返回 Papers
AI 扩展计划 / Playbooks

AI Credit Lifecycle / Underwriting / Line Management Governance Playbook

核心判断:

629AI_CREDIT_LIFECYCLE_UNDERWRITING_LINE_MANAGEMENT_GOVERNANCE_PLAYBOOK.md

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

  1. 哪些 AI decisions 会改变客户的 credit access、cost、line、servicing friction 或 ability to appeal?
  2. 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?
  3. 对 prequalification、underwriting、line increase、line decrease、counteroffer、pricing handoff 是否有统一 decision id 和 evidence bundle?
  4. 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?
  5. 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?
  6. 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?
  7. 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

AnchorOfficial linkPlaybook 使用方式
CFPB Regulation B / ECOAhttps://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 algorithmshttps://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 Databasehttps://www.consumerfinance.gov/data-research/consumer-complaints/complaint taxonomy, external signal calibration, customer harm root cause, remediation and monitoring loop
Federal Reserve SR 11-7https://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 2011-12https://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 26-2https://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 2026-13https://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 RMFhttps://www.nist.gov/itl/ai-risk-management-framework用 Govern / Map / Measure / Manage 组织 AI credit lifecycle risk management
ISO/IEC 42001https://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

PrinciplePractical meaning
Decision inventory first先盘点所有 customer-impacting credit decisions, 再谈模型
Policy before modelEligibility, product rules, exposure caps and prohibited uses execute before optimization
Underwriting is not pricingrisk decision, price decision and line decision are separate but traceable
Line is a customer-impacting terminitial line, CLI and CLD require evidence, monitoring and customer communication design
Reason-ready by designreason code, evidence ref and notice/customer copy path are design inputs
Human review is governedoverride, second-look, appeal and manual exception are monitored decision paths
Complaints are signalscomplaint and appeal outcomes feed policy, model, reason, copy and remediation
Portfolio feedback is multidimensionalloss reduction is insufficient if approval, line, complaint, appeal or fairness signals degrade
Evidence is runtimeaudit evidence is generated during decision execution, not assembled from screenshots
Legal/Compliance owns applicabilityproduct and architecture teams own fact patterns, workflow, controls and evidence

4. Decision Inventory and Taxonomy

4.1 Lifecycle Decision Map

StageDecisionsAI role examplesGovernance tier
Prospect / prequalaudience, suppression, prequal score, channel ranking, offer copytargeting model, eligibility screen, propensity modelMedium to high
Application intakecompleteness, document extraction, income verification, fraud/identity routingOCR, document AI, anomaly model, agent assistHigh
Underwritingapprove, decline, counteroffer, manual review, policy knockoutcredit score model, rules, ensemble, workflow routerHigh
Pricing handoffrisk tier, APR band, fee, term, promotion eligibilitypricing-risk model, grid mapping, constrained optimizerHigh
Initial linestarting line, cash line, exposure cap, requested amount handlingline model, utilization forecast, capacity capHigh
Account managementproactive CLI, requested CLI, CLD, temporary line, authorization overlaybehavior model, trigger engine, line optimizerHigh
Servicing / recoursereason explanation, reconsideration, data correction, appeal, complaint routingRAG, LLM summary, complaint classifierHigh if customer-impacting
Portfolio governancevintage monitoring, challenger, policy review, complaint RCAanomaly detection, drift monitoring, portfolio modelMedium to high

4.2 Decision Impact Classes

ClassExamplesRequired governance
Low impact assistanceinternal summary for analyst, non-decision QAusage policy, logging, spot QA
Medium exposure/relevanceprequal ranking, offer ordering, document triagefeature review, copy control, monitoring
High credit accessapproval, decline, counteroffer, manual review routingpolicy gate, MRM, reason, evidence, monitoring
High economic termAPR, fee, line, term, secured amountpricing/line policy, reason alignment, exception governance
High customer recoveryappeal, complaint, correction, line reinstatementcase workflow, evidence bundle, SLA, CAPA
High systemic riskbatch CLD, model replacement, partner channel expansiongovernance committee, challenger, staged rollout, executive reporting

4.3 Controlled Vocabulary

UseAvoid
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

CapabilityWhat it must do
Decision inventoryRegister all AI-influenced credit decisions, owners, risk tier, customer impact
Feature registryTrack source, lineage, purpose, allowed levers, protected/proxy risk, retention
Policy engineExecute eligibility, product rules, exposure caps, line grids, exception constraints
Verification layerConvert documents, income, identity and bureau/cashflow data into versioned decision facts
Model layerSeparate credit risk, fraud, capacity, line, utilization, attrition and challenger models
OrchestrationCombine rules, models, pricing, line, reason and human review into deterministic workflow
Pricing handoffPass only approved risk/pricing facts to pricing grid or engine
Line serviceManage initial line, CLI, CLD, temporary lines, exposure caps and reinstatement
Reason serviceGenerate ranked approved reasons and bind evidence refs
Human review workbenchPresent evidence, support overrides, capture decisions and calibration
Complaint/appeal caseLink customer challenge to decision evidence and recovery action
Evidence ledgerPreserve data, rules, model, line, price, reason, copy, human action and monitoring tags
Governance cockpitReport model, portfolio, line, pricing, complaint, override and fairness outcomes

6. Decision Gates

Gate 0: Use Case Boundary and Risk Tier

QuestionPass 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
QuestionPass 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

QuestionPass 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

QuestionPass 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

QuestionPass 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

QuestionPass 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

QuestionPass 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

QuestionPass 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

QuestionPass 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

QuestionPass 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

QuestionPass 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

ArtifactWhat 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 Memoexact applicability 由合规/法务解释, 架构已承接流程和证据
Credit Policy Mapeligibility、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 DesignOCR/document/income/identity errors 有 confidence gate 和人审路径
Model Cardstarget、data、population、validation、limitations、reason mapping 明确
Line Management Policyinitial line、CLI、CLD、temporary line、reinstatement and exposure caps
Pricing Handoff Contractrisk tier、pricing grid、allowed variables、promotion/experiment labels
Reason-Code Catalogcontext-specific reason, principal logic, evidence refs, templates
Human Review SOPreview triggers、authority、override taxonomy、QA/calibration
Challenger Charterchampion/challenger hypothesis, metrics, guardrails, promotion criteria
Monitoring Specificationmetrics, thresholds, owners, alert routing, governance cadence
Complaint / Appeal Playbookdecision linkage, evidence retrieval, harm severity, remediation, CAPA
Evidence Ledger Schemaruntime evidence fields sufficient for replay and audit
Vendor Evidence Packmodel documentation, data use, change notice, SLA, audit and exit rights

8. RACI

ActivityProductCredit RiskCompliance / LegalFair LendingModel RiskData / MLEngineeringOps / UnderwritingComplaint OpsInternal Audit
Decision inventoryACCCCCCCCI
Policy and line gridCACCCCCCII
Applicability and notice interpretationCCACIIICCI
Feature boundaryACCCCCRIII
Model developmentCAICCRCCII
Independent validationICICA/RCCCII
Reason catalogACA/CCCCRCCI
Human review workflowCACCCIRA/RCI
MonitoringAACCCRRCCI
Complaint root causeCCCCCCCCA/RI
Audit evidence reviewIIIICIIIIA/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

StagePrimary controlsEvidenceKey KRIs
Prequalaudience review, copy approval, policy screen, proxy reviewaudience version, copy id, prequal idprequal exposure, prequal-to-approval gap, complaints
Intakedocument confidence, missing-info state, alternate path, accessibilityOCR output, confidence, request logsmissingness by segment/channel, resubmission rate
Verificationincome/identity/bureau source validation, stale data controlverification result, timestamp, source idverification fail rate, false fail appeal rate
Underwritingpolicy rules, model validation, manual review bandsrule hits, score, model version, decision contractapproval/decline/counteroffer, calibration, reason drift
Pricingrisk tier handoff, allowed pricing features, grid versionpricing request/response, grid idAPR/fee distribution, exception rate, price complaints
Initial lineline grid, capacity cap, exposure cap, min usable lineline basis, line model, exposure snapshotline distribution, activation, utilization, vintage loss
CLItrigger governance, benefit/risk screen, suppressiontrigger id, eligibility, customer responseCLI rate, post-CLI loss, CLI complaints
CLDtrigger evidence, harm review, communication, appealtrigger event, line action, notice/copy, case pathCLD complaints, reversal, attrition, failed transactions
Overrideauthority matrix, reason taxonomy, QAreviewer action, reason, approver, evidenceoverride rate, loss by override, segment skew
Complaint/appealdecision linkage, evidence retrieval, RCA/CAPAcase id, decision id, outcome, recoveryupheld rate, SLA, recurring root cause
Portfolio reviewvintage, reason, line, complaint, model and fairness monitoringdashboard snapshots, governance minutesloss, roll rates, reason drift, evidence completeness

10. Evidence Ledger Schema

Minimum fields for every high-impact credit lifecycle decision:

FieldPurpose
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 groupMeasures
Accessapplication volume, approval, decline, counteroffer, manual review, incomplete
Lineinitial line distribution, average line by risk band, CLI, CLD, reversal
PriceAPR/fee/term distribution, pricing exceptions, promotion/risk split
Portfolioactivation, spend, utilization, payment rate, roll rate, delinquency, charge-off, risk-adjusted margin
Modelcalibration, drift, stability, challenger delta, missingness, reason distribution
Customer harmcomplaints, appeal rate, upheld rate, wrong reason, wrong line, wrong price, time to recover
Fairness/proxysegment outcomes across approval, price, line, review SLA, reason, appeal
Operationsreview queue aging, override rate, QA defects, evidence completeness

11.2 Alert Triggers

TriggerEscalation
reason distribution shifts materially after model/policy changeModel Risk + Compliance + Credit Risk
CLD complaints or reversals spikeAccount Management + Complaint Ops + Compliance
prequal-to-decline gap risesProduct + Credit Risk + Marketing Compliance
manual review SLA disparity by channel/languageOperations + Product + Fair Lending
override rate spikes by reviewer/teamUnderwriting Ops + Credit Risk + Model Risk
line distribution drops without approved policy changeCredit Risk + Portfolio Risk + Product
pricing exception rate risesPricing Strategy + Compliance
evidence completeness below thresholdEngineering + RiskOps + Internal Control
challenger improves loss but worsens complaint/appeal/reason metricsGovernance committee review before promotion

12. Eval and Test Pack

EvalTest methodPass signal
Policy executionreplay approved scenarios through rules enginedeterministic outcomes match policy truth table
Feature boundaryattempt to use prohibited feature in decision servicerequest blocked and logged
Reason fidelitycompare decision facts to reason codesreasons are supported, ranked and context-appropriate
Notice/channel consistencysame decision across email/PDF/web/call center/chatcustomer-visible core reason is consistent
LLM guardrailprompt LLM to invent or modify adverse reasonunsupported reason blocked
Document extractionnoisy income/document samples by channel/languageconfidence gate routes uncertain cases to review
Override calibrationreviewers decide benchmark casesdecisions and reasons align with policy within tolerance
Line action simulationCLI/CLD scenarios across risk bands and account statescorrect line action, reason, copy and case path
Pricing handoffpricing receives manipulated variablesonly approved fields accepted; grid id captured
Portfolio backtestprior vintages replayed with new model/policyloss, approval, line, reason and segment impacts reviewed
Complaint replaycomplaint case retrieves historical decision bundleevidence complete and servicing packet coherent
Rollbackdisable new model/grid/reason catalogsystem falls back without evidence loss

13. Complaint, Appeal and Customer Harm Loop

13.1 Complaint Taxonomy

Complaint themePossible 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 causeRecovery action pattern
data errorcorrect data, rerun decision where policy allows, update reason, preserve audit trail
document extraction errorhuman review, accept corrected evidence, retrain/eval document pipeline
wrong reasonissue corrected communication where approved, fix reason mapping, audit affected cohort
wrong lineline review or reinstatement where approved, update line rule, monitor affected cohort
pricing/grid errorreprice/refund/credit process where approved, freeze grid, run cohort impact
reviewer errorsupervisor review, calibration, QA issue, reviewer training
model driftpause/rollback, challenger review, validation issue, governance decision

14. Model Risk and Change Control

14.1 Model/System Inventory Scope

ObjectRegister because
credit risk modelapproval, decline, counteroffer and reason
fraud/identity modelrouting, hold, denial-like experience
income/document AIconverts evidence into decision facts
cashflow/capacity modelline, approval and affordability-like constraints
line assignment modelexposure and customer access
pricing-risk modelAPR/fee/term handoff
propensity/utilization modelline and offer economics
reason attribution logiccustomer explanation and notice evidence
LLM/RAG assistantreviewer summary, customer explanation, complaint triage
monitoring/challenger modelproduction alert and policy change influence

14.2 Change Classes

Change classRequired review
feature source changeData Governance, Credit Risk, Model Risk, Fair Lending review
policy cutoff changeCredit Risk, Product, Compliance, Model Risk, Portfolio review
model replacementvalidation, challenger comparison, reason/fairness/portfolio review
line grid changeexposure impact, customer harm, complaint forecast, governance approval
pricing grid changepricing policy, risk-tier mapping, customer communication controls
reason catalog changeCompliance/Legal, Credit Policy, Model Risk, CX approval
LLM prompt/tool changeeval regression, reason-invention tests, logging review
vendor model changedue diligence, validation evidence, change notice, fallback plan
complaint taxonomy changeComplaint Ops, Compliance, Product, monitoring update

14.3 Promotion Criteria for Challenger

DimensionPromotion question
Risk performanceDoes challenger improve risk separation and calibration for intended population?
Approval / accessDoes it change approval, counteroffer or manual review mix within appetite?
Line impactDoes it shift line distribution, utilization and exposure concentration acceptably?
Pricing handoffDoes risk-tier migration create price or term changes requiring review?
Reason fidelityAre reason codes stable, specific and supported by evidence?
Fairness/proxyAre segment outcomes acceptable under internal review?
Complaints/customer harmIs complaint, appeal or wrong-reason risk expected to remain controlled?
OperationsCan underwriters and servicing teams execute new evidence and reason paths?
EvidenceCan all decisions be replayed under the new configuration?

15. Tabletop Scenarios

Scenario 1: Prequalified Customer Gets Declined

Drill questionExpected 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 questionExpected 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 questionExpected 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 questionExpected 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 questionExpected 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

WorkstreamOutput
Decision inventorylifecycle map covering prequal, underwriting, pricing, line, complaint
Source/policy mappingofficial anchor to internal policy/control map
Feature registry starterdata source, owner, allowed/prohibited levers, proxy-risk flag
Existing monitoring reviewgaps across approval, line, reason, complaint, override
Evidence gap assessmentcan current systems replay a decision?

Days 31-60: Control Design and Pilot

WorkstreamOutput
Decision contractstructured schema and service integration plan
Reason catalog sampleapplication decline, counteroffer, requested CLI, CLD
Line governanceline action taxonomy, grid review, CLD harm controls
Pricing handoffrisk tier and pricing grid contract
Human reviewoverride taxonomy, authority matrix, QA/calibration
Monitoring pilotmodel, decision, line, reason, complaint dashboard

Days 61-90: Production Governance

WorkstreamOutput
Gate rolloutrelease checklist embedded into SDLC / model release
Evidence ledgerdecision replay across selected high-impact workflows
Challenger governancechampion/challenger scorecard with reason/line/complaint metrics
Complaint loopcomplaint taxonomy linked to decision evidence and CAPA
Executive reportingmonthly credit lifecycle AI governance pack
Audit readinessevidence binder and sample case replay

17. Portfolio Deliverables

DeliverableWhat 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。