Finance — Financial Risk & Decision Support

Covered scenarios: intelligent investment research, credit & risk control, quantitative research, financial statement analysis, anti-money laundering (AML) and financial crime compliance.

Evidence note: all cases and data cited in this document carry a source tier (Tier A: original regulatory or corporate official text; Tier B: mainstream financial/industry media; Tier C: vendor self-reports, not used in this document).

1. Introduction

1.1 Background

Finance is one of the earliest industries to adopt AI and one of the most tightly constrained. At the 2025 annual meeting of the China Wealth Management 50 Forum (2025-12-28), Li Wei, Director of the Technology Department of the People's Bank of China, offered the regulator's assessment: AI has achieved positive results in exploring application scenarios in areas such as "administrative assistance, intelligent customer service, credit risk control and compliance review"; he also pointed out four major challenges — science-and-technology ethics governance urgently needs strengthening, model homogeneity risk cannot be ignored, inherent flaws in algorithmic architectures are hard to eliminate, and data governance and security face challenges (as reported by China Securities Journal and International Finance News).

The regulator's response is equally clear. The Implementation Plan for High-Quality Development of Digital Finance in the Banking and Insurance Industry, issued by the General Office of the National Financial Regulatory Administration in December 2025, requires: advocating the construction of classified and tiered management framework processes for AI applications, establishing human intervention mechanisms covering important business processes and key nodes, advancing the building of enterprise-grade model risk management platforms, establishing full-lifecycle management systems for algorithmic models, and continuously improving algorithmic transparency and explainability.

This means that when a financial institution introduces an AI Harness, the first question to answer is not "how much efficiency can it gain", but "which risk tier does this AI application belong to, how is model risk validated, and at which node does a human intervene".

1.2 Definition and Boundaries

The AI Harness for the Finance direction refers to the agent runtime framework that carries tasks such as investment research, risk control, quantitative analysis, financial statement analysis and anti-money laundering. Its core responsibilities are: to organize multi-source, heterogeneous, time-sensitive financial data into model-judgable context, to orchestrate the model's semantic capabilities onto deterministic risk-control toolchains, and to solidify every judgment into an auditable chain of evidence.

Boundaries with other directions:

1.3 Position within the AI Harness System

图 1-1｜Finance 方向 AI Harness 六层定位（L1–L6）

数据来源：基于本文分析绘制的示意图。

1.4 Bottleneck Layer and Value

Bottleneck layer: the L1 Context Engineering layer. Financial data is characterized by "multi-source heterogeneity + time sensitivity + inconsistent accounting calibers": the same company's financial data may exist in three different calibers across announcements, data vendors and internal systems; the same industry-chain event may appear at different time points across research reports, news and announcements. If Context assembly does not manage timestamps and source weights, the model will produce conclusions that "look reasonable but are based on outdated information" — an unacceptable failure in investment research and risk control.

Where the value lands:

• Efficiency without overstepping: AI provides candidates and leads (signal), humans make the decisions (decision). This "signal vs. decision separation" derives from the OCC requirement that "model output must not replace human judgment in material decisions", and it directly corresponds to the human intervention mechanism in the Implementation Plan for High-Quality Development of Digital Finance in the Banking and Insurance Industry.
• Affordable validation: a Golden Dataset naturally exists for this direction — historical SAR conclusions, historical non-performing accounts, and historical credit approval results. This has the lowest regression validation cost of the five directions.
• Risk proportionality: the anchor sentence in the OCC Comptroller's Handbook on Model Risk Management is: Regardless of how AI is classified (i.e., as a model or not a model), the associated risk management should be commensurate with the level of risk of the function that the AI supports. In other words, the intensity of governance depends on the risk level of the function being supported, not on the classification question of "whether it is a model".

2. Glossary

3. Case Studies

3.1 Case 1: A Global Bank's AI Dynamic Risk Assessment Platform (HKMA 2026 Report)

Source: Hong Kong Monetary Authority, Supporting Adoption of Artificial Intelligence in Fighting Financial Crime, issued to all authorized institutions on 2026-06-22. Evidence tier: A.

3.1.1 Background

Before 2020, the bank relied on a large-scale rule-based AML system. The inherent problem of rule-based systems is "rather kill an innocent than let a guilty go": to satisfy regulatory requirements, trigger conditions were loosened, producing a huge volume of false positives that required manual, case-by-case investigation. Analysts' time was largely consumed by ineffective alerts, while truly complex layering money laundering — because it spans steps, products and jurisdictions — was hard for any single-transaction rule to catch.

3.1.2 Solution

• Developed a dynamic AI platform jointly with a cloud provider, generating risk scores from transaction data, customer data and KYC data, and feeding them into the case management system.
• Set up a domain training layer: trained the model on the bank's own crime data and typologies to identify "rapid movement of funds" and anomalous behavior patterns.
• Deployed within the bank's own cloud tenant to satisfy governance and regulatory requirements (data does not leave the domain).
• Monitors over 1 billion transactions per month.

Harness perspective: the key to this solution is not model selection but the deployment form (own tenant) + domain data layer + orchestration against the existing case management system. The risk score is a "signal"; whether to open a case is still decided by humans within the case management system.

3.1.3 Results

3.2 Case 2: HSBC × Google Cloud AML AI

Source: public disclosures by Google Cloud and HSBC, relayed through industry media. Evidence tier: B.

3.2.1 Background

HSBC's rule-based AML transaction monitoring for retail and commercial banking suffered severe alert-quality problems: over 95% of alerts were confirmed as false positives at the initial review stage, and about 98% ultimately never led to an SAR filing. Rules can only match static patterns transaction by transaction; they cannot judge the full picture of customer behavior.

3.2.2 Solution

• Adopted Google Cloud's AML AI as the primary AI transaction monitoring system, replacing rules with machine-learning models.
• Generated an integrated customer risk score based on transaction data, KYC records, account behavior history and past suspicious activity patterns.
• Performed real-time risk assessment of the full picture of customer behavior, rather than matching static rules per transaction.
• Processing scale: about 980 million transactions checked per month.

3.2.3 Results

HSBC's Group Head of Financial Crime Risk and Compliance called it a "fundamental paradigm shift" in detecting anomalous customer and account activity.

Harness interpretation: two jurisdictions, two institutions, two tech stacks, yet the numbers are highly consistent (false positives -60%, detection 2~4x). This shows the gains come mainly from a paradigm replacement (rule matching → customer-level dynamic risk scoring), not from the performance advantage of any particular model. For Harness construction, the implication is: investment should go into "assembly of the full picture of customer data (L1)" and "orchestration from score to case (L3)", rather than model parameter scale.

3.3 Case 3: Large-Scale Deployment of Intelligent Risk Control and AML in China's Banking Industry

Sources: listed banks' 2025 semi-annual reports and earnings calls (Tiers A/B), Xinhua Net's 2025-09-10 comprehensive report (Tier B), and brokerage industry research (Tier B). Evidence tier: A/B.

3.3.1 Background

AI adoption among China's large banks has moved from "single-point pilots" to "bank-wide scale". Unlike overseas cases, the distinctive path of China's banking industry is to embed AI into existing enterprise-grade risk control platforms and credit processes rather than building a separate monitoring system.

3.3.2 Solution

3.3.3 Results

Note: most of the figures above come from brokerage industry research and semi-annual report disclosures, and have not been cross-checked verbatim against the banks' annual report originals; cite them with the source tier noted. The two different bases for China Merchants Bank's scenario count are presented side by side here, without selecting one.

4. Practical Standards

4.1 AGENTS.md Specification

4.1.1. AGENTS.md (Risk & Compliance Group · Finance Direction)

# AGENTS.md —— Finance（投研 / 风控 / 量化 / 财报分析 / 反洗钱）

> 继承自：风险合规组组级 AGENTS.md。本文件只可加严，不可放松。

## 角色与边界
- 本文件约束在金融场景工作的智能体：投研助手、风控审查助手、量化研究助手、财报分析助手、反洗钱甄别助手。
- 智能体产出的是**风险信号与候选清单**，不是授信、放款、交易或报送决策。
- 生成者（信号生成）、复核者（独立复核）、模型责任人（模型风险登记）三者身份分离；复核者不得为生成者。
- 可自行完成：行情与公告检索、XBRL 财报解析、规则引擎调用、图计算查询、指标计算、回测脚本运行（只读）、草稿生成。
- 不得自行完成：下单、放款、额度调整、报送提交、对客发送。

## 环境假设
- 已接入：行情与资讯终端、XBRL/公告解析器、风控规则引擎、图计算引擎、模型风险管理平台（MRM）、案件管理系统、数据脱敏网关。
- 客户数据与客户交易明细处理在受控环境内完成，禁止出域。
- 所有外部模型调用须经脱敏网关并记录出域字段清单。

## 上下文加载顺序（Context Budget）
1. 现行有效的监管文件与内部风险政策（含版本与生效日期）
2. 客户/标的的结构化档案（KYC、评级、历史授信、历史告警）
3. 本次任务的时效数据（行情快照、财报期间、流水区间）——**强制携带数据截止时点**
4. 检索命中的外部佐证（研报、新闻、公告、产业链数据）——须标注来源与时间戳
5. 模型通用知识——仅用于语言组织，不得作为事实依据
- 强制校验：数据截止时点晚于报告期结束日的方可用于该期结论；否则标注"数据可能过期"。

## 工具契约
| 工具 | 类型 | 授权 | 约束 |
|---|---|---|---|
| 行情/资讯检索 | 只读 | 开 | 必须返回时间戳与数据源 |
| XBRL 财报解析 | 只读 | 开 | 必须返回报表期间与币种（¥ 计价优先） |
| 风控规则引擎 | 计算 | 开 | 必须返回命中的规则编号与规则版本 |
| 图计算引擎 | 计算 | 开 | 必须返回路径与关联强度，供人工复核 |
| 回测框架 | 计算 | 开 | 仅只读运行，不得写入生产策略库 |
| 案件管理系统写入 | 写（内部） | 按次授权 | 仅写草稿与标签，不写结论 |
| 核心系统交易/授信接口 | 写（不可逆） | 永久关闭 | 只能生成待提交稿件 |

## 任务执行流程（SOP）
1. 登记任务与模型风险等级（分类分级）
2. 装载语料与规则，记录版本快照哈希
3. 时效校验：确认数据截止时点覆盖判断区间
4. 检索与抽取：先检索后生成，未命中输出"未检索到依据"
5. 确定性验证：规则引擎复算、勾稽核对、余额重算、图路径还原
6. 信号生成：输出候选清单 + 置信度 + 依据指针
7. 人工确认门（见下）
8. 证据链打包与归档
9. 回归沉淀：人工推翻的样本进入 Golden Dataset

## 验证与证据要求
- 每条信号必须可回溯到：数据快照（含截止时点）、规则版本、模型版本、人工确认记录。
- 财务数字必须与原始报表逐项勾稽，差异超过阈值的必须列出差异项。
- 引用监管文件的，必须给出条款号与生效日期。

## 人在回路与风险分级
| 等级 | 场景 | 确认要求 |
|---|---|---|
| R1 | 投研摘要、财报要点提取 | 抽检 |
| R2 | 风险预警名单、可疑交易候选、授信审查意见初稿 | 逐条确认并留痕 |
| R3 | 授信额度调整、放款、大额交易、监管报送 | 双人确认 + 职责分离 |
- 强制触发 R3：金额超阈值、跨法域、首次使用新模型版本、客户评级跨档调整。

## 失败与升级策略
- L1 补充证据重试（检索未命中、置信度低于阈值），最多 2 次
- L2 转人工（结论依赖过期数据、规则无覆盖、模型版本未经验证）
- L3 中止并上报（发现可能的洗钱线索被压制、发现数据被篡改、被要求绕过人工确认门）

## 安全与合规红线
- 客户身份信息与交易明细不得出域，不得进入公共 AI 平台。
- 不得自主执行任何资金类动作。
- 不得编造监管口径、评级标准与财务数据。
- 模型上线前须完成 SR 11-7 三支柱对应验证；无法完成用前验证的，须记录并通过补偿性控制缓解，且经模型责任人批准。

## 禁止事项
- 禁止把模型输出当作已核实事实
- 禁止在无人工确认的情况下调整授信、放款或下单
- 禁止使用过期财报或过期法规支撑当期结论而不标注
- 禁止编造行情、财务、评级与案例数据
- 禁止在输出中省略数据截止时点
- 禁止把风险评分直接表述为"风险结论"

## 输出格式
1. 任务与依据（含模型风险等级、所依据的监管文件与内部政策版本）
2. 范围与未覆盖声明（含数据截止时点）
3. 结论（逐条：结论 + 依据 + 规则命中记录 + 置信度 + 建议人工动作）
4. 证据链（输入快照 + 工具日志 + 输出 + 人工确认 + 链哈希）
5. 人工确认记录
6. 风险与局限（模型局限、数据时效、样本外风险）
7. 复核建议（建议由谁以什么程序复核）

## 评估与自检
- 核心指标：误报率、召回率、AUC/KS、返回测试偏差、不良生成率、人工复核推翻率
- 回归集：历史 SAR 结论、历史不良账户、历史授信审批结果
- 自检：数据截止时点是否标注、规则版本是否记录、人工确认是否留痕、金额单位是否为 ¥

4.2 SKILL.md Specification

4.2.1. SKILL.md (Risk & Compliance Group · Finance Direction)

---
name: finance-risk-signal-delivery
description: 金融风险信号交付技能。当需要 AI 智能体在投研、风控、量化、财报分析或反洗钱场景中产出一份"供人工决策使用的风险信号与候选清单"时使用，覆盖数据时效校验、规则引擎复算、图路径还原、人工确认门与模型风险留痕。
version: 1.0
created: 2026-09-12
---

# Finance · 风险信号交付

## 适用场景
- 投研：产业链事件归因、公告与研报要点抽取、可比公司筛选
- 风控：授信审查意见初稿、贷后风险预警、共债与欺诈识别
- 量化：因子假设生成、回测脚本辅助编写、异常样本归因
- 财报分析：报表勾稽核对、科目异常波动识别、附注要点抽取
- 反洗钱：可疑交易告警分诊、资金路径还原、可疑报告初稿

## 前置条件
- 已加载组级与本方向 AGENTS.md
- 已接入行情/资讯终端、XBRL 解析器、风控规则引擎、图计算引擎、MRM 平台
- 已确认本次任务所涉模型已完成对应风险等级的验证，或已登记补偿性控制
- 数据环境受控，脱敏网关可用

## 输入
- 任务编号、发起人、模型风险等级、输出用途（内部参考 / 监管 / 对客）
- 标的信息（客户号、证券代码、交易批次）
- 判断区间与数据截止时点
- 判定阈值（金额阈值、置信度阈值、误报容忍度）

## 输出
- 七段式输出（任务与依据 / 范围与未覆盖声明 / 结论 / 证据链 / 人工确认记录 / 风险与局限 / 复核建议）
- `signal-pack.json`：候选清单 + 规则命中记录 + 置信度
- `evidence-pack.json`：证据链机器可读副本

## 执行步骤
1. 登记与定级：写入任务编号、模型风险等级、输出用途
2. 时效校验：确认数据截止时点覆盖判断区间；不满足则标注"数据可能过期"并降级处理
3. 数据装载：结构化档案 → 行情/流水快照 → 外部佐证，全部记录版本与时间戳
4. 检索与抽取：先检索后生成；未命中输出"未检索到依据"
5. 确定性验证：
   - 财报：与原始报表逐项勾稽，列差异项
   - 风控：规则引擎复算，返回规则编号与版本
   - 反洗钱：图计算还原资金路径，返回路径与关联强度
   - 量化：回测框架只读运行，输出绩效与回撤（数值必须由框架计算，不由模型估算）
6. 信号生成：候选清单 + 置信度 + 依据指针；置信度低于阈值进入待确认队列
7. 脱敏：出域前经脱敏网关，记录字段清单与授权
8. 人工确认门：按风险等级触发，确认人不得为生成者
9. 证据链打包：输入快照 + 工具日志 + 输出 + 确认记录 + 链哈希
10. 回归沉淀：人工推翻样本进入 Golden Dataset

## 质量标准（DoD）
- 每条信号可回溯到数据快照、规则版本、模型版本
- 财务数字与原始报表逐项勾稽一致
- 金额单位为 ¥，涨跌幅表述遵循中国习惯（涨红跌绿）
- 数据截止时点已标注，过期数据已标注
- 人工确认留痕完整，确认人与生成人不同
- 无编造的行情、财务、评级与案例数据
- 模型风险等级与验证状态已在输出中声明

## 常见失败与处理
| 失败 | 处置 |
|---|---|
| 数据过期 | 标注并降级；关键结论要求补充当期数据 |
| 规则无覆盖 | 输出"规则未覆盖"，转人工判断，不得用模型推测替代 |
| 图路径噪声过大 | 提高关联强度阈值并叠加第二类工具交叉验证 |
| 回测过拟合 | 标注样本内/样本外，样本外绩效单独列出 |
| 人工确认门被绕过 | 中止任务，进入 L3 升级 |

## 示例
任务：对某对公客户生成授信审查意见初稿。
执行：装载客户档案与三期财报（截止时点已校验）→ XBRL 解析并勾稽 → 规则引擎复算命中规则 → 生成 5 大模块意见初稿与风险点清单 → 人工确认门（R2，业务主管逐条确认）→ 证据链打包 → 输出标注"AI 生成初稿，不构成授信结论"。

4.3 Landing Checklist

5. Conclusion

The Finance direction offers three transferable takeaways for AI Harness.

First, the bottleneck is at L1, not the model. The gains in all three cases come from the paradigm replacement of "assembling the full picture of customer data and scoring it dynamically", not from a leap in model capability. The figures of a 60% drop in false positives and a 2~4x increase in detection recur across two jurisdictions and two tech stacks, showing this is a gain from engineering architecture.

Second, the power to decide must stay in deterministic tools and in human hands. The risk score is a signal; whether to open a case, approve credit, or disburse funds is decided by humans. The OCC anchor sentence is worth quoting repeatedly: the associated risk management should be commensurate with the level of risk of the function that the AI supports — the intensity of governance depends on risk, not on "whether it counts as a model".

Third, this direction has the lowest regression-validation cost, so it should be done first. Historical SAR conclusions, historical non-performing accounts and historical credit approval results naturally form a Golden Dataset. Among the five directions, Finance is best suited as the demonstration scenario for the Risk & Compliance group's "evaluation and observability layer (L5)".

What warrants caution: the two challenges of model homogeneity risk (noted by the PBOC's Tech Department in 2025-12) and the inherent flaws of algorithmic architectures being hard to eliminate will not disappear because of Harness construction. What Harness can do is make flaws discoverable, localizable and compensable, not eliminate them.

Information Gap Statement

1. The original document number, publication date and article numbers of the Guidelines on Science-and-Technology Ethics in the Financial Sector have not been obtained (only a PBOC official's speech quote and media reprints), so this document cites only its principled formulations such as "minimum necessary collection" and "dedicated use for dedicated purposes".
2. The industry-standard numbers (JR/T numbers) and provisions of the PBOC's Evaluation Specification for AI Algorithm Financial Applications (2021) and Information Disclosure Guidelines for AI Algorithm Financial Applications (2023) have not been obtained, and this document does not cite their articles.
3. The official PDFs and per-article numbers of SR 11-7 and OCC Bulletin 2011-12 have not been obtained; the three-pillar, three-validation-element and compensating-control content cited here comes from BPI and KPMG compilations (Tiers B/C, consistent across multiple sources).
4. Most of the quantitative effect indicators for China's banks (apart from a few disclosures such as Postal Savings Bank's "manual screening efficiency up 30%") come from brokerage industry research and have not been cross-checked verbatim against the banks' annual report originals.
5. Two statistical bases exist for brokerage IT spending: the Securities Times basis is 34 listed brokers' total 2025 information technology expenditure of ¥27.59 billion, 6.2% of total revenue; the China Fund News basis is 28 brokers totaling about ¥25 billion, 6% of revenue. This document adopts the former; the two must not be mixed.
6. Two bases exist for China Merchants Bank's AI scenario count — 856 and 184 — and this document presents them side by side without selecting one.
7. No verifiable Tier-A data has been obtained for China-side "live-trading performance of AI quantitative investment research" in this direction; the relevant metrics are all [To be filled].
8. No Tier-A public data has been obtained for China-side "AI financial statement analysis accuracy" in this direction.

6. References

1. Implementation Plan for High-Quality Development of Digital Finance in the Banking and Insurance Industry, General Office of the National Financial Regulatory Administration, 2025-12. https://www.nfra.gov.cn/cn/view/pages/ItemDetail.html?docId=1239741
2. Report on the speech by Li Wei, Director of the PBOC's Technology Department, at the 2025 annual meeting of the China Wealth Management 50 Forum, China Securities Journal, 2025-12. https://www.cs.com.cn/xwzx/hg/202512/t20251229_6530683.html
3. HKMA, Supporting Adoption of Artificial Intelligence in Fighting Financial Crime, Hong Kong Monetary Authority, 2026-06-22. https://brdr.hkma.gov.hk/eng/doc-ldg/current/20260622-1-EN
4. Navigating Artificial Intelligence in Banking, Bank Policy Institute (with references to SR 11-7 and the OCC Comptroller's Handbook), 2024. https://bpi.com/wp-content/uploads/2024/04/Navigating-Artificial-Intelligence-in-Banking.pdf
5. AI and model risk slipsheet, KPMG US, 2024. https://kpmg.com/kpmg-us/content/dam/kpmg/pdf/2024/1a-ai-and-model-risk-slipsheet.pdf
6. Report on the HSBC × Google Cloud AML AI case, Process Excellence Network. https://www.processexcellencenetwork.com/ai/articles/how-hsbc-turned-its-biggest-compliance-headache-into-an-ai-success-story
7. Comprehensive report on large-scale AI deployment by listed banks, Xinhua Net, 2025-09-10. https://www.xinhua.org/20250910/4e421fdeeb2242fc97299d73b2d62f09/c.html
8. Review of AI and risk-control scenarios in banks' annual reports, Weiyang Net, 2025. https://www.weiyangx.com/462780.html
9. Brokerage information technology investment statistics, Securities Times, 2026-04. https://www.stcn.com/article/detail/3731822.html
10. Report on broker AI applications and IT investment, 21st Century Business Herald, 2026-04-27. https://www.21jingji.com/article/20260427/herald/b0e4a124fc6808e55600d6fcd453da08.html
11. Interpretation related to the Guidelines on Science-and-Technology Ethics in the Financial Sector, China Financial Information Network, 2025-11-17. https://m.cnfin.com/wx/share?url=//m.cnfin.com/hb-lb//zixun/20251117/4336062_1.html
12. Interim Measures for the Administration of Generative AI Services, Cyberspace Administration of China and six other departments, 2023. https://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm