王峰 AI × 经营 约一次诊断

总目 / 文章 / WF-0355

文章 · 存档

深度解读《会计师事务所反洗钱工作管理办法》——把风控装进业务流程

财政部和中国人民银行在2025年9月3日联合发布了《会计师事务所反洗钱工作管理办法》。

这部管理办法的发布预示着,中国反洗钱监管体系正在从“金融机构主导”转向“全链条中介参与”

过去十几年,反洗钱(AML)主要集中在银行、证券、保险等金融机构,因为它们直接接触资金流。

但近年国际趋势明显变化——资金流与信息流脱钩,中介机构(审计、税务、评估、信托、律师等)成为洗钱活动的重要“前端”与“隐蔽环节”。

事务所在客户尽职调查(CDD)、资金核查、审计报告、跨境结构设计等环节拥有穿透信息,若不纳入监管体系,就会成为洗钱与逃税的盲区。

所以这不是一份“可选合规”,而是一份把反洗钱(AML/CFT)要求嵌入执业全流程的“基础设施型规则”。

它改变的是事务所的组织架构、客户管理、技术栈、证据链与监督关系五件事。

一、监管坐标:从“合规任务”升级为“治理工程”

《办法》明确财政部门统筹、人民银行协同、行业协会自律的三层治理结构,并要求按照风险为本(RBA)实施检查与自律管理。

含义很直接:以后监督不是“一刀切抽查”,而是看你是否有一套与自身风险相称的制度+流程+记录证据。

这对不同规模与业务结构的所,意味着因业制宜、差异化控制成为新常态。

二、义务清单:从“知道客户”到“持续了解客户”

《办法》把事务所的反洗钱义务梳成一条可执行的链路:

- 风险评估:新技术/新业务上线前必须做风险评估并设缓释措施;

- 客户尽职调查(CDD):建立业务关系前完成身份识别、目的性质、资金/资产来源等核验;持续尽调贯穿整个业务周期;

- 强化尽调(EDD):高风险国家/地区、政要及其关联、被调查名单等情形,须加密审查、提频监测,并需经高管批准;

- 可疑交易报告(STR):不论金额大小,怀疑即报,由行业统一监管平台汇总后报送;

- 记录保存:客户身份与业务记录在业务终止后至少保存十年;

- 特别预防措施:对列示对象立即停止提供服务并履行报告义务。

这条链路把“客户-业务-记录-报告”的每一步都标准化、留痕化和可追溯化。

三、组织与分工:把“不相容分离”落到权限、账号与证据

《办法》虽以义务为主,但其内在要求事务所把不相容岗位分离、授权审批、内部审计与检查做成“可被验证”的制度与记录:谁评估了风险、谁批准了高风险客户、谁复核了EDD证据、谁提交了STR,都要在系统中有链路。

这会倒逼事务所完成三件事:

1. 职责图谱:合规负责人/业务合伙人/项目经理/质控复核的“权责矩阵”;

2. 流程固化:把CDD/EDD/STR“卡点”配置为系统节点(非纸面制度);

3. 证据闭环:把身份证明、受益所有人(UBO)穿透材料、名单筛查结果、访谈纪要、审批截图,纳入统一的案件卷宗。

四、与AI的“正确连接”:既借力、也设“护栏”

1)AI可以显著提升三类能力

- 身份识别与穿透:OCR读取证件、知识图谱穿透UBO、名单与制裁库自动筛查与复核队列;

- 持续尽调与预警:Agent定期抓取公开负面信息、舆情与司法文书,触发“风险等级上调→EDD复审→高管再批”;

- 可疑点生成与报告辅助:大模型将访谈、函证、资金轨迹等非结构化信息自动形成可疑要点与STR草稿(仍需人工把关)。

2)但必须有三道“护栏”

- 数据合规:客户信息与项目材料不得输入公共大模型,采用企业私域模型/专用合规API,并与日志、访问控制对齐;

- 人类监督:任何AI输出的EDD结论、名单比对结果与STR草稿必须由项目合伙人/合规负责人复核签名;

- 模型治理:建立提示词与输出留痕、红队对抗测试、偏差与幻觉案例库,纳入年度内部审计范围。

换言之,AI是放大器,不是“免责符”。合规与职业怀疑仍是不可外包的核心价值。

五、对业务版图的影响:四类“高敏场景”要前置评估

在审计、税务、咨询与验资等板块,以下四类场景要特别关注:

1. 涉及大量现金/高频资金划转的行业客户(如贸易、工程分包、贵金属、虚拟资产相关服务商),初始风险分值应上调,EDD成为--“默认开启”--;

2. 跨境结构复杂、SPV层级多的委托事项,要把UBO穿透、资金真实用途与来源证据做成清单化模板;

3. 非常规服务(如配合上市地尽调、特殊税务筹划、被执法/监察调查对象关联服务),需项目立项即同步合规会签与高管批准;

4. 新技术/新类型业务(含数据算法评估、加密技术审计等)上线前,必须完成场景级风险评估与缓释方案。这些要求与《办法》对“新技术/新业务先评估后开展”的条款一一对应。

六、与行业自律的“新关系”:协会不是旁观者

《办法》把中注协的角色从“倡议者”升级为自律管理执行者:制定反洗钱工作指引、完善执业准则中反洗钱要求、开展风险提示、实施自律惩戒并向财政部门报告。

预期效果是:行业标准将更快迭代,“以案促改”将成为常态。

对各所的启示是:别等惩戒通知来学习,要主动对标行业指引版本迭代。

七、最令人“长记性”的两条:STR与十年保存

两点最容易被忽视:

- 可疑交易报告(STR)“怀疑即报、金额不设门槛”,并通过行业统一平台报送(可追踪、可量化);

- 客户身份与业务记录“终止后至少保存十年”,这要求你的档案系统、云存储与密钥管理足以承受长周期、可审计与可恢复。

它们既是底线,也是技术与投入的“锚点”。

八、给管理层的三条战略建议

建议一:把反洗钱做成“流程产品”

把CDD/EDD/STR、名单筛查、UBO穿透、审批与归档,设计为统一的流程产品(SOP+系统节点+审计日志),让每一个“该发生的动作”都在系统里“被迫发生”。

建议二:用AI做“效率×合规”的双提升

用私域大模型+规则引擎承接证据收集、文档摘要、名单比对、负面舆情抓取;把“EDD复核→高管批准→卷宗固化”做成自动工作流。但记住:AI前有制度,后有审计。

建议三:把“风险为本”刻进绩效

把高风险客户识别率、EDD复核及时率、STR质量与退件率、十年保全合规率、AI使用合规通过率,纳入部门KPI与合伙人绩效。让制度变成行为,行为变成文化。

/ 60 / 90 天落地清单(可直接执行)

天:搭骨架

- 发布《反洗钱合规责任书》与权责矩阵;

- 上线客户/业务风险评级表(行业×地域×结构×资金来源×媒体舆情五维度);

- 采购或接入名单筛查与UBO穿透工具;

- 制定STR编制与复核指引模板。

天:固流程

- 把CDD/EDD/STR审批做成系统化节点,强制留痕;

- 将高风险情形须高管批准配置为系统必经关卡;

- 建立私域模型(或合规API)+日志闭环,禁止客户数据进入公有大模型;

- 上线十年保全的电子卷宗与密钥管理方案。

天:可验证

- 组织一次内部审计/穿透式抽测:随机选10个项目,从身份识别到STR全链路复盘;

- 形成年度风险评估报告与改进计划并向管理层汇报;

- 与协会/监管口径对齐,关注行业指引与案例并更新SOP。

这部《管理办法》真正改变的是事务所的工程观:反洗钱不是“合规部的事”,而是每一条业务流程的事、每一段系统链路的事、每一次判断取舍的事。

当我们把规则做成流程、把流程做进系统、把系统留住证据,再用AI去放大效率与质量,反洗钱就不再是成本,而是事务所品牌与信任的乘数。

Deep look at “Administrative Measures for Anti–Money Laundering Work of Accounting Firms” — put risk control into the workflow

The Ministry of Finance and the People's Bank of China jointly issued the “Administrative Measures for Anti–Money Laundering Work of Accounting Firms” on September 3, 2025.

The release of these measures shows that China's AML supervision is moving from “led by financial institutions” to “full-chain participation by intermediaries.”

Over the past decade, anti–money laundering (AML) focused on banks, securities, and insurance firms, because they deal directly with money flows.

But in recent years the global trend has changed — money flows and information flows are no longer the same, and intermediaries (audit, tax, valuation, trust, lawyers, etc.) have become key “front ends” and “hidden parts” of money laundering.

Firms hold penetrating info in client due diligence (CDD), fund checks, audit reports, and cross-border structure design. If they are not in the oversight system, they become blind spots for money laundering and tax evasion.

So this is not “optional compliance,” but an “infrastructure-type rule” that embeds AML/CFT demands into the whole practice process.

It changes five things for firms: the org chart, client management, tech stack, evidence chain, and oversight ties.

I. Regulatory position: upgrade from a “compliance task” to a “governance project”

The measures set a three-layer governance setup: the Finance authority leads, the PBOC works with it, and the industry body self-regulates. They also ask for risk-based (RBA) checks and self-discipline.

The meaning is clear: supervision will no longer be “one-size-fits-all spot checks,” but will look at whether you have rules + flows + recorded proof that match your own risk.

For firms of different sizes and business mixes, this means “fit-for-purpose” and different controls will be the new normal.

II. Duty list: from “know your customer” to “keep knowing your customer”

The measures line up a workable AML duty chain for firms:

Risk assessment: before any new tech/new business goes live, you must do a risk review and set mitigations;

Customer due diligence (CDD): before a business tie starts, finish ID checks, purpose and nature, and source of funds/assets; keep due diligence going through the full business life;

Enhanced due diligence (EDD): for high-risk countries/regions, politically exposed persons (PEPs) and their ties, watch-list hits, etc., you must tighten review, monitor more often, and get senior approval;

Suspicious transaction report (STR): report on doubt no matter the amount, and send it through the unified industry platform;

Record keeping: keep client ID and business records at least ten years after the business ends;

Special preventive steps: stop service at once for listed targets and meet report duties.

This chain makes each step — “client–business–records–report” — standard, with logs, and traceable.

III. Organization and roles: make “segregation of duties” real in rights, accounts, and proof

Though the measures focus on duties, they in fact ask firms to turn incompatible duties split, approval rights, and internal audit/checks into “verifiable” rules and records: who did the risk review, who okayed a high-risk client, who re-checked EDD proof, who filed the STR — all must have a trail in the system.

This will push firms to finish three things:

Role map: a “rights-and-duties matrix” for the compliance lead/business partner/project manager/quality reviewer;

Process hardening: make CDD/EDD/STR “gates” system nodes (not paper rules);

Evidence loop: put ID proofs, UBO (ultimate beneficial owner) look-through files, list-screen results, interview notes, and approval screenshots into one case file.

IV. The “right link” with AI: use its power, set guardrails too

AI can lift three types of ability a lot

ID check and look-through: OCR reads IDs, knowledge graphs look through UBOs, watch and sanctions lists auto screen and push to review queues;

Ongoing due diligence and alerts: agents pull public bad news, media talk, and court files on a cycle, to trigger “risk level up → EDD re-check → senior re-approval”;

Suspicion points and report help: large models turn interviews, confirms, and fund trails and other unstructured info into suspicion points and STR drafts (still need human review).

But three “guardrails” are a must

Data compliance: do not put client info and project files into public large models; use enterprise private models/compliant special APIs, aligned with logs and access control;

Human oversight: any AI output on EDD results, list match results, and STR drafts must be signed off by the project partner/compliance lead;

Model governance: set prompt/output logs, red-team tests, and a bias/hallucination case base, and include them in the yearly internal audit.

In short, AI is an amplifier, not a “get-out-of-jail card.” Compliance and professional skepticism are still core values you cannot outsource.

V. Impact on the business map: four “high-sensitivity scenes” need up-front review

In audit, tax, advisory, and capital-verification work, watch these four scenes:

Clients in fields with lots of cash or high-frequency fund moves (like trade, project sub-contracting, precious metals, virtual-asset service firms): raise the start risk score, make EDD on by default;

Engagements with complex cross-border setups and many SPV layers: build a checklist template for UBO look-through, true use of funds, and source proofs;

Non-routine services (like due diligence for a listing place, special tax planning, or work tied to a party under law/discipline probe): start the project with compliance co-sign and senior approval in step;

New tech/new type business (including data/algorithm reviews, crypto tech audits, etc.): finish scene-level risk reviews and mitigations before go-live. These match the clause “assess first, then start” for new tech/new business in the measures.

VI. A new tie with industry self-rule: the association is not a bystander

The measures lift CICPA's role from “advocate” to self-regulation executor: set AML work guides, improve practice rules with AML asks, send risk alerts, do self-discipline penalties, and report to the Finance authority.

The expected result: industry rules will update faster, and “fix by case” will be normal.

For firms the hint is: don't wait for a penalty notice to learn — benchmark new guide versions on your own.

VII. Two rules you will remember: STR and ten-year keeping

Two easy-to-miss points:

STRs are “report on suspicion, no amount floor,” and go through the unified industry platform (trackable, countable);

Keep client ID and business records “at least ten years after end.” This asks your archive system, cloud storage, and key management to handle long time, auditability, and recovery.

These are both the bottom line and the “anchors” for tech and spend.

VIII. Three strategy tips for management

Tip 1: Turn AML into a “process product.”

Design CDD/EDD/STR, list screening, UBO look-through, approval, and filing as one process product (SOP + system nodes + audit logs), so every “should happen” step is forced to happen in the system.

Tip 2: Use AI for a “dual lift of efficiency × compliance.”

Use a private large model + rules engine to handle evidence capture, doc summary, list match, and bad-news fetch; make “EDD review → senior approval → file lock-down” an auto workflow. But remember: before AI stands rules, after AI stands audit.

Tip 3: Carve “risk-based” into performance.

Put high-risk client hit rate, EDD on-time review rate, STR quality and return rate, ten-year keep compliance rate, and AI use compliance pass rate into department KPIs and partner performance. Turn rules into acts, and acts into culture.

/ 60 / 90-day rollout checklist (ready to use)

days: build the frame

Issue the “AML compliance letter of duty” and the rights-duty matrix;

Launch the client/business risk score sheet (five lenses: industry × region × structure × fund source × media talk);

Buy or connect list screening and UBO look-through tools;

Make STR drafting and review guide templates.

days: harden the flow

Make CDD/EDD/STR approvals system nodes with forced logs;

Set “high-risk needs senior approval” as a must-pass gate in the system;

Build a private model (or compliant API) + log loop, and ban client data from public large models;

Go live with e-files and key management that can keep for ten years.

days: make it verifiable

Run an internal audit/look-through spot test: pick 10 projects at random and replay the full chain from ID to STR;

Form the yearly risk review report and improvement plan and brief management;

Align with the association/regulator, watch guide updates and cases, and update SOPs.

These measures change the firm's engineering view: AML is not “the compliance department's job,” but the job of every business flow, every system link, and every judgment call.

When we turn rules into flows, flows into systems, systems into evidence, and then use AI to scale speed and quality, AML is no longer a cost — it becomes a multiplier for the firm's brand and trust.