返回 Papers
AI 底层逻辑 / 经典论文

AI Account Opening / KYC:开户与准入决策架构

开户 AI 不是选择一个 IDV 组件,而是一套受监管的决策与证据架构。难点在于把准入资格、身份可信度、CIP/KYC/CDD、欺诈、AML 交接、账户激活、客户沟通和审计证据分清楚,避免所有失败都被压成一句含混的“KYC 未通过”。

434ai-foundations/papers/143-ai-account-opening-kyc-onboarding-decision-architecture.md

AI Account Opening / KYC / Onboarding Decision Architecture 解读

配对阅读:本篇的操作手册版(模板/RACI/门禁/runbook)是 docs/AI_ACCOUNT_OPENING_KYC_ONBOARDING_DECISION_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。

核心导读

开户 AI 不是选择一个 IDV 组件,而是一套受监管的决策与证据架构。难点在于把准入资格、身份可信度、CIP/KYC/CDD、欺诈、AML 交接、账户激活、客户沟通和审计证据分清楚,避免所有失败都被压成一句含混的“KYC 未通过”。

AI 能改变的决策链,是把表单、身份主张、证件与活体信号、地址和税务属性、KYB/UBO、设备与行为图谱、制裁/负面信息筛查、CDD/EDD、fraud queue、AML referral、账户限制、入金/出金激活和客户状态沟通放进同一个 state machine。每个状态都要说明它是缺资料、低置信度、政策不满足、欺诈疑点、AML 敏感路径、人工复核、受限激活还是最终拒绝;这些状态不能被一个 vendor pass/fail 或一个 LLM reason 覆盖。

系统价值来自把模型输出定义成 decision signal,而不是最终结论。OCR、liveness、document authenticity、entity match、fraud score、risk tier 和 RAG/LLM 辅助说明都必须带 source、confidence、method、policy mapping、expiry、owner、allowed use 和 evidence reference。架构上要能重放申请快照、同意和用途、第三方响应、规则版本、review note、communication template、restriction release 和 early-life monitoring,才能解释为什么某个客户被要求补料、进入复核、受限开户、放弃或被拒绝。

控制边界围绕客户伤害和金融犯罪风险同时设计。误把 IDV 通过等同于 CIP 通过会制造合规缺口,误把不确定性直接拒绝会制造 false reject 和准入伤害,把 AML/SAR-sensitive 逻辑暴露给客户会破坏保密边界。评估证据要覆盖 false reject、manual overturn、abandonment、RFI burden、accessibility、文档类型/语言/地区差异、early-life fraud/AML outcomes、投诉和供应商变更;开户 AI 的目标不是最高转化,而是可解释、可救济、可审计的客户接纳与激活控制。


0. 适用边界

本文是学习和内部架构训练材料, 不构成法律意见、BSA/AML 合规结论、KYC/CIP 判断、CDD/EDD 结论、SAR 建议、客户拒绝建议、账户关闭建议、身份供应商推荐、模型验证报告或监管解释。

规则、阈值、客户通知、资料要求、报告义务、SAR-sensitive handling、数据保留和适用性判断必须由 Legal / Compliance / BSA-AML owner / Fraud Risk / Product / Operations / Model Risk / Privacy / Information Security 按机构、产品、客户、渠道和司法辖区确认。

本文只讨论架构能力: 如何让开户旅程中的状态、决策、证据、人工复核、客户沟通、激活控制和后续监测可重放、可解释、可审计。


Source Anchors

SourceLink用途
FFIEC BSA/AML Manual - Customer Identification Programhttps://bsaaml.ffiec.gov/manual/AssessingComplianceWithBSARegulatoryRequirements/01参考 CIP written program、risk-based identity verification、account-opening risk factors、缺少可验证身份时的处理路径(访问日期: 2026-07-01)
FFIEC BSA/AML Manual - Customer Due Diligencehttps://bsaaml.ffiec.gov/manual/AssessingComplianceWithBSARegulatoryRequirements/02参考 customer risk profile、nature and purpose、ongoing monitoring 与 KYC/CDD 信息更新(访问日期: 2026-07-01)
FFIEC BSA/AML Manual - Beneficial Ownership Requirementshttps://bsaaml.ffiec.gov/manual/AssessingComplianceWithBSARegulatoryRequirements/03参考 legal entity customer beneficial owner / control person information、verification、recordkeeping 和无法形成合理信念时的处理(访问日期: 2026-07-01)
FFIEC BSA/AML Manual - Suspicious Activity Reportinghttps://bsaaml.ffiec.gov/manual/AssessingComplianceWithBSARegulatoryRequirements/04参考 suspicious activity escalation、SAR decision ownership、supporting documentation、confidentiality 和 management notification(访问日期: 2026-07-01)
FinCEN CDD Final Rule resourceshttps://www.fincen.gov/resources/statutes-and-regulations/cdd-final-rule参考 CDD 四个核心要求、beneficial ownership、customer risk profile 和 ongoing monitoring;具体适用性由 Legal / Compliance 判断(访问日期: 2026-07-01)
FinCEN BOI resourceshttps://www.fincen.gov/boihttps://www.fincen.gov/boi/Reference-materials参考 BOI reporting / access / safeguards 的官方材料;开户 KYB/UBO 架构不应把 FinCEN BOI registry、CDD rule 和本机构客户资料混成同一事实源(访问日期: 2026-07-01;注意 2025-03 IFR 已豁免美国境内实体的 BOI 申报,见文末 SOTA 检查)
FinCEN BSA Filing Information / SAR resourceshttps://www.fincen.gov/resources/filing-information参考 SAR filing operations、SAR form resources 和 BSA E-Filing handoff;AI 不拥有 SAR filing decision(访问日期: 2026-07-01)
NIST SP 800-63-4 Digital Identity Guidelineshttps://pages.nist.gov/800-63-4/参考 identity proofing、authentication、federation、fraud controls、forged media、customer experience 和 privacy(正式版 2025-07)
NIST SP 800-63A-4 Identity Proofing and Enrollmenthttps://pages.nist.gov/800-63-4/sp800-63a.html用 identity resolution、evidence validation、attribute validation、identity verification、enrollment 和 exception handling 组织 proofing gate(正式版 2025-07,随 SP 800-63-4 同批发布)
NIST AI RMFhttps://www.nist.gov/itl/ai-risk-management-framework用 Govern / Map / Measure / Manage 组织 AI decisioning、model/vendor risk、monitoring、human oversight 和 incident learning(访问日期: 2026-07-01)
ISO/IEC 42001https://www.iso.org/standard/42001用 AI management system 视角组织 roles、operation、performance evaluation、internal audit、management review 和 continual improvement(访问日期: 2026-07-01)

访问日期按 2026-06-30 记录。


1. 核心问题: 开户不是 KYC 工具串联

很多数字开户项目把问题定义为:

form + document OCR + liveness + sanctions screen + fraud score -> open / reject

这个定义会把不同责任线压成一个模糊结论。开户决策至少包含:

Product eligibility
  + Identity assurance
  + CIP / KYC / CDD evidence
  + Financial crime risk
  + Fraud and synthetic identity control
  + Funding and activation control
  + Customer communication and recourse
  + Evidence and monitoring governance

低成熟度系统把所有异常都显示为 KYC failed。这会制造四类问题:

  • owner 混乱: product eligibility、CIP、fraud、AML、funding hold 和 customer abandonment 被同一个状态覆盖。
  • 客户伤害: 技术失败、资料缺失、欺诈怀疑和 AML review 使用同一文案, 客户不知道下一步。
  • 控制失效: AML/SAR-sensitive 信息可能被暴露给普通客服或客户。
  • 审计困难: 事后无法证明当时哪些证据、规则、模型、人工判断导致 hold、decline、RFI、restricted activation 或 account closure。

成熟问题定义是:

Can we prove which gate, owner, evidence, model, policy, human action and communication
produced each onboarding state transition,
while preserving customer-safe explanations and restricted AML/fraud confidentiality?

2. 方法和架构贡献

本文的架构贡献是把开户设计成 stateful decision-and-evidence fabric。

digital channel / branch assisted / contact center
  -> consent and disclosure service
  -> application journey orchestrator
  -> document and evidence capture
  -> identity proofing service
  -> CIP verification service
  -> CDD / KYB profile service
  -> fraud and synthetic identity graph
  -> sanctions / AML referral interface
  -> policy and decision orchestration
  -> exception queue workbench
  -> core account opening
  -> funding and activation controls
  -> customer communication service
  -> evidence ledger and audit binder
  -> monitoring, QA, model risk and management reporting

关键设计是 decision orchestration 不直接吞掉所有模型输出。它应接收结构化信号:

decision_signal:
  source: identity_proofing_vendor
  signal_type: evidence_validation
  result: inconclusive
  confidence: calibrated_bucket_medium
  reason_codes: [DOCUMENT_GLARE, ADDRESS_MISMATCH]
  customer_safe_reason: additional_information_needed
  evidence_refs: [doc_id, session_id, vendor_response_id]
  owner: identity_platform
  expires_at: 2026-07-15

架构原则:

PrinciplePractical meaning
State first先定义 application、review、hold、restricted、active、declined、abandoned 等状态, 再接入模型
Gate ownership每个 gate 有真实 owner, 不是所有异常都归 KYC
Evidence before conclusion证据、规则、模型和人工判断分开保存
Customer-safe communication客户文案解释下一步, 不泄露 fraud/AML/SAR-sensitive logic
Activation-aware开户批准不等于资金、卡、转账、限额全部激活
Recourse and friction monitoringabandonment、RFI burden、manual overturn 和 complaint 是治理信号

3. 机制原理: End-to-End State Machine

开户不是一次性 yes/no, 而是一组 state transitions。

lead / pre-application
  -> consent and purpose capture
  -> application intake
  -> eligibility screen
  -> identity proofing
  -> CIP information collection and verification
  -> CDD risk profile
  -> legal entity / UBO / authority review if applicable
  -> fraud and synthetic identity assessment
  -> sanctions / AML screening and referral triggers
  -> decision orchestration
  -> account creation or hold / decline / RFI
  -> funding and activation
  -> early-life monitoring
  -> ongoing CDD update triggers

3.1 状态必须区分

State含义不应混用
application_started客户开始申请, 尚未形成开户关系不等于 customer record completed
consent_captured已取得特定用途同意/披露确认不等于可任意复用身份数据
identity_proofing_pending身份证明流程未完成不等于 AML hold
cip_verification_pending关键识别信息待验证不等于欺诈结论
cdd_review_pending需补足 nature/purpose、expected activity、risk profile不等于产品资格失败
fraud_review_pending欺诈信号需人工复核不等于 SAR decision
aml_review_pendingAML/CDD/SAR-sensitive review不应把原因暴露给普通客服或客户
approved_restricted可创建账户但有入金、出金、转账或渠道限制不等于 fully active
opened_not_funded账户已建但未入金/未激活不等于 onboarding complete
active账户可按政策使用仍需 early-life monitoring
declined申请被拒或未开户reason catalog 和通知 owner 必须明确
abandoned客户中途退出或超时可能是 hidden harm / friction signal

状态设计的核心是避免系统把“未完成、待审核、拒绝、限制、关闭、可疑活动”混成一个 UI 文案。

3.2 Outcome Table

OutcomeWhen usedGuardrail
Approve and activate低风险、必要信息与证据足够、资金路径清晰early-life monitoring 仍要保留
Approve restricted身份/KYC 可接受但资金或早期行为风险需限制限制、解除条件、客户文案和监控必须版本化
Request information缺少可补证据或信息冲突可解释只索取必要材料, 避免无限 RFI loop
Hold for review需要人工判断、增强审核或跨团队确认必须有 queue owner、SLA、客户状态文案
Decline / do not open不符合资格、无法验证、风险超出政策或政策禁止reason code 和客户沟通由对应 owner 批准
Exit / close after opening已开账户但后续无法完成必要验证或出现政策触发账户关闭、资金退回、SAR-sensitive 处理需权责方确认
Abandon / timeout客户未完成流程或未响应计入 friction/harm monitoring, 不自动视为低风险成功

4. Decision Domains 和 Gate Map

每个 gate 要明确 decision question、AI role、owner、evidence 和客户沟通边界。

GateDecision questionAI roleHuman / policy ownerEvidence
Consent and purpose是否可采集、处理、共享、保存这些数据classify data purpose, detect missing consentPrivacy / Legal / Productconsent version, purpose tag, channel proof
Product eligibility客户是否符合产品、地区、年龄、账户类型、业务类型等规则prefill, recommend alternate pathProduct / Legal / Complianceeligibility rule id, application snapshot
Identity proofing申请人是否与真实自然人建立足够关联document/liveness/media/device risk scoringIdentity platform + KYC Opsidentity claim, evidence validation, verification result
CIP verification是否可形成对客户真实身份的合理信念non-documentary match, discrepancy triageBSA/AML ComplianceCIP fields, verification method, discrepancy resolution
CDD risk profile是否理解关系性质和目的、可建立风险档案risk-tier recommendation, expected activity anomalyBSA/AML ownercustomer profile, expected activity, risk factors
KYB / UBO小企业/法人客户是否有实体、授权人、控制人、实益拥有人证据entity extraction, ownership graph consistencyKYB / BSA Opsentity docs, ownership attestation, BO verification
Fraud是否存在 synthetic identity、device farm、stolen identity、mule riskfraud score, link analysis, velocityFraud Risksignal bundle, graph links, reviewer decision
AML handoff是否需要 AML/CDD/EDD/SAR-sensitive reviewred-flag detection, evidence summaryBSA/AML Compliancereferral reason, supporting documentation
Funding activation是否允许入金、解冻、发卡、转账、提高限额funding risk score, return riskDeposit Ops / Fraud / Paymentsfunding source, account status, restriction policy
Communication客户看到什么原因、下一步、申诉/补件路径approved template fill, translationLegal / Compliance / CX / Productnotice id, reason code, delivery proof

责任线要分清:

问题真实 owner典型系统动作客户沟通边界
产品资格不符合Product / Eligibility policydecline / offer alternate product可解释产品规则, 不暴露内部风控阈值
无法形成身份合理信念KYC/CIP owner + Operationsrequest evidence / hold / decline / close per policy说明需要验证信息, 不推断欺诈
身份证明流程无法完成Identity platform + CX + Accessibilityretry / alternate method / assisted path避免把技术失败表述为客户风险
欺诈风险高Fraud Riskmanual review / restricted activation / decline per policy客户安全语言, 避免泄露规则
AML/CDD 风险需增强审核BSA/AML ComplianceCDD/EDD queue / relationship manager review / SAR-sensitive handlingSAR 相关信息不能向客户泄露
UBO/授权资料矛盾KYB / BSA OperationsRFI / authority review / entity verification明确缺少或不一致信息, 不替客户下结论
首次入金风险高Deposit Operations / Fraud / Paymentshold funding / limit account / delayed activation解释资金可用性或验证步骤, 由政策批准
客户放弃或被困住Product + CX Riskrescue path / assisted onboarding / abandon recovery不把摩擦率下降误认为风险下降

5. Evidence Bundle

开户 evidence bundle 要回答三个问题:

  1. 当时系统知道什么?
  2. 哪个规则、模型、人员或队列作了什么动作?
  3. 客户、审计、合规、模型风险和运营复盘分别能看到什么?
Evidence object内容设计要求
Application snapshot表单、渠道、设备、时间、产品、客户输入immutable version, PII access control
Consent record披露、同意、用途、第三方共享、biometric/IDV notice where applicablepurpose-bound, versioned, retrievable
Identity claim申请人核心属性、证据类型、proofing path不把 proofing pass 当成 CIP pass
Evidence validation result证件/文件/属性验证结果source, method, vendor version, discrepancy
Identity verification resultface match、liveness、manual check、trusted refereeaccessibility fallback and retry history
CIP verification recorddocumentary / non-documentary method and resultBSA/AML policy mapping
CDD profilerelationship purpose、expected activity、occupation/business、risk tierupdated on trigger, not only at onboarding
UBO / authority packageentity docs、ownership/control person、attestation、verificationKYB queue and beneficial ownership owner
Fraud signal bundledevice, network, velocity, synthetic, mule, deepfake indicatorscustomer-safe separation from internal rules
AML referral recordred flags, referral reason, supporting docs, analyst decisionSAR confidentiality and restricted access
Decision logoutcome, reason code, policy/model versions, human overridedeterministic replay as far as practical
Communication proofmessage template, channel, delivery, customer action pathwording approved and reason-owned
Activation recordaccount created, restrictions, funding status, holds, releaseavoids active account with unresolved critical gate

Evidence bundle 必须按权限分层。普通客服不应看到 SAR-sensitive rationale; 客户不应看到 fraud rule; 审计和合规需要能抽样重放; 模型风险需要版本、eval、monitoring 和 issue evidence。


6. Small Business / UBO / KYB Flow

小企业开户的难点不是多收几份文件, 而是把三件事分清:

entity exists
  + applicant has authority
  + natural persons behind control / ownership are identified and risk-assessed where applicable
Decision pointEvidenceCommon AI supportGuardrail
Entity identityformation record, EIN/TIN, business address, registration statusdocument extraction, registry match, name normalizationregistry mismatch enters KYB review; AI summary not final entity verification
Authorized representativeofficer title, resolution, signer authority, role evidenceauthority document extraction, signature package triageauthority ambiguity cannot be hidden behind “business verified”
Beneficial owner / control personownership attestation, control prong info, ID evidenceownership graph consistency, duplicate person detectionCDD/BO applicability and thresholds are Compliance-owned
Business nature and purposeNAICS/MCC, website, expected activity, cash intensitybusiness classification, website risk summarydo not over-rely on web presence for legitimacy
Ownership discrepancyconflicting names, addresses, percentages, hidden nomineesanomaly detection, graph linksescalates to KYB/BSA queue; customer wording stays evidence-based

FinCEN BOI resources should be treated as official source anchors, not as a substitute for institution-owned CDD/KYB procedures. The architecture should preserve source provenance:

customer attestation != state registry != FinCEN BOI data != third-party KYB vendor != internal CDD profile

KYB/UBO 的核心是 provenance 和 authority。多个来源不一致时, 系统不能简单用“最新值覆盖旧值”。应保留 source、time、confidence、owner、resolution action 和 customer-safe RFI。


7. Customer Harm、Abandonment 和 Activation

开户 AI 的客户伤害常隐藏在“未完成”里:

Harm patternSignalControl
False rejecthigh decline / upheld appeal / manual overturnsegment review, reason QA, reviewer calibration
Endless evidence looprepeated upload, same rejection reason, no alternate pathmax retry policy, assisted path, evidence reuse
Accessibility failureliveness or document capture fails for legitimate usersalternate proofing, trusted referee, branch/contact center path
Privacy overcollectionasking for documents beyond decision needfield-level necessity review, purpose tags
Confusing holdcustomer sees generic “under review” for daysSLA-based status, safe explanation, escalation path
Abandoned onboardingdrop-off after IDV, CDD, UBO, fundingrescue analytics, friction-vs-risk dashboard
Wrong denial communicationproduct decline, fraud decline and KYC failure share one messagereason taxonomy and message owner
Premature activationaccount opened before critical gate or funding risk resolvedrestricted activation and early-life controls

North-star metric 不应只是 conversion rate。更稳的指标组合:

eligible activation rate
  + verified identity completion
  + false reject / overturn rate
  + RFI completion burden
  + review SLA
  + segment friction disparity
  + early-life fraud / AML quality
  + complaint and recourse outcome

Activation 是独立控制面。可以存在 approved_restrictedopened_not_fundedfunding_holdoutbound_transfer_blockedcard_not_issuedearly_life_monitoring_active 等状态。这样既能避免过度拒绝, 也能避免高风险账户直接全功能激活。


8. Model Risk and AI Governance

开户链路里的 AI 不止一个模型:

AI / model assetRiskRequired governance
ID document classifierwrong evidence type or missed tampereval by document class, quality, region, channel
Liveness / PAD / media integrityfalse pass, false fail, bias, injection bypassadversarial eval, segment QA, vendor monitoring
Attribute matchingname/address/DOB false mismatchmatch policy, thresholds, manual review
Synthetic identity graphunfair or opaque network inferencefeature lineage, reviewer evidence, calibration
Fraud scorefalse decline or mule account passoutcome monitoring, override QA, early-life loss linkage
CDD risk tieringunder-risking high-risk profiles or overburdening low-risk customerstypology coverage, scenario review, Compliance owner
LLM evidence assistanthallucinated summary or unsafe customer wordinggrounded summary only, citation requirement, output policy
Routing modelqueues overloaded or wrong ownerqueue telemetry, SLA, reviewer feedback

Use NIST AI RMF for lifecycle risk management and ISO/IEC 42001 for operating system discipline:

AI inventory -> intended use -> risk tier -> eval plan -> release gate
  -> runtime monitoring -> incident / harm review -> management review
  -> control improvement

治理证据应包含:

  • use case inventory and risk tier。
  • approved use / prohibited use。
  • data source, consent/purpose tag, retention, PII and biometric handling where applicable。
  • vendor model version, change notice, SLA, audit/evidence rights。
  • eval suite by document type, region, channel, accessibility, spoof/media attack, customer segment。
  • human review, override reason, queue SLA and reviewer calibration。
  • false reject, manual overturn, early-life fraud/AML outcomes, complaints and abandonment。

9. 为什么有效: 分离责任线, 保留证据线

这套架构有效, 因为它把开户从“通过/失败”改成多 gate、多 owner、多 evidence 的 state machine。

第一, state machine 防止流程语义坍缩。identity_proofing_pendingfraud_review_pendingaml_review_pendingapproved_restrictedabandoned 代表不同风险、不同 owner 和不同客户沟通方式。

第二, decision signal contract 让模型输出进入治理边界。OCR、liveness、fraud graph、CDD risk tier、routing model 和 LLM assistant 都只提供带来源、置信度、reason、evidence、owner 和 expiry 的信号, 不直接产生最终状态。

第三, evidence bundle 让事后复盘可重放。开户争议通常跨 channel、vendor、core、case management、AML queue、funding system 和 customer communication。没有统一证据包, 审计和投诉只能靠截图拼接。

第四, customer-safe communication 降低伤害。客户需要知道下一步、需要什么材料、预计多久、如何获得帮助; 但 fraud/AML/SAR-sensitive 逻辑必须按权限隔离。

第五, activation-aware design 避免过度二元化。系统可以在证据足够时开户但限制功能, 或在 funding risk 未清时延迟激活, 从而同时管理客户体验和风险。


10. 局限和误用

Anti-patternWhy it failsBetter architecture
KYC failed as universal reasonowner、evidence 和 communication collapsecontrolled reason taxonomy by decision domain
IDV vendor pass equals CIP passproofing signal is not full compliance decisionCIP decision service with policy mapping
Decline before queue triagefalse rejects and poor customer recoursehold/RFI/review state with SLA and evidence
AML, fraud and product decline share messageconfidentiality, fairness and customer harm riskscustomer-safe message catalog by owner
Funding before critical gates clearmule and synthetic accounts can activaterestricted opening and activation control
LLM writes denial reasoninvented reasons and inconsistent noticesstructured reason code source of truth
BO data overwritten by latest sourcedestroys provenance and auditabilitysource-ranked evidence graph
Conversion optimized alonehides abandoned harmed customersbalanced risk, harm and activation metrics
Vendor evidence not retainedcannot replay document/liveness/fraud outcomesvendor response id, model version, audit rights

局限还包括 identity proofing 的公平体验问题。摄像头质量、姓名结构、地址历史、非传统身份文件、残障辅助需求、语言和渠道能力都会影响完成率。架构上需要 alternate proofing、assisted onboarding、trusted referee 或 branch/contact center fallback, 但具体政策由机构 owner 批准。


11. 架构和产品价值

Value areaArchitecture output
Regulated journey designstate machine、gate map、owner、SLA、customer-safe status
Evidence governanceapplication snapshot、consent、identity claim、CIP/CDD/KYB、fraud/AML、decision、communication、activation
Customer experienceRFI burden control、alternate path、assisted onboarding、clear next step
Financial crime controlfraud graph、AML referral、SAR-sensitive access boundary、early-life monitoring
Product growth qualityeligible activation rate, not raw conversion; friction and false reject monitoring
Model/vendor governanceinventory、eval、vendor change notice、override QA、segment performance
Audit readinessdeterministic replay where practical, source provenance, policy/model/template versions

对金融零售软件经验者而言, 这篇的核心价值是把开户旅程从“前端转化漏斗”升级为“受监管的客户接纳和账户激活系统”。产品目标、风险控制、客户负担和证据保留必须同时成立。


12. 金融零售系统案例: Digital Checking + Small Business Checking

场景: 一家银行计划把个人 digital checking 和 small business checking 合并到同一 onboarding platform。产品团队希望提高转化, 风险团队担心 synthetic identity、mule accounts、KYB/UBO 不完整和 early-life fraud。

目标架构:

DomainDesign choiceEvidence
Journey个人和小企业共用 orchestrator, 但 gate map 不同journey version、product type、state history
Consent每个数据源和第三方调用绑定 purpose tagconsent version、purpose、channel proof
IdentityIDV vendor 输出 proofing signal, 不直接等同 CIP conclusionvendor response id、method、confidence、discrepancy
CIP/CDDCIP verification service 和 CDD profile service 独立版本化policy version、verification result、risk profile
Frauddevice、velocity、synthetic graph 进入 fraud review queuegraph evidence、review note、restriction action
KYB/UBOentity、authority、beneficial owner、business purpose 分开 reviewentity docs、ownership package、source provenance
AML handoffred flags 进入 restricted queue, 与普通客服权限隔离referral record、restricted evidence、analyst decision
Activation允许 approve restricted: 入金、出金、卡、ACH、wire、限额分层释放activation policy、restriction reason、release evidence
Communication客户看到材料需求、状态和下一步, 不看到敏感规则template id、delivery proof、language
Monitoring同时看 conversion、eligible activation、RFI burden、manual overturn、early-life fraud、complaintsmetric contract、threshold、owner

典型事件: 小企业客户完成表单和证件上传, entity registry match 通过, 但 UBO ownership percentages 与 attestation 不一致, 同时申请人设备与多个近期申请有关联。低成熟度系统可能显示“KYC failed”并拒绝。成熟系统会把 entity identity、authorized representative、UBO discrepancy 和 fraud signal 分开记录; 对 UBO discrepancy 发送 customer-safe RFI; 对 fraud signal 进入 restricted review queue; 在资金风险未清时使用 approved_restricted 状态; 记录 evidence refs、reviewer actions、communication template 和 activation release criteria; 并监控该类 discrepancy 是否造成高放弃率或人工 overturn。这样系统既避免“风险未清直接全功能开户”, 也避免“所有不确定性直接拒绝”。


13. 学习验证

完成本文后, 应能解释开户 AI 为什么是 stateful decision-and-evidence fabric, 并区分 started、consent、identity pending、CIP pending、CDD review、fraud review、AML review、restricted、active、declined、abandoned。还应能拆清 eligibility、identity proofing、CIP、CDD、KYB/UBO、fraud、AML、funding、communication 的 owner, 定义 decision signal 和 evidence bundle, 说明 customer-safe communication、activation control、harm monitoring 与 model/vendor governance。

最小可交付架构包应包括: account opening state machine、CIP/CDD/KYB/UBO gate map、decision signal contract、evidence bundle schema、customer-safe reason/status catalog、exception queue SLA、activation restriction matrix、AI model/vendor eval plan、friction/harm/early-life risk dashboard 和 audit evidence replay checklist。真正掌握这篇的标志是: 能把开户体验、KYC/AML、身份、欺诈、资金激活和客户沟通放进同一个可治理状态机, 同时不牺牲证据、权限边界和客户可行动性。


SOTA 检查 (2026-07-01)

  • 身份证明基线已从草案换成正式版:NIST SP 800-63-4 / 63A-4 于 2025-07 发布最终版(结束近四年修订、两轮公开草案、近 6000 条公开意见),本篇 Source Anchors 引用的即该正式版页面;正文第 3-4 节按 identity resolution → evidence validation → identity verification 组织 proofing gate 的方式与 63A-4 最终结构一致,主线仍成立。
  • FinCEN BOI 事实源发生实质变化 (2025-03):FinCEN 于 2025-03-26 发布 interim final rule,豁免所有美国境内设立实体(原 domestic reporting companies)的 BOI 申报义务,"reporting company" 定义收窄为在美注册经营的外国实体。这强化了本篇第 6 节的核心警告——FinCEN BOI registry 不能替代机构自有的 CDD/KYB beneficial ownership 程序(银行侧 CDD rule 义务独立于 BOI registry 覆盖范围),customer attestation != registry != FinCEN BOI != internal CDD profile 的 provenance 分层在 registry 覆盖收窄后更加必要。
  • 威胁面 2025-2026 主线:deepfake 防御趋于成熟、synthetic identity 成为主要缺口:行业报告(ID-Pal、Sumsub、Biometric Update,2025-12 至 2026 上半年)称 AI 生成身份欺诈同比增长约 700%,出现全自动化 KYC 绕过 bot(从生成假证到通过 liveness 再到与客服对话全链路代理);同时多数 KYC 栈对 deepfake 已有分层防御,而 synthetic identity 尝试仍同比 +31%。这印证本篇把 synthetic identity graph 设为独立 gate、以及从单点检查转向 KYC+KYB+持续监测的分层 risk-based workflow 的判断。
  • Agentic AI 进入 AML/KYC 生产线,但 HITL 边界与本篇结论一致:2025-09 的 Co-Investigator AI(arXiv 2509.08380)与 2026 年行业实践(Hawk、Bretton AI 等)把 SAR 叙述草拟、KYC refresh、证据汇总做成 agent 工作流,但 SAR 提交、EDD 判定、AML 账户关闭等高风险决策一律要求 human-in-the-loop、由 BSA officer 签署——本篇「AI 不拥有 SAR filing decision」「LLM 只做 grounded summary 不写 denial reason」的控制边界不仅未过时,反而成为 2026 年主流治理共识("the model said so" 不是有效监管抗辩)。库内配套实操见 docs/aipa/day1-aml-copilot-product-discovery-jtbd.mddocs/aipa/day2-aml-ai-competitive-landscape.md(2026-06,AIPA 计划)。
  • 不随版本过时的框架性结论:state machine 状态分层(proofing/CIP/CDD/fraud/AML/restricted/abandoned 不混用)、decision signal contract(source/confidence/owner/expiry/evidence_refs)、evidence bundle 可重放、customer-safe communication 与 SAR-confidentiality 权限隔离、activation 独立控制面——这些是责任线与证据线的组织方式,不绑定任何 IDV 供应商、模型版本或具体阈值;会过时的是具体 vendor 能力、liveness 攻防手段和监管申报口径,复查时只需刷新这三类。