, QA/calibration Monitoring pilot model, decision, line, reason, complaint dashboard
Days 61-90: Production Governance
Workstream Output Gate rollout release checklist embedded into SDLC / model release Evidence ledger decision replay across selected high-impact workflows Challenger governance champion/challenger scorecard with reason/line/complaint metrics Complaint loop complaint taxonomy linked to decision evidence and CAPA Executive reporting monthly credit lifecycle AI governance pack Audit readiness evidence binder and sample case replay
17. Portfolio Deliverables
Deliverable What a hiring manager sees AI Credit Lifecycle Architecture Diagram 能把数据、政策、模型、line、pricing、reason、人审和监控串起来 Decision Inventory Spreadsheet 有产品治理颗粒度, 不只懂模型 Decision Contract JSON Sample 能把合规/风险要求转成工程接口 Reason-Code Catalog 理解 adverse action and complex algorithm reason handoff Line Management Governance Memo 理解 initial line、CLI、CLD 的客户和组合影响 Monitoring Dashboard Mock 同时看 portfolio, model, line, complaints, overrides and evidence Tabletop Pack 能带跨职能团队演练真实故障 Executive One-Pager 能对高管讲清楚价值、风险、控制和上线条件
18. Interview-Ready Language
Q1: “你如何治理 AI 信贷生命周期?”
30 秒版本 :
我会从 decision inventory 开始, 把 prequal、underwriting、pricing handoff、initial line、CLI、CLD、reason、override、complaint 和 portfolio monitoring 放到同一套操作架构里治理。核心不是让模型更聪明, 而是让每个客户影响性决策都有 policy basis、model evidence、line/pricing trace、reason code、human accountability 和 monitoring feedback。
Q2: “你如何防止 underwriting 模型变成黑箱?”
30 秒版本 :
我会把 reason attribution 设计在决策服务里, 而不是让模型或 LLM 事后解释。每个 approve/decline/counteroffer/line action 都保存 data snapshot、policy result、model version、line basis、pricing handoff、ranked reason codes 和 evidence refs。
Q3: “额度管理为什么是治理重点?”
30 秒版本 :
额度直接改变客户可用信用和机构风险暴露。初始额度、提额和降额都可能造成客户伤害、投诉、组合漂移和公平性问题, 所以我会单独设计 line action taxonomy、exposure caps、reason mapping、CLD appeal path 和 line-level monitoring。
Q4: “如何把投诉纳入 AI 信贷治理?”
30 秒版本 :
投诉不是客服噪音, 是生产监控信号。我会把 complaint theme 连接到 decision_id 和 evidence bundle, 分类为 prequal mismatch、wrong reason、wrong line、pricing issue、document friction 等 root cause, 再进入 CAPA、模型监控、文案修正、政策调整和客户恢复。
Q5: “如何处理模型风险?”
30 秒版本 :
我会按 risk-based MRM 思路登记所有影响信用决策的 assets: credit model、fraud model、document AI、line model、pricing-risk model、reason attribution、LLM assistant 和 monitoring/challenger。每个 asset 都有 intended use、validation、limitations、ongoing monitoring、change control、vendor evidence 和 effective challenge。