Compliance 合规

1. Introduction

1.1 Background

The essence of compliance work is to translate external regulatory requirements into routine, day-to-day actions inside an institution that are executable, checkable, and provable. This translation process has long faced three structural difficulties: rules are many and changing (laws, regulations, departmental rules, regulatory documents, industry standards, and self-regulatory rules stack in multiple layers and are frequently updated), evidence is scattered (customer identity data, transaction records, and approval trails are spread across multiple systems), and responsibility is real (penalties reach down to specific positions and individuals).

AI changes how the first two difficulties are solved, but does nothing about the third — on the contrary, AI's involvement makes the third more prominent. Because when a submitted document is generated by a model, the question of "who is responsible for the authenticity of that document" becomes harder to answer than ever.

From a regulatory perspective, 2025 to 2026 is a period in which the intensity of financial compliance in China has risen significantly:

• The newly revised Anti-Money Laundering Law of the People's Republic of China took effect on January 1, 2025, which for the first time specifies hard requirements for core obligations such as customer due diligence, retention of transaction records, and reporting of suspicious transactions, extends the customer-data retention period to at least 5 years, and details the risk-based supervisory principle (per industry interpretations reported by China Business Journal; the specific provisions are subject to the original text of the law).
• The Measures for Financial Institutions' Customer Due Diligence and the Management of Customer Identity Data and Transaction Records were issued by the People's Bank of China, the National Financial Regulatory Administration, and the China Securities Regulatory Commission on November 28, 2025, and formally took effect on January 1, 2026, upgrading the previous "customer identity identification" to "customer due diligence."
• Penalty intensity has risen markedly. According to data from Enterprise Early Warning (企业预警通), relayed by China Business Journal, in 2025 banks and their employees received 6,521 fines totaling RMB 2.641 billion, up 44.95% from 2024; among these, fines related to AML business violations grew 185.92% from 2024.

Against this backdrop, AI applications in the Compliance direction have two clear value orientations: first, expanding coverage (turning sampling checks into full-coverage screening), and second, maintaining consistency of specification (avoiding drift in how the same rule is understood across time and across people). There is also a clear red line: the model must not become the reporting subject.

1.2 Definition

The AI Harness for the Compliance direction is an engineered operational layer that carries compliance tasks such as regulatory reporting, anti-money laundering (AML), customer identity identification and due diligence (KYC/CDD), and internal control. Its core responsibility is not to "judge whether something is compliant," but to prepare the facts, rules, and evidence needed for a compliance judgment completely, accurately, and traceably, and to keep irreversible actions firmly in human hands.

This direction covers four categories of tasks:

Key boundary: all outputs in the Compliance direction are materials pending review. Submitting reports to regulators, changing the outward reporting specification, reporting suspicious transactions, and making final determinations of customer risk levels must all be completed by named positions.

1.3 Position in the AI Harness System

图 1-1｜合规方向六层定位模型：瓶颈在 L2，L6 为主约束层

数据来源：基于本文分析绘制的示意图。

The bottleneck is at L2.

This is the most distinctive feature that sets the Compliance direction apart from others. In directions such as software engineering and content creation, the resistance to opening the tool layer comes mainly from technical implementation; in the Compliance direction, the first resistance to opening the tool layer is attribution of responsibility.

The reason is that "dual penalties" have become the norm. As reported by China Business Journal and Securities Times, most of the fines publicly disclosed by the central bank and the Financial Regulatory Administration in 2025 simultaneously publicized penalties against the responsible individuals; both ICBC and Bank of Communications had up to two dozen or more people penalized, and within the year several banking practitioners also received industry-barring penalties. The widespread implementation of "dual penalties" means that the careers of the responsible persons are directly tied to the institution's compliance performance.

Under this responsibility structure, it is not feasible, in terms of accountability, to open irreversible actions such as submission to a model-driven Agent. Hence the L2 design principle for the Compliance direction: read as openly as possible, write under strict tiering, and submission disabled by default.

L6 is the second constraint: audit trails exist not for internal review but to provide evidence in regulatory inspections, penalty defenses, and judicial disputes. This is why logs in the Compliance direction are not "operation logs" but an "evidence chain."

1.4 Value and Limitations

Value:

1. From sampling to full coverage. Traditional compliance checks are limited by headcount and can only sample; AI can run consistency checks across all transactions and all customer files, significantly improving coverage.
2. Consistency of reporting specifications. The same rule is often interpreted differently by different people at different times; AI combined with a versioned rule library keeps the specification stable.
3. Speed of rule-update propagation. After a regulatory rule is updated, AI can quickly locate the affected internal policies and business processes, shortening the propagation cycle.
4. Reducing repetitive work. High-frequency repetitive tasks such as assembling due-diligence materials, validating reporting data, and tracking rectification of returned items can be significantly accelerated.

Limitations (which must be faced):

1. The model does not have complex compliance-judgment capability. Yang Haiping, researcher at the Shanghai Institute of Finance and Law, points out that current suspicious-transaction identification suffers from "weak effectiveness and notably poor timeliness"; existing transaction monitoring and manual review mechanisms are not mature enough, and the ability to identify concealed and complex abnormal fund flows is insufficient to precisely capture new money-laundering risks (as reported by China Business Journal). This describes a difficulty that still exists even after human-machine collaboration — AI has not fundamentally solved it.
2. The institutional adaptation period is itself a source of risk. The Measures for Financial Institutions' Customer Due Diligence and the Management of Customer Identity Data and Transaction Records have upgraded "customer identity identification" to "customer due diligence," requiring a shift from statically "checking the ID card" to dynamically "knowing your customer," with the need to identify beneficial owners along the chain and continuously monitor transaction behavior. According to an analysis by Shi Shuo, assistant researcher at Fudan Development Institute, many banks have not yet adapted to this shift, frequently crossing the line on "failing to report suspicious transactions as required" and "transacting with clients of unidentified identity," while under-investing in anti-money-laundering technology systems and talent (as reported by China Business Journal).
3. The "local knowledge" of compliance specifications is hard to encode. Between the principle-level wording of regulatory rules and the judgment standards used on the front line there is a large body of unwritten experience that is hard to bring into the rule library and hard for models to learn.
4. The introduction of AI itself becomes a new compliance subject. Using AI to process customer identity data raises new compliance issues such as data leaving the domain, automated decision-making, and algorithm explainability; institutions need to build a compliance case for the "manner in which AI is used."

2. Glossary of Terms

3. Case Studies

Effectiveness figures cited in the case studies come from institutional self-reporting or media accounts, not independent third-party verification; the source specification should be noted when citing.

3.1 Compliance Capability Outsourcing for Small and Medium Financial Institutions: Modular Intelligent Compliance Services

3.1.1 Background

Building compliance capability has pronounced economies-of-scale characteristics. Large institutions can build their own anti-money-laundering technology systems and dedicated teams, but small and medium institutions such as regional banks, futures companies, and fund subsidiaries often lack both systems and people. Against a backdrop of rising regulatory penalty intensity, this capability gap translates directly into compliance risk. As reported by China Business Journal, multiple analysts have noted that many banks under-invest in anti-money-laundering technology systems (such as abnormal-transaction monitoring models) and talent, making it hard to effectively identify complex money-laundering patterns.

3.1.2 Approach

According to a review of the 2026 half-year reports compiled by Economic Information Daily, Huachuang Yunxin's SaaS platform "Yunxin Workshop" (云信工坊) already provides intelligent risk control, compliance audit, and anti-money-laundering modeling module services to 213 regional banks, 67 futures companies, and 32 fund subsidiaries. In the same period the company's R&D expense reached RMB 321 million, up 41.8% year-on-year, accounting for 18.3% of revenue; among its R&D staff, those holding machine-learning doctorates accounted for 36% (institutionally self-reported data).

From a Harness perspective, the engineering point of this model is to make compliance capability modular rather than project-based:

• L1: Distill regulations, policies, and historical cases into searchable corpora, loading them trimmed by institution type and business scope.
• L2: Provide capabilities in module form (risk control, compliance audit, anti-money-laundering modeling), with each module's tool permissions approved independently and customer data processed within the tenant-isolation boundary.
• L3: Codify processes such as submission validation, due-diligence material generation, and issue screening into reusable workflows.
• L6: Data isolation and access audit trails in the multi-tenant scenario are the precondition for whether this model can work at all.

3.1.3 Results

Under the disclosure specification of the above half-year reports, the platform serves three types of institutions — regional banks, futures companies, and fund subsidiaries — totaling 312 institutions. This figure is self-reported by the institution and relayed by the media, and no independent third-party verification has been seen.

It should be noted objectively: this model solves the availability of capability problem, not the attribution of responsibility problem — institutions using external intelligent services remain the responsible subjects for compliance.

3.2 Rule-Driven: From "Customer Identity Identification" to "Customer Due Diligence"

3.2.1 Background

On November 28, 2025, the People's Bank of China, the National Financial Regulatory Administration, and the China Securities Regulatory Commission jointly issued the Measures for Financial Institutions' Customer Due Diligence and the Management of Customer Identity Data and Transaction Records, which took effect on January 1, 2026. This measure is not a simple term substitution but a substantive upgrade of the content of obligations.

According to the analysis of Shi Shuo, assistant researcher at Fudan Development Institute, in an interview with China Business Journal, the Measures upgrade the previous "customer identity identification" to "customer due diligence," requiring banks to move from statically "checking the ID card" to dynamically "knowing your customer," with the need to identify beneficial owners along the chain and continuously monitor transaction behavior.

3.2.2 Approach

This institutional change places clear requirements on Harness design:

The engineering implementation path is: first establish a knowledge representation for tracing beneficial owners (shareholders, ownership percentages, contractual control, ultimate natural persons), then establish trigger mechanisms for continuous monitoring (transaction behavior deviating from baseline, equity changes, negative-news hits), and finally push monitoring results to human review in the form of "leads" rather than "conclusions."

3.2.3 Results

No official unified quantitative assessment of the effects of this institutional adjustment has yet been published, but directional signals can be observed from the regulatory specification. As reported by China Business Journal, the central bank's 2026 financial stability work conference stressed deepening and solidifying technology-empowerment efforts and strengthening financial risk monitoring, assessment, early warning, and early correction; based on this, Shi Shuo analyzed that regulators are no longer satisfied merely with whether banks have policies, but with whether those policies operate effectively, and that regulatory tools will also raise the level of digitalization to further enhance regulatory penetration.

Direct implication for institutions: the mere existence of a compliance policy no longer constitutes grounds for exemption; the ability to provide evidence that "the policy operates effectively" becomes the new core of compliance. This is precisely the value of L4 (retraceable state records) and L5 (quantifiable operational metrics) in Harness.

3.3 Failures and Risks: Surging AML Penalties and the Penetration of Dual Penalties

3.3.1 Background

From 2025 to 2026, anti-money laundering became one of the areas of banking with the most concentrated penalties in China. This is not incidental, but the result of a systemic upgrade of the institutional foundation, regulatory technology, and enforcement model.

3.3.2 Approach (the Regulatory Side's Upgrade Path)

According to an analysis reported by China Business Journal, this round of regulatory upgrading is driven by three factors:

1. A more precise ruler: in preparation for a new round of international mutual evaluations on anti-money laundering, relevant regulations were aligned with international standards, suspicious-transaction reporting standards became more precise, retention periods were significantly extended, and "gray areas that may have been tolerated in the past now clearly constitute violations."
2. Deep application of regulatory technology: off-site and penetration-style monitoring become the norm.
3. Enforcement model transformation: dual penalties move from principle to practice; major fines target not only institutions but also penetrate to directly responsible management and even operational staff.

3.3.3 Results (Penalty Facts That Have Occurred)

The following are specific penalty cases publicly reported by the media (as reported by China Business Journal and Securities Times):

Most of the above institutions have responded that the relevant issues occurred in 2023 or earlier and have largely completed rectification (as reported by Securities Times).

Lessons from this case study:

1. The fine amount is not the biggest cost; accountability reaching individuals is. Dual penalties tie compliance responsibility to individual careers, fundamentally changing the incentive structure inside institutions and underlying why it is so difficult to open up Agent tool permissions in the compliance direction.
2. "Having a policy" is no longer enough. Regulatory attention has shifted from "whether a policy exists" to "whether the policy operates effectively," requiring institutions to provide evidence of operation — precisely what Harness's evidence chain and audit-trail capabilities provide.
3. AI is not the cure for this problem. AI can improve coverage and consistency, but if the underlying data is wrong and the specification is messy, AI will only amplify and accelerate the errors.

4. Practice Standards

4.1 AGENTS.md Specification

The following is the proposed full text of AGENTS.md for the Compliance direction. This is a draft proposal; there is no official industry standard. It may be copied directly and trimmed to fit each institution's actual situation.

# AGENTS.md —— Compliance 合规

> 本文为建议稿，业界无官方标准。引用法规为真实规范文本，落地方式为工程建议。

## 角色与边界
- 角色：规则检索者、材料准备者、数据校验者、疑点发现者、退回项整改跟踪者。
- 不是：报送人、合规审批人、客户风险等级终判人、可疑交易上报人。
- 不得：自主向监管机构提交任何报送；自主变更报送口径；自主完成可疑交易定性或上报。
- 边界判定：凡传统工作流中需双人复核、分级审批、加盖公章的动作，智能化后仍需同等人工程序。

## 环境假设
- 存在权威法规库与内部制度库，可返回效力层级与生效日期。
- 存在名单筛查服务与交易监测模型，可返回命中结果与命中理由。
- 存在报送系统，但提交接口默认对 Agent 关闭。
- 存在不可变日志存储与具名主体身份服务；日志保存期不短于法定要求。

## 上下文加载顺序（Context Budget）
1. 任务契约（报送期次/尽调对象、数据分级、允许工具、确认点、报送口径版本）。
2. 适用规则（法律法规、部门规章、规范性文件、内部制度），必须携带**效力层级与生效日期**。
3. 报送口径定义（本次任务所适用的口径版本，不得混用版本）。
4. 历史同类结论与历史退回原因（保持口径一致性）。
5. 证据材料（客户档案、交易记录），按数据分级脱敏后加载。
6. 补充资料（预算不足时最先裁剪）。
硬约束：第 1~4 类不可裁剪；已废止或未生效条款禁止作为依据加载。

## 工具契约
| 工具 | 风险等级 | 说明 |
|---|---|---|
| 法规/制度检索 | R0 | 只读，必须返回效力层级与生效日期 |
| 名单筛查 | R0 | 只读，须返回命中理由与名单版本 |
| 交易监测模型查询 | R0 | 只读，须返回模型版本与阈值 |
| 报送数据校验 | R1 | 输出校验报告，不写入报送系统 |
| 尽调材料草稿生成 | R1 | 写入草稿区，可回滚 |
| 客户档案读取 | R2 | 受控，须有任务关联授权记录 |
| 报送系统提交/撤回接口 | R3 | **默认关闭**，需审批令牌 + 双人复核 |

## 任务执行流程（SOP）
1. 定级：识别任务类型、数据分级、是否涉及不可逆动作。
2. 定规：锁定适用规则的效力层级、生效日期与报送口径版本，全程不得混用版本。
3. 取数：按上下文顺序加载，客户身份资料脱敏后使用。
4. 生成：产出校验报告、尽调材料或疑点清单，每条结论附来源与生效日期。
5. 自校验：法规与制度引用逐条回源核验；未通过核验的引用直接删除。
6. 双人复核：制单与复核分离，复核人须为具名岗位人员。
7. 提交：仅由授权岗位在人工界面完成提交，Agent 不参与提交动作本身。
8. 归档：输入快照、规则版本、口径版本、复核记录、提交回执一并留痕。

## 验证与证据要求
- 法规与制度引用 100% 可溯源，且携带效力层级与生效日期。
- 报送数据校验须覆盖完整性、格式、勾稽关系与跨期一致性四类检查。
- 筛查类任务以召回优先：宁可多报被人工排除，不可漏报。
- 客户风险等级仅可输出初判建议，终判必须由人工完成并记录理由。
- 所有对外材料须可还原到生成时的规则版本与数据快照。

## 失败与升级策略
- 检索不到适用规则 → 输出"未取得结论"，不得以非权威源替代。
- 规则版本或口径版本不明确 → 停止任务，交合规负责人确认版本。
- 引用核验不通过 → 删除该引用及依赖它的结论；累计触发阈值上报合规负责人。
- 触发 R3 工具调用尝试 → 拦截并告警至合规与 IT 安全负责人。
- 同一客户前后风险等级初判不一致 → 冻结自动流程，交人工复核。

## 安全与合规红线
- 严格遵守《人工智能生成合成内容标识办法》（2025-09-01 施行）：生成内容须保留显式标识与元数据隐式标识；依法留存相关日志不少于六个月（第 9 条）；不得恶意删除、篡改、伪造、隐匿标识（第 10 条）。
- 不得自主提交、撤回或变更任何监管报送。
- 不得自主完成可疑交易定性或上报。
- 客户身份资料与交易记录的处理遵循最小必要原则，保存期限不短于法定要求。
- 禁止编造法规条文、规章编号、报送口径、统计数据与来源。
- 禁止以模型记忆提供法规条文内容，必须回源核验。

## 禁止事项
1. 禁止向监管系统提交任何数据或材料。
2. 禁止自主变更报送口径或统计方法。
3. 禁止自主判定客户风险等级终值。
4. 禁止自主完成可疑交易上报或撤回报送。
5. 禁止混用不同版本的规则或报送口径而不显式标注。
6. 禁止引用已废止或尚未生效的条款作为依据。
7. 禁止删除或弱化生成合成内容标识与审计日志。
8. 禁止以"AI 生成"为由免除人工复核义务。
9. 禁止越权读取与任务无关的客户档案或交易明细。
10. 禁止把本文件的建议表述为监管的强制要求。

## 输出格式
- 结论：明确、可判定；无法得出结论时写明"未取得结论"及原因。
- 依据：表格化，含来源类型、来源标识、**效力层级**、**生效日期**、条款或片段位置、核验状态。
- 不确定性：列出影响可靠性的事项与建议处置。
- 留痕：任务标识、规则与口径版本、模型与提示词版本、制单人、复核人及时间、产物哈希。

## 评估与自检
- [ ] 所有法规与制度引用可溯源且携带效力层级与生效日期
- [ ] 报送口径版本已锁定且未混用
- [ ] 筛查类任务召回优先，关键事实召回率达标
- [ ] 未执行任何 R3 动作；报送提交由人工完成
- [ ] 双人复核已完成，制单人与复核人分离
- [ ] 生成合成内容标识已保留，日志已写入只追加存储
- [ ] 未编造任何条文、编号、口径或统计数据

4.2 SKILL.md Specification

The following is the proposed full text of SKILL.md for the Compliance direction. This is a draft proposal; there is no official industry standard.

---
name: compliance-cdd-material-preparation
description: 客户尽职调查材料准备与受益所有人穿透梳理。整理客户身份资料、还原股权链路、标注受益所有人识别疑点，输出待人工研判的尽调材料包。适用于对公客户准入、存量客户定期重检、高风险客户强化尽调场景。
version: 1.0
created: 2026-09-12
---

# 客户尽职调查材料准备与受益所有人穿透梳理

## 适用场景
- 适用：对公客户准入材料整理、存量客户定期重检、股权链路还原、受益所有人识别疑点标注、持续监测线索归集。
- 不适用：客户风险等级终判、自主决定建立或终止业务关系、自主上报可疑交易。

## 前置条件
- 已明确尽调对象（客户编号、统一社会信用代码）与尽调类型（准入/定期重检/强化尽调）。
- 工商与股权数据源、内部客户档案系统可用。
- 名单筛查服务可用，且能返回命中理由与名单版本。
- 尽调草稿区可写入（R1），且复核人角色已指定。

## 输入
| 输入项 | 必填 | 说明 |
|---|---|---|
| 尽调对象标识 | 是 | 客户名称、客户编号、统一社会信用代码 |
| 尽调类型 | 是 | 准入 / 定期重检 / 强化尽调，决定材料深度 |
| 适用规则版本 | 是 | 《管理办法》及内部制度版本，含生效日期 |
| 数据分级判定 | 是 | 客户身份资料通常为敏感级，决定部署形态 |
| 历史尽调记录 | 否 | 用于口径一致性与变更识别 |

## 输出
| 输出项 | 说明 |
|---|---|
| 尽调材料清单 | 应获取材料项、已获取项、缺失项及缺失原因 |
| 股权链路图 | 股东—出资比例—最终自然人链路，标注穿透断点 |
| 受益所有人初判 | 候选自然人、识别依据、置信度、疑点说明 |
| 名单筛查结果 | 命中项、命中理由、名单版本、同名异人排查建议 |
| 疑点清单 | 需人工研判的事项，含疑点类型与建议核实路径 |
| 留痕信息 | 任务标识、规则版本、数据源与查询时点、模型版本、制单与复核人 |

## 执行步骤
1. 锁定适用规则版本与生效日期，明确本次尽调的法定要求边界。
2. 获取客户基础身份资料，脱敏后进入上下文（直接标识符与组合准标识符一并处理）。
3. 还原股权链路：逐层获取股东信息与出资比例，直至自然人或穿透断点。
4. 识别受益所有人候选：按规则阈值与协议控制关系综合初判，标注置信度与疑点。
5. 执行名单筛查：对客户、股东、受益所有人候选、法定代表人执行筛查，记录命中理由与名单版本。
6. 归集持续监测线索：交易行为偏离、股权变更、负面舆情命中等。
7. 自校验：所有外部数据回源核对；无法核对的一律删除而非标注"待核实"。
8. 输出材料包并转人工研判，标注需重点核实项。
9. 归档留痕，研判结论回流为评估样本。

## 质量标准（DoD）
- 一票否决：自主判定客户风险等级终值；自主上报或决定建立业务关系；使用了未脱敏的敏感数据；引用了已废止或未生效条款。
- 引用覆盖率 100%，引用核验通过率 ≥ 99%。
- 股权链路穿透率：可穿透至自然人的比例须达机构设定阈值；未穿透项必须逐一说明断点原因。
- 疑点清单不得为空时静默通过；无发现须显式声明"未发现疑点"及检查范围。
- 名单筛查须包含同名异人排查建议，不得仅凭名称匹配下结论。

## 常见失败与处理
| 失败模式 | 表现 | 处置 |
|---|---|---|
| 穿透断点 | 股权链路在某层无法继续，无法定位最终自然人 | 显式标注断点与原因，转人工；不得以"无受益所有人"结案 |
| 条款时效错误 | 引用已废止办法的条款 | 强制携带生效日期；废止条款禁止加载 |
| 同名异人 | 名单筛查因同名误命中 | 输出排查建议（证件号、地址、成立日期比对），不得直接定性 |
| 协议控制遗漏 | 只看股权比例，忽略协议控制与实际控制 | 检索协议控制与一致行动安排线索并标注 |
| 材料缺失静默 | 缺失材料未列出，看不出缺口 | 缺失项必须显式列出并说明原因 |
| 数据出域 | 客户身份资料被发往不匹配的模型服务 | 分级判定前置；敏感数据仅私有化/本地推理 |

## 示例
输入：
- 尽调对象：客户 B，统一社会信用代码 [以实际为准]
- 尽调类型：定期重检
- 适用规则：《金融机构客户尽职调查和客户身份资料及交易记录保存管理办法》（2026-01-01 实施）+ 内部制度 vX.X
- 数据分级：敏感，仅私有化部署

输出（节选）：
| 项目 | 内容 |
|---|---|
| 已获取材料 | 营业执照、法定代表人证件、章程、股东名册 |
| 缺失材料 | 受益所有人声明书（客户未提交，已列入催收） |
| 股权链路 | 自然人甲 60% → 公司乙 40%；穿透层级 2 层，无断点 |
| 受益所有人初判 | 自然人甲（持股 60%，超阈值）；置信度 高；疑点：需核实是否存在代持安排 |
| 名单筛查 | 命中 0 项；名单版本 [以检索结果为准]；已做同名异人排查 |
| 疑点清单 | 1 项：需核实是否存在代持或一致行动安排 |

> 说明：上表中标识与比例为示例格式，实际必须以数据源返回结果填充，不得沿用示例值。

4.3 Implementation Checklist

5. Summary

The AI Harness for the Compliance direction is, in essence, an engineering system that makes "the effective operation of policies" provable.

Three conclusions:

First, the bottleneck in the Compliance direction is not model capability but tool permissions. Once dual penalties push responsibility down to individuals, no model-driven irreversible action can find a home in terms of accountability. Hence the correct posture in the Compliance direction is "open reading, tiered writing, closed submission" — concentrating AI's capabilities on material preparation and issue discovery rather than replacing the human judgment at the end of the process.

Second, version management is the lifeline of the Compliance direction. Regulations go effective and are repealed, reporting specifications have version iterations, and internal policies are amended. Errors caused by mixing versions cannot be explained away as a "technical oversight" in a regulatory inspection. Therefore rule citations must carry the legal hierarchy and effective date, and submissions must lock a single specification version — this is not engineering pedantry but a compliance baseline.

Third, recall takes priority over precision. In tasks such as suspicious-transaction identification, watchlist screening, and control-point compliance checks, the cost of under-reporting far exceeds the cost of over-reporting. Metrics should tolerate high recall and use human review for precision, rather than pursuing a flattering single-point accuracy.

A limitation that must be recognized: the penalty data cited in this direction comes from third-party statistics relayed by the media, with the specification set by the original publisher; some of the specific violations in the cases are regulatory determinations, and the institutions have responded that the relevant issues occurred in 2023 or earlier and have largely completed rectification. This article cites them only to illustrate the regulatory landscape and does not constitute an assessment of any institution's compliance status. In addition, AI cannot solve the fundamental difficulty of "insufficient ability to identify concealed and complex abnormal fund transactions" — it improves coverage and consistency, not judgment.

Information Gap Statement

For the data below, no reliable secondary source could be obtained while writing this direction, and it has been annotated and processed per the convention, pending supplementation:

1. The full text, article by article, of the Measures for Financial Institutions' Customer Due Diligence and the Management of Customer Identity Data and Transaction Records: the provisions cited here are media accounts and industry interpretations; the specific text is subject to the official released version, marked as [To be verified].
2. The location of the provision in the newly revised Anti-Money Laundering Law of the People's Republic of China stating "customer data retention period of at least 5 years": this expression comes from an industry interpretation relayed by the media; the specific article number has not been located, marked as [To be verified].
3. The effectiveness data of the 312 institutions served by Huachuang Yunxin's "Yunxin Workshop": only the self-reported count of served institutions is available; no compliance-quality metrics (such as improvements in under-reporting or rejection rates) have been disclosed, marked as [To be filled].
4. The specific application of regulatory technology (SupTech): the relevant statements from the central bank's 2026 financial stability work conference are media accounts; the original text of the meeting has not been obtained, marked as [To be verified].
5. The specific numbers and cases of "responsible persons barred from the industry" in AML penalties: the media gave only general descriptions without case-by-case disclosure, marked as [To be filled].

6. References

1. Banks fined 2.6 billion in 2025! Fines for this business surged 185% — China Business Journal, 2026. https://news.qq.com/rain/a/20260106A03BU300
2. 586 institutions fined 576 million! Q1 banking AML and data-security oversight tightens — China Business Journal, 2026. https://new.qq.com/rain/a/20260331A0872S00
3. High-pressure AML! Multiple banking institutions penalized for violations — China Business Journal, 2026. https://big5.cctv.com/gate/big5/finance.cctv.cn/2026/05/14/ARTIW9Ahh5SiqIeV4yxLDFRI260514.shtml
4. Heavy blows! Nearly 30 ten-million-yuan fines! A roundup of 2025 banking penalties — Securities Times, 2026. https://stcn.com/article/detail/3568591.html
5. AML regulatory measures stepped up again! — Securities Times. https://egs.stcn.com/news/detail/2127567.html
6. Self-developed large models become a focus of competition among brokerages — Economic Information Daily, 2026. https://k.sina.com.cn/article_5953189932_162d6782c06704xss6.html
7. Notice on Issuing the Measures for the Identification of AI-Generated Synthetic Content (Guoxinbantongzi No. 2 of 2025) — Cyberspace Administration of China and three other departments, 2025. https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
8. CAC deploys a special action to "Clean Cyberspace: Rectify Abuse of AI Technology" — Office of the Central Cyberspace Affairs Commission, 2025. https://www.cac.gov.cn/2025-04/30/c_1747719097461951.htm
9. Enabling AI adoption with ISO 42001 — Deloitte, 2024. https://www.deloitte.com/us/en/services/consulting/articles/iso-42001-standard-ai-governance-risk-management.html
10. ISO/IEC 42001: The Complete Guide to the AI Management System Standard — Konfirmity, 2026. https://www.konfirmity.com/blog/iso-42001
11. From Compliance Project to Operating Model: Building an AI Governance System with ISO/IEC 42001 — McKenna Consultants, 2025. https://www.mckennaconsultants.com/from-compliance-project-to-operating-model-building-an-ai-governance-system-with-iso-iec-42001
12. Secure Software Development Practices for Generative AI and Dual-Use Foundation Models (NIST SP 800-218A) — NIST CSRC, 2024. https://csrc.nist.gov/pubs/sp/800/218/a/final