AI 底层逻辑 / 经典论文
Banking AI Reference Models：BIAN / FIBO / ISO 20022
一句话:
297 行ai-foundations/papers/92-banking-ai-reference-models-bian-fibo-iso20022.md
Banking AI Reference Models / BIAN / FIBO / ISO 20022 解读
面向对象: Financial Services AI Architect / Banking Product Architect / Business Architect / Senior BA / AI Product Manager。 核心问题: 金融 AI 项目如果只从当前系统字段和页面流程出发, 很容易把银行内部历史包袱误当成业务本质。BIAN、FIBO、ISO 20022 这类行业参考模型可以帮助你识别能力边界、金融概念语义、消息结构和跨系统集成语义。 学习目标: 把银行行业参考模型用于 AI 需求、RAG 知识边界、工具契约、语义层、数据治理、审计证据和跨系统集成。
Source Anchors
| Source | Link | 用途 |
|---|---|---|
| BIAN | https://bian.org/ | 参考银行业务能力、service domain、service landscape 和银行架构标准化边界 |
| FIBO | https://spec.edmcouncil.org/fibo/ | 参考金融行业本体、概念关系和语义一致性 |
| ISO 20022 | https://www.iso20022.org/ | 参考金融消息标准、业务概念、支付/证券/现金管理消息语义 |
| NIST AI RMF | https://www.nist.gov/itl/ai-risk-management-framework | 将行业模型映射到 AI 风险、治理、测量和管理 |
| W3C RDF | https://www.w3.org/RDF/ | 为金融语义图和概念关系提供基础 |
一句话:
Banking AI Reference Models 是让 AI 系统不要只学习某个银行的局部表结构, 而是对齐行业级业务能力、金融概念和消息语义。
1. 为什么金融 AI 需要行业参考模型
金融零售系统通常有这些问题:
- 一个概念在多个系统里叫法不同。
- 一个字段在不同产品线含义不同。
- 历史系统把流程、权限、数据结构混在一起。
- 监管报告、支付消息、客户服务、风险系统使用不同语义。
- AI RAG 读取文档时容易把政策、产品、操作手册、监管定义混用。
- Agent 调用工具时不知道系统字段和真实业务概念的差异。
行业参考模型能提供:
| Reference model | 解决的问题 |
|---|---|
| BIAN | 银行业务能力和服务边界 |
| FIBO | 金融概念、关系、定义和本体 |
| ISO 20022 | 金融消息、交易、支付、现金管理和证券业务语义 |
AI 产品/架构价值:
- 定义 RAG 知识边界。
- 对齐 domain vocabulary。
- 设计 tool/API contract。
- 建立 semantic layer。
- 识别系统字段与业务概念差距。
- 形成审计可解释证据。
- 减少 AI 在金融概念上的误解。
2. BIAN for AI Capability and Service Boundaries
BIAN 的关键启发是: 银行业务可以按 service domains 和能力边界来组织, 而不是按历史应用系统组织。
AI 映射:
| BIAN-inspired lens | AI architecture usage |
|---|---|
| Service domain | 定义 AI tool/API 边界, 避免跨域工具过宽 |
| Business capability | 识别 AI use case 应支持的能力 |
| Service operation | 设计 agent 可调用的操作契约 |
| Service landscape | 识别跨域依赖和平台化机会 |
| Control record | 建立业务对象和状态边界 |
示例:
| Banking area | AI use case | Boundary question |
|---|---|---|
| Customer profile | KYC assistant | AI 能读哪些客户资料, 是否能更新 KYC status |
| Card dispute | dispute agent | AI 只能 triage, 还是能创建 case note |
| Loan servicing | credit policy assistant | AI 是否能解释贷款条款, 是否能修改还款计划 |
| Fraud operations | alert triage copilot | AI 能否合并 alerts, 谁做最终 decision |
| Payments | payment investigation copilot | AI 能否读取 ISO 20022 message details |
BIAN 的价值不是照抄模型, 而是把 AI capability 放进更稳定的银行业务服务边界。
3. FIBO for Financial Semantics
FIBO 适合处理金融概念和关系:
- legal entity。
- contract。
- obligation。
- instrument。
- account。
- transaction。
- party role。
- product。
- risk exposure。
- regulatory concept。
AI 架构中 FIBO 的使用方式:
| AI component | FIBO usage |
|---|---|
| RAG taxonomy | 用金融概念组织文档和 metadata |
| Knowledge graph | 建立 entity/relation/schema |
| Semantic layer | 对齐指标、字段和业务概念 |
| Eval set | 防止模型混淆 instrument/account/contract/party |
| Tool contract | 用业务概念命名字段, 而不只是系统字段 |
| Audit evidence | 解释回答或决策使用了哪些金融概念 |
示例: 信贷 AI copilot
| Concept | Common confusion | Semantic correction |
|---|---|---|
| Borrower | applicant、customer、obligor 混用 | 区分申请人、借款人、责任方 |
| Loan account | contract、facility、account 混用 | 区分合同、额度、账户 |
| Delinquency | late fee、default、forbearance 混用 | 区分逾期状态、违约、宽限安排 |
| Interest rate | APR、nominal rate、effective rate 混用 | 标注计算语义和适用上下文 |
4. ISO 20022 for Message and Payment Semantics
ISO 20022 对 AI 的价值在于金融消息和交易语义。
适用场景:
- 支付调查 copilot。
- reconciliation assistant。
- cash management assistant。
- securities operations assistant。
- regulatory reporting helper。
- fraud and sanctions investigation。
AI 映射:
| ISO 20022 aspect | AI design usage |
|---|---|
| Business message | RAG/source parsing and message explanation |
| Message element | structured extraction and validation |
| Business process | event-driven workflow and case routing |
| Code set | controlled vocabulary in schema/eval |
| Supplementary data | exception handling and audit evidence |
支付调查例子:
| Question | AI needs |
|---|---|
| 这笔付款为什么被退回 | parse payment status/report message, reason code, party data |
| 谁是 debtor/creditor/intermediary | understand party roles and agent banks |
| 是否可能是 sanctions/AML issue | connect message data to screening workflow |
| 是否需要人工升级 | risk tier, policy, exception workflow |
ISO 20022 不等于“AI 要读 XML”。它提供的是业务消息语义和标准化字段关系, 可用于 extraction、validation、tool contract 和 audit。
5. Reference Model Mapping Matrix
| AI artifact | BIAN | FIBO | ISO 20022 |
|---|---|---|---|
| Capability map | service domain / service landscape | business concept dependency | business process area |
| RAG source taxonomy | banking domain category | ontology class and relation | message/business area |
| Tool contract | service operation boundary | semantic field meaning | message element / code set |
| Eval set | domain-specific task | semantic confusion cases | message interpretation cases |
| Evidence graph | business service impacted | concept lineage | message instance evidence |
| Product metrics | service outcome | concept-specific metric | process/message SLA |
高级能力是能把三者组合:
BIAN tells where the capability belongs.
FIBO tells what the financial concept means.
ISO 20022 tells how financial messages express events and data.
6. Financial Retail Case: Payment Investigation Copilot
目标: 帮助运营人员解释付款状态、失败原因、参与方角色、下一步动作和升级路径。
6.1 Reference Mapping
| Layer | Mapping |
|---|---|
| Business capability | Payment operations / payment investigation |
| BIAN boundary | payment execution / payment order / correspondence-style domains |
| FIBO semantics | account, party, obligation, transaction, contract, role |
| ISO 20022 semantics | payment status, return reason, debtor, creditor, agent, message id |
| AI components | message parser, RAG policy retrieval, tool gateway, case summarizer |
| Controls | PII masking, sanctions boundary, human escalation, audit trace |
6.2 AI Task Boundary
| Task | AI can do | AI cannot do |
|---|---|---|
| Explain status | Explain reason code and message path | Override bank status |
| Summarize evidence | Gather message, case, policy context | Invent missing data |
| Recommend next action | Suggest investigation checklist | Execute payment reversal |
| Draft customer note | Draft internal/customer response | Send without approval |
6.3 Eval Cases
| Eval type | Example |
|---|---|
| Message interpretation | Correctly explain status/reason code |
| Party role | Distinguish debtor, creditor, debtor agent, creditor agent |
| Policy grounding | Cite internal payment operations policy |
| Risk boundary | Escalate sanctions-related or AML-related uncertainty |
| Tool correctness | Use message lookup before summarizing |
| No hallucination | Refuse to infer missing intermediary details |
7. Banking AI Portfolio Uses
| Use case | Reference model leverage |
|---|---|
| KYC assistant | BIAN customer/party capability + FIBO party/legal entity semantics |
| AML investigation | FIBO entity/relationship + BIAN risk/operations boundary |
| Credit policy assistant | FIBO contract/obligation/product semantics |
| Payment investigation | ISO 20022 message semantics + BIAN payment domains |
| Wealth suitability assistant | FIBO instrument/product/risk semantics |
| Regulatory reporting helper | FIBO regulatory concepts + ISO 20022 message lineage |
8. Templates
Reference Model Mapping
| Field | Content |
|---|---|
| AI use case | Use case name |
| Business capability | BIAN-inspired capability / domain |
| Key financial concepts | FIBO-aligned terms |
| Message / transaction semantics | ISO 20022 message or element if relevant |
| RAG boundary | Approved sources and taxonomy |
| Tool boundary | Allowed operations and side effects |
| Eval cases | Concept/message/domain-specific tests |
| Evidence | Source lineage, message id, trace, approval |
Semantic Gap Log
| Gap | Example | Impact | Remediation |
|---|---|---|---|
| Field ambiguity | customer_id means party in one system, account holder in another | wrong retrieval/tool action | semantic mapping and glossary |
| Process ambiguity | payment return vs cancellation | wrong customer explanation | ISO 20022 reason code mapping |
| Product ambiguity | facility vs loan account | credit policy error | FIBO concept alignment |
| Domain boundary | fraud ops vs AML investigation | wrong escalation | BIAN capability mapping |
9. Common Failure Modes
| Failure mode | 表现 | 修正 |
|---|---|---|
| System-field driven AI | AI 学会历史字段, 没学会金融概念 | map fields to FIBO/semantic glossary |
| Generic chatbot over banking docs | 跨域回答混乱 | BIAN domain boundary + source registry |
| Message parsing without business meaning | 只抽 XML 字段 | ISO 20022 business concept mapping |
| Ontology overkill | 为小场景建庞大本体 | use thin semantic slice |
| Reference model theater | 引用标准但不影响架构 | connect to tool, eval, RAG, evidence |
10. 面试表达
30 秒版本:
我会用 BIAN、FIBO、ISO 20022 帮金融 AI 建立行业语义边界。BIAN 帮我定义银行能力和服务边界, FIBO 帮我统一金融概念, ISO 20022 帮我理解支付和金融消息语义。这样 RAG、tool contract、eval、semantic layer 和 audit evidence 不会只依赖某个银行的历史字段和局部流程。
2 分钟版本:
以 payment investigation copilot 为例, 我会先用 BIAN 定义它属于支付运营/调查能力, 再用 ISO 20022 对付款状态、参与方角色、reason code 和 message id 建立结构化语义, 用 FIBO 对 party、account、transaction、obligation 等金融概念做统一。RAG 只接入 approved payment policy, tool contract 使用标准化业务字段, eval 包含 message interpretation、party role、policy grounding 和 risk escalation。这样 AI 可以解释和汇总, 但不能越权改状态或执行退款。
11. Practice Assignment
选一个金融 AI use case, 完成:
- BIAN-style capability/domain mapping。
- FIBO concept slice。
- ISO 20022/message semantics mapping, 如适用。
- RAG source taxonomy。
- Tool/API contract semantic field list。
- 10 条 semantic eval cases。
- Semantic gap log。
- 2 分钟面试叙事。
完成标准:
- 至少 10 个核心术语有稳定定义。
- 每个工具字段都映射到业务概念。
- 每个高风险概念混淆都有 eval case。
- 能解释行业模型如何降低 AI 语义风险。