AI Document Intelligence：非结构化数据与证据质量架构
Document intelligence 不是 “OCR + LLM summary”。在金融运营里，它是一套 evidence system：把非结构化文档转换成有 provenance、source coordinate、confidence calibration、validation、human review、records/hold metadata、workflow policy 和
AI Document Intelligence / Unstructured Data / Evidence Quality Architecture 解读
配对阅读：本篇的操作手册版（模板/RACI/门禁/runbook）是
docs/AI_DOCUMENT_INTELLIGENCE_UNSTRUCTURED_DATA_EVIDENCE_QUALITY_PLAYBOOK.md。第一遍读本篇建立原理与架构判断；第二遍做案例时再用 playbook 查表落地，两者不需要重复精读。
重要说明: 本文只讨论 bank statement、paystub、claim package、dispute evidence、KYC/KYB 文件、insurance / loan servicing 文档和运营来信中的 AI document intelligence 产品与架构设计，不构成法律、监管、记录保留、e-discovery、KYC/KYB 充分性、贷款/保险承保、消费者争议处置、欺诈处置、模型验证或供应商推荐结论。记录、证据、legal hold、KYC/KYB、信贷、保险、投诉、争议、索赔、跨境数据和 e-discovery 的具体适用性必须由 Legal、Compliance、Records Management、Privacy、Model Risk、Fraud Risk、Operations、Data Governance、Vendor Management、Internal Audit 等共同确认。
Source Anchors
| Source | Link | 用途 |
|---|---|---|
| NIST AI Risk Management Framework | https://www.nist.gov/itl/ai-risk-management-framework | 用 Govern / Map / Measure / Manage 组织 document AI 的风险治理、eval、monitoring、human oversight、incident and evidence controls |
| NIST Privacy Framework | https://www.nist.gov/privacy-framework | 用 privacy risk management、data minimization、purpose、processing、access and monitoring 设计文档数据采集、抽取、使用和保留边界 |
| NARA Records Management | https://www.archives.gov/records-mgmt | 用 records lifecycle、disposition、records program 和 accountability 作为 records retention / evidence management 的官方锚点 |
| NARA Electronic Records Management | https://www.archives.gov/records-mgmt/policy/transfer-guidance-tables.html | 用电子记录格式、metadata、transfer/readiness guidance 作为 electronic records architecture and preservation discussion 的锚点 |
| CFPB Consumer Complaint Database | https://www.consumerfinance.gov/data-research/consumer-complaints/ | 用消费者投诉和 complaint operations 视角校验 document evidence trace、dispute handling、case explanation 和 operational learning loop |
| FFIEC Authentication and Access to Financial Institution Services and Systems | https://www.ffiec.gov/press/pr081121.htm | 用金融机构认证、访问控制、风险评估和 layered security 思路设计 document intake、reviewer access、workflow action 和 privileged operation controls |
| ISO/IEC 42001 overview | https://www.iso.org/standard/42001 | 用 AI management system、roles、operation、performance evaluation、internal audit 和 continual improvement 建立 document AI operating model |
核心导读
Document intelligence 不是 “OCR + LLM summary”。在金融运营里，它是一套 evidence system：把非结构化文档转换成有 provenance、source coordinate、confidence calibration、validation、human review、records/hold metadata、workflow policy 和 audit replay 的决策输入。
问题定义
文档智能的目标不是“更快读文档”，而是把文件变成可依赖的 operational evidence：
uploaded PDF / scanned image / email attachment / photo / fax
-> evidence envelope
-> document provenance
-> page/layout map
-> extracted entities with source coordinates
-> normalized facts
-> confidence and validation results
-> fraud/tamper signals
-> human review decisions
-> workflow action
-> records retention / legal hold metadata
-> replayable audit trail
关键边界：
Text recognized does not mean fact established.
Model summary does not mean evidence accepted.
Confidence score does not mean business risk resolved.
Human review is a control only when it is designed, sampled and evidenced.
Document storage does not mean records compliance.
金融零售常见文档包括贷款收入证明、insurance claim photos/invoices、chargeback packages、KYC/KYB registration、death certificate、power of attorney、complaint letters、merchant correspondence、call transcripts 和 servicing forms。一个字段错误可能影响 credit、claim payout、dispute outcome、account authority、complaint remediation 或 regulatory response。
核心原理/方法
document AI 至少要区分八类对象：
| Object | Definition | 控制含义 |
|---|---|---|
| Raw document | 客户/商户/员工/系统提交原始文件 | hash、source channel、received time、case binding |
| Rendered page | 渲染后的 page image / normalized PDF | renderer version、page count、quality |
| Layout element | 表格、段落、checkbox、signature、stamp、logo | coordinates、reading order、structure |
| Extracted field | 从文档抽取的字段和值 | page、bounding box、confidence、model version |
| Normalized entity | 标准化姓名、地址、金额、日期、企业名 | normalization rule、entity resolution |
| Derived fact | 多字段或规则计算的事实 | formula、input fields、calculation version |
| Decision evidence | 被业务 policy 接受或人工确认的证据 | policy reason、reviewer/action trace |
| Summary | 面向 reviewer/客户的 source-linked 摘要 | 不得替代原证据 |
字段级 metadata：
field_name
document_id / document_version
page_number / bounding_box_or_anchor
raw_text
normalized_value
extraction_method
model_or_rule_version
confidence_score
calibration_bucket
validation_results
cross_document_match
fraud_or_tamper_signals
human_review_status
policy_acceptance_status
evidence_retention_rule
legal_hold_flag
系统/架构模型
参考架构：
intake channels
-> document capture and provenance service
-> file normalization / rendering / malware and content safety scan
-> document classification and package splitting
-> OCR + layout understanding + table extraction
-> multimodal extraction / entity extraction / relationship mapping
-> normalization and business validation
-> confidence calibration and quality scoring
-> fraud / tamper / duplicate / synthetic-document checks
-> evidence policy engine
-> human review and exception queues
-> workflow integration: KYC, claims, disputes, servicing, complaints
-> records retention / legal hold / disposition integration
-> evidence ledger, monitoring, QA and model governance
核心组件：
| Component | 职责 |
|---|---|
| Intake gateway | upload、email、fax、branch scan、mobile capture、API、vendor feed，并绑定 case/session |
| Provenance service | document id、hash、timestamp、source、custody events、version |
| Classification service | document type、issuer/source、language、package boundary、record category candidate |
| OCR/layout service | text、reading order、tables、checkboxes、signatures、coordinates |
| Multimodal extraction | schema-constrained fields and relationships，输出必须 source-linked |
| Entity normalization | 金额、日期、账号、企业名、地址、policy/claim id 标准化 |
| Validation/reconciliation | 跨页、跨文档、内部系统、第三方数据一致性 |
| Confidence engine | field/document/case confidence and calibration |
| Fraud/tamper service | metadata、visual、duplicate、arithmetic、synthetic signals |
| Evidence policy engine | 判断 extraction 是否可用于当前 workflow |
| Human review workbench | source-first UI、conflicts、model rationale、structured decision |
| Records/hold connector | record class、retention schedule、legal hold、disposition |
| Evidence ledger | document、model、rule、review、workflow、final communication trace |
关键机制与取舍
Document classes 决定 extraction schema、review threshold、fraud checks、retention 和 workflow route：
| Document class | High-value fields | 风险 | 控制 |
|---|---|---|---|
| Bank statements | account holder、period、balances、deposits、NSF | altered PDF、missing pages、fake template | completeness、transaction table validation、issuer checks |
| Paystubs | employer、employee、period、gross/net、YTD | fake paystub、YTD inconsistency | arithmetic、period consistency、employer validation |
| Claims documents | loss date、policy、coverage、invoice、photos | inflated invoices、reused photos | timeline、media metadata、duplicate search |
| Dispute packages | transaction、merchant proof、customer assertion | weak evidence、over-summary | reason-code evidence checklist |
| KYC/KYB docs | identity、registration、ownership、license、signer authority | stale docs、entity mismatch | freshness、entity resolution、review boundary |
| Servicing authority docs | death certificate、POA、court order | entitlement/privacy/legal risk | privileged workflow、dual review |
| Complaints/correspondence | allegations、harm、attachments | missed complaint、wrong product | complaint taxonomy and final response linkage |
Confidence 不能只看一个分数：
| Confidence level | 示例 | 用途 |
|---|---|---|
| Character/text | 8 vs B in account mask | OCR quality |
| Field | pay period end date | field acceptance/review |
| Layout | deductions table structure | table validation |
| Classification | paystub vs payroll summary | workflow/retention route |
| Cross-validation | YTD income vs pay period | consistency |
| Fraud/tamper | PDF metadata anomaly | review signal |
| Case evidence | package sufficient for next step | workflow decision |
Human review 是控制，不是橡皮图章。Reviewer 应看到 model suggestion、source location、validation failures、fraud signals、prior versions、policy use 和 required reason code；对 high-impact fields 如 income、authority、claim amount、beneficial owner，应有更高阈值或 second review。
Records/legal hold 机制：
- raw docs、OCR text、layout JSON、extracted fields、summaries、review notes、exports 可能有不同 record status。
- legal hold 必须传播到 raw、derived artifacts、case decisions、exports、vendor copies 和 deletion/training jobs。
- reprocessing 不能 silent overwrite 历史结果；旧版本应按 policy 保留 lineage。
证据与控制
Document evidence envelope：
document_id
case_id
customer / business reference
source channel
received timestamp
submitter / system reference
raw file hash
file type and size
page count
document class / subtype
classification confidence
language / locale
quality score
processing lineage
fraud/tamper signals
record class
retention rule
legal hold flag
access restrictions
derived artifacts
workflow decisions
complaint / audit links
控制矩阵：
| Control objective | Control activity | Evidence |
|---|---|---|
| 保留原始证据 | immutable raw artifact、hash、received timestamp、source | document hash、intake event |
| 正确分类 | type/subtype model with ambiguity routing | classification result、review decision |
| 字段绑定来源 | 每个 extracted field 包含 page/coordinate/anchor | extraction JSON、UI source link |
| 验证关键字段 | amounts、dates、identity、authority 的 rule/cross-doc checks | validation log |
| 校准 confidence | 与 reviewer outcomes and sampling 对比 | calibration report |
| 防 unsupported automation | field criticality thresholds and policy gates | policy decision id |
| 控制 summaries | source-linked summaries with prohibited conclusions | model run、citations、eval |
| route human review | ambiguity、high impact、authority、fraud、records sensitivity | task id、reviewer、rationale |
| 检测 fraud/tamper | metadata、visual、duplicate、arithmetic checks | risk signals、fraud case link |
| 管理 records/hold | record class、retention、hold propagation | record metadata、hold log |
| 支持 complaints/audit | docs、AI runs、review、decision、final message linked | evidence bundle、complaint id |
指标应看 field-level precision/recall、confidence calibration、review overturn、evidence completeness、hold propagation、fraud/tamper yield、privacy access defects、downstream rework、complaint defects、audit replay success，而不只看自动化率。
金融零售/AI产品场景
- Loan income review：paystub and bank statement extraction 只生成 income evidence package；affordability 或 credit decision 仍需 policy and review。
- Insurance claim：photos、invoices、police report 形成 timeline，duplicate media 和 inflated invoice signals 进入 adjuster review。
- Payment dispute：merchant receipt、tracking proof、customer statement 通过 reason-code evidence checklist 组织，AI summary 不覆盖原证据。
- KYC/KYB onboarding：business registration、license、authorized signer docs 做 entity resolution and freshness check；beneficial ownership ambiguity 升级。
- Bereavement servicing：death certificate、POA、court order 进入 privileged workflow，减少重复提交并控制敏感访问。
- Complaint handling：attached letters and screenshots 被链接到 complaint_id、AI run、final response 和 CAPA。
反模式
| 反模式 | 风险 | 更好的控制 |
|---|---|---|
| OCR text treated as truth | 金额、日期、姓名可能读错 | source-linked fields、validation、review |
| document-level confidence 决定所有字段 | 平均分掩盖关键字段错误 | field criticality thresholds |
| LLM summary 成为 decision record | 遗漏冲突或编造结论 | source-linked summary + structured evidence |
| wrong classification | 错 workflow、SLA、retention class | ambiguity queue and QA |
| silent reprocessing overwrite | audit 无法解释历史决定 | versioned artifacts and lineage |
| legal hold 不传播 | derived artifacts 被删除 | hold-aware artifact graph |
| reviewer rubber-stamping | automation bias | source-first UI、reason codes、QA |
| fraud model 直接阻断客户 | false positive harm | risk signal + human/fraud review |
| vendor retains docs unexpectedly | privacy/records/legal hold exposure | contract controls and deletion evidence |
| prompt injection in documents | 模型遵循文档内恶意指令 | tool isolation and output validation |
最终心智模型
Document intelligence 的本质是 evidence operating architecture。成熟系统应能证明：每个自动或人工辅助的文档决定都基于正确原件、正确 source-linked fields、正确 confidence/validation、正确 human review boundary、正确 fraud/records/legal hold treatment，以及当时适用的 workflow policy。缺的不是更强 OCR，而是 provenance、evidence quality、records integration 和 AI governance 的闭环。
SOTA 状态标注 (2026-07-01)
本篇属于第二、三遍深读池（参考架构/深读笔记），未列入 12 周主线必读。时效基线为写作时点；引用前请按 CLAUDE.md 全局时效性硬规则复查最新进展。模块级 SOTA 对照见 docs/AI_SYSTEMATIC_LEARNING_ROADMAP_2026.md 各周「2026 SOTA 对照」行与文末「SOTA 检查」。