AI 底层逻辑 / 经典论文
AI Account Opening / KYC：开户与准入决策架构
访问日期按 2026-06-30 记录。本文不输出法律或合规结论；规则、阈值、客户通知和报告义务的最终适用性由 Legal / Compliance / BSA-AML owner 按机构、产品、客户、渠道和司法辖区确认。
321 行ai-foundations/papers/143-ai-account-opening-kyc-onboarding-decision-architecture.md
AI Account Opening / KYC / Onboarding Decision Architecture 解读
面向对象: AI PM / Product Architect / Senior BA / AML-KYC Product Lead / Identity Platform Architect / Digital Banking Transformation Lead。 核心问题: 数字开户不是把证件 OCR、活体检测、制裁筛查、欺诈模型和开户表单串起来。成熟系统要把 eligibility、CIP/KYC/CDD、identity proofing、fraud、AML handoff、例外队列、客户沟通、资金激活和审计证据设计成一个可解释、可复核、可治理的 decision architecture。 学习目标: 设计 account opening state machine、decision gate、risk-tiering、evidence package、small business UBO/KYB flow、fraud/AML handoff、hold/decline ownership、model risk controls 和面试表达。
Source Anchors
| Source | Link | 用途 |
|---|---|---|
| FFIEC BSA/AML Manual - Customer Identification Program | https://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 |
| FFIEC BSA/AML Manual - Customer Due Diligence | https://bsaaml.ffiec.gov/manual/AssessingComplianceWithBSARegulatoryRequirements/02 | 参考 customer risk profile、nature and purpose、ongoing monitoring 与 KYC/CDD 信息更新 |
| FFIEC BSA/AML Manual - Beneficial Ownership Requirements | https://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 |
| FFIEC BSA/AML Manual - Suspicious Activity Reporting | https://bsaaml.ffiec.gov/manual/AssessingComplianceWithBSARegulatoryRequirements/04 | 参考 suspicious activity escalation、SAR decision ownership、supporting documentation、confidentiality 和 management notification |
| FinCEN CDD Final Rule resources | https://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 |
| FinCEN BOI resources | https://www.fincen.gov/boi 和 https://www.fincen.gov/boi/Reference-materials | 参考 BOI reporting / access / safeguards 的当前官方材料；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 |
| FinCEN BSA Filing Information / SAR resources | https://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 |
| NIST SP 800-63-4 Digital Identity Guidelines | https://pages.nist.gov/800-63-4/ | 参考 identity proofing、authentication、federation、fraud controls、forged media、customer experience 和 privacy |
| NIST SP 800-63A-4 Identity Proofing and Enrollment | https://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 |
| NIST AI RMF | https://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 |
| ISO/IEC 42001 | https://www.iso.org/standard/81230.html | 用 AI management system 视角组织 roles、operation、performance evaluation、internal audit、management review 和 continual improvement |
访问日期按 2026-06-30 记录。本文不输出法律或合规结论；规则、阈值、客户通知和报告义务的最终适用性由 Legal / Compliance / BSA-AML owner 按机构、产品、客户、渠道和司法辖区确认。
一句话:
Account opening AI is a governed decision-and-evidence fabric, not an IDV widget selection problem.
1. 核心心智模型
开户决策至少包含四条不同责任线:
Product eligibility
+ Identity assurance
+ Financial crime risk
+ Customer activation
+ Evidence and communication governance
很多项目失败，是因为把所有异常都压成一个 KYC failed。高级架构要把不同问题分开:
| 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 policy | decline / offer alternate product | 可解释产品规则，但不暴露内部风控阈值 |
| 无法形成身份合理信念 | KYC/CIP owner + Operations | request 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 + Accessibility | retry / alternate method / trusted referee / branch path | 避免把技术失败表述为客户风险 |
| 欺诈风险高 | Fraud Risk | manual 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 Compliance | CDD/EDD queue / relationship manager review / SAR-sensitive handling | SAR 相关信息不能向客户泄露 |
| 资料或实益拥有人信息矛盾 | KYB / BSA Operations | RFI / 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 / Payments | hold 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 Risk | rescue path / assisted onboarding / abandon recovery | 不把摩擦率下降误认为风险下降 |
开户不是一次性 yes/no，而是一组 state transitions。AI 可以排序、推荐、摘要和路由，但不能模糊 owner。
2. End-to-End State Machine
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
2.1 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_pending | AML/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 | 原因 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 |
状态设计的核心不是漂亮流程图，而是避免系统把“未完成、待审核、拒绝、限制、关闭、可疑活动”混成一个 UI 文案。
3. Decision Domains
3.1 Gate Map
| Gate | Decision question | AI role | Human / policy owner | Evidence |
|---|---|---|---|---|
| Consent and purpose | 是否可采集、处理、共享、保存这些数据 | classify data purpose, detect missing consent | Privacy / Legal / Product | consent 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 path | Product / Legal / Compliance | eligibility 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 scoring | Identity platform + KYC Ops | identity 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 triage | BSA/AML Compliance | CIP 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 anomaly | BSA/AML owner | customer 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 consistency | KYB / BSA Ops | entity docs, ownership attestation, BO verification |
| Fraud | 是否存在 synthetic identity、device farm、stolen identity、mule risk | fraud score, link analysis, velocity | Fraud Risk | signal bundle, graph links, reviewer decision |
| AML handoff | 是否需要 AML/CDD/EDD/SAR-sensitive review | red-flag detection, evidence summary | BSA/AML Compliance | referral 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 risk | Deposit Ops / Fraud / Payments | funding 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, translation | Legal / Compliance / CX / Product | notice id, reason code, delivery proof |
3.2 Outcome Table
| Outcome | When used | Guardrail |
|---|---|---|
| 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 Legal/Compliance/Ops 确认 |
| Abandon / timeout | 客户未完成流程或未响应 | 计入 friction/harm monitoring，不应自动视为低风险成功 |
4. Reference Architecture
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
5. Evidence Bundle
开户 evidence bundle 要能回答三个问题:
- 当时系统知道什么?
- 哪个规则、模型、人员或队列作了什么动作?
- 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 applicable | purpose-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 result | face match、liveness、manual check、trusted referee | accessibility fallback and retry history |
| CIP verification record | documentary / non-documentary method and result | BSA/AML policy mapping |
| CDD profile | relationship purpose、expected activity、occupation/business、risk tier | updated on trigger, not only at onboarding |
| UBO / authority package | entity docs、ownership/control person、attestation、verification | KYB queue and beneficial ownership owner |
| Fraud signal bundle | device, network, velocity, synthetic, mule, deepfake indicators | customer-safe separation from internal rules |
| AML referral record | red flags, referral reason, supporting docs, analyst decision | SAR confidentiality and restricted access |
| Decision log | outcome, reason code, policy/model versions, human override | deterministic replay as far as practical |
| Communication proof | message template, channel, delivery, customer action path | wording approved and reason-owned |
| Activation record | account created, restrictions, funding status, holds, release | avoids active account with unresolved critical gate |
6. Small Business / UBO 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
6.1 KYB / UBO Decision Table
| Decision point | Evidence | Common AI support | Guardrail |
|---|---|---|---|
| Entity identity | formation record, EIN/TIN, business address, registration status | document extraction, registry match, name normalization | registry mismatch enters KYB review; AI summary not final entity verification |
| Authorized representative | officer title, resolution, signer authority, role evidence | authority document extraction, signature package triage | authority ambiguity cannot be hidden behind “business verified” |
| Beneficial owner / control person | ownership attestation, control prong info, ID evidence | ownership graph consistency, duplicate person detection | CDD/BO applicability and thresholds are Compliance-owned |
| Business nature and purpose | NAICS/MCC, website, expected activity, cash intensity | business classification, website risk summary | do not over-rely on web presence for legitimacy |
| Ownership discrepancy | conflicting names, addresses, percentages, hidden nominees | anomaly detection, graph links | escalates 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
7. Customer Harm and Abandonment
开户 AI 的客户伤害常隐藏在“未完成”里:
| Harm pattern | Signal | Control |
|---|---|---|
| False reject | high decline / upheld appeal / manual overturn | segment review, reason QA, reviewer calibration |
| Endless evidence loop | repeated upload, same rejection reason, no alternate path | max retry policy, assisted path, evidence reuse |
| Accessibility failure | liveness or document capture fails for legitimate users | alternate proofing, trusted referee, branch/contact center path |
| Privacy overcollection | asking for documents beyond decision need | field-level necessity review, purpose tags |
| Confusing hold | customer sees generic “under review” for days | SLA-based status, safe explanation, escalation path |
| Abandoned onboarding | drop-off after IDV, CDD, UBO, funding | rescue analytics, friction-vs-risk dashboard |
| Wrong denial communication | product decline, fraud decline and KYC failure share one message | reason taxonomy and message owner |
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
8. Model Risk and AI Governance
开户链路里的 AI 不止一个模型:
| AI / model asset | Risk | Required governance |
|---|---|---|
| ID document classifier | wrong evidence type or missed tamper | eval by document class, quality, region, channel |
| Liveness / PAD / media integrity | false pass, false fail, bias, injection bypass | adversarial eval, segment QA, vendor monitoring |
| Attribute matching | name/address/DOB false mismatch | match policy, thresholds, manual review |
| Synthetic identity graph | unfair or opaque network inference | feature lineage, reviewer evidence, calibration |
| Fraud score | false decline or mule account pass | outcome monitoring, override QA, early-life loss linkage |
| CDD risk tiering | under-risking high-risk profiles or overburdening low-risk customers | typology coverage, scenario review, Compliance owner |
| LLM evidence assistant | hallucinated summary or unsafe customer wording | grounded summary only, citation requirement, output policy |
| Routing model | queues overloaded or wrong owner | queue 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
9. Anti-Patterns
| Anti-pattern | Why it fails | Better architecture |
|---|---|---|
KYC failed as universal reason | owner, evidence and communication collapse | controlled reason taxonomy by decision domain |
| IDV vendor pass equals CIP pass | proofing signal is not full compliance decision | CIP decision service with policy mapping |
| Decline before queue triage | false rejects and poor customer recourse | hold/RFI/review state with SLA and evidence |
| AML, fraud and product decline share message | confidentiality, fairness and customer harm risks | customer-safe message catalog by owner |
| Funding before critical gates clear | mule and synthetic accounts can activate | restricted opening and activation control |
| LLM writes denial reason | invented reasons and inconsistent notices | structured reason code source of truth |
| BO data overwritten by latest source | destroys provenance and auditability | source-ranked evidence graph |
| Conversion optimized alone | hides abandoned harmed customers | balanced risk, harm and activation metrics |
10. 面试表达
30 秒版本:
我会把 AI 开户设计成 decision-and-evidence architecture，而不是把 OCR、活体、名单筛查和欺诈分数简单串联。核心是建立状态机和 gate: consent、eligibility、identity proofing、CIP verification、CDD risk profile、KYB/UBO、fraud、AML handoff、funding activation 和 customer communication。每个 gate 都有 owner、reason code、evidence bundle、人工复核和客户安全文案。AI 可以辅助评分、路由和摘要，但不能替代 Legal/Compliance 的适用性判断、SAR decision 或正式 denial/adverse communication ownership。
2 分钟版本:
我的架构会先分离四条责任线: 产品资格、身份保证、金融犯罪风险、账户激活。身份 proofing 按 NIST SP 800-63-4 拆成 resolution、evidence validation、attribute validation、identity verification 和 enrollment；CIP/CDD 按 FFIEC/FinCEN source anchors 转成机构政策和证据要求；小企业 KYB 要区分 entity exists、authorized representative、beneficial owners/control person 和 business purpose。 技术上我会设计 journey orchestrator、policy decision service、identity/CIP/CDD services、fraud graph、AML referral interface、exception workbench、customer communication service 和 evidence ledger。所有模型输出都进入结构化 decision signal，不允许 LLM 发明拒绝原因。 运营上我会重点看 false reject、manual overturn、RFI burden、abandonment、review SLA、segment friction disparity、early-life fraud/AML quality 和 complaint/appeal outcome。开户成功不是“更多人过审”，而是合规、可解释、客户负担可控、风险可管理、事后能 replay。
11. Portfolio Exercise
为一个 digital checking + small business checking onboarding 设计:
- Account opening state machine。
- CIP / CDD / KYB / UBO decision gate map。
- Risk-tiering table and activation restrictions。
- Exception queue taxonomy and SLA。
- Customer-safe hold / decline reason catalog。
- Evidence bundle schema。
- AI model inventory and eval plan。
- 2 页 architecture memo + 2 分钟面试回答。