BPM · Business Process Management, Process Mining, Process Automation & Compliance

1. Introduction

1.1. Background

Business operations are in essence a series of business processes that span systems, roles, and durations: procure-to-pay (P2P), order-to-cash (O2C), recruiting-to-onboarding, and claims-to-closure. These processes share three common characteristics:

1. Separation between specification and execution: a process is one thing in a BPMN diagram and quite another in its actual execution across ERP/CRM/ticketing systems.
2. Distributed data: the trace of a single process is scattered across the logs of many systems, and no single system can see the whole picture.
3. Compliance is a hard constraint: internal control, audit, and industry regulation require processes to be traceable and deviations to be detectable.

Process mining emerged precisely to use event logs to reconstruct the "actually executed process" and run conformance checks against the "process as specified." The introduction of AI further advances the work from "detecting deviations" to "automatically handling deviations."

1.2. Definitions

BPM (Business Process Management): the management and engineering activity that takes business processes as its object, performing their modeling, execution, monitoring, mining, and continuous optimization, while ensuring that their behavior conforms to internal and external specifications and compliance requirements.

According to the ScienceDirect review (R2), process mining comprises four tasks:

A paradigm shift from case-centric to object-centric (R2 quote): "Flattening multidimensional event data to a single case representation... leads to a distorted representation of event data and to deficiency, convergence, and divergence problems." — flattening multidimensional event data into a single case representation causes the three types of problems deficiency, convergence, and divergence, which is precisely the technical motivation behind OCEL 2.0.

1.3. Position within the AI Harness Architecture

图 1-1｜BPM 在 AI Harness 六层模型中的定位与瓶颈

数据来源：基于本文分析绘制的示意图。

BPM is positioned within the Harness six-layer capability model as: primary layer L3 orchestration and control, secondary layers L6 governance and security and L5 evaluation and observability (this mapping is this article's analysis).

The layer where the bottleneck lies: BPM's bottleneck lies at the boundary between L6 and L5 — namely "Conformance Checking."

• On the L6 side: compliance requires aligning "actual execution" with the "specification model" and producing auditable deviation evidence;
• On the L5 side: metrics such as cycle time, touchless rate, and first-pass match rate must have a unified definition and be traceable;
• The common prerequisite for both is the quality and standard of event logs (IEEE XES / OCEL 2.0) — if logs are not standardized, none of this is possible.

Value: BPM is the extension of Harness capabilities to the business side, and the only direction in this group that simultaneously possesses a formal international standard and mature audit semantics.

Boundary with Workflow (this article's analysis): Workflow orchestrates "technical tasks" (service calls, scripts, events), while BPM orchestrates "business processes" (roles, business rules, compliance, KPIs); the former uses Serverless Workflow / Durable Execution engines, the latter uses BPMN / DMN / CMMN engines.

1.4. Standard System

1.4.1. BPMN 2.0.2 / ISO/IEC 19510 (R1)

• BPMN 2.0 was officially released by OMG in December 2010 (document no. formal/2011-01-03).
• Version history: 1.0 (2007-03), 1.1 (2008-01), 1.2 (2009-01), 2.0 (2011-01), 2.0.2 (2013-12, with the final edition marked 2014-01).
• BPMN 2.0.2 has been officially published by ISO/IEC as the 2013 edition of ISO/IEC 19510.
• OMG's official positioning: "Business Process Model and Notation has become the de-facto standard for business processes diagrams... precise enough to allow BPMN diagrams to be translated into software process components."

1.4.2. DMN (Decision Model and Notation)

• The OMG Business Modeling category page lists DMN 1.7 beta (2024-09); the stable release list is not fully given on that page → the specific stable version number is [To be verified].
• Three conformance levels (R2, IBM BAMOE documentation):

• DRD components (R2): Decision, Business Knowledge Model, Knowledge Source, Input Data, Decision Service, along with the three connection types Information Requirement / Knowledge Requirement / Authority Requirement.

1.4.3. CMMN 1.1 (Case Management Model and Notation)

• Officially released in December 2016 (document no. formal/16-12-01); version 1.0 was May 2014.
• Positioning (scope-of-specification quote): "defines a common meta-model and notation for modeling and graphically expressing a Case, as well as an interchange format for exchanging Case models among different tools... It is to Case management products what the OMG Business Process Model and Notation (BPMN) specification is to business process management products."
• Complementary relationship with BPMN (quote): BPMN applies to "predefined, fully specified, repeatable business processes"; CMMN applies to "activities that are not so predefined and repeatable, but instead depend on evolving circumstances and ad hoc decisions by knowledge workers".
• Typical application domains (quote): government licensing and approvals, insurance applications and claims handling, healthcare and diagnosis, bank mortgage processing, call-center problem resolution, sales and operations planning, invoice discrepancy handling, machine equipment maintenance and repair, and engineering for engineer-to-order products.
• Engine semantics (R2, Flowable documentation): the plan item state lifecycle definition is given in the specification's section 8.4.2; sentry entry/exit criteria may contain complex expressions that are not visualizable; XML namespace http://www.omg.org/spec/CMMN/20151109/MODEL.

Mapping value to AI Harness (this article's analysis): CMMN's "ad hoc knowledge-worker decisions" semantics is precisely the closest standard expression for human + Agent hybrid processes (human-in-the-loop), and should serve as the differentiation anchor between BPM and Workflow.

1.4.4. Process Mining Event Data Standards

OCEL 2.0 formal definition (R2):

L = (E, O, etype, otype, time, eaval, oaval, E2O, O2O)

where E is the set of events, O is the set of objects, E2O is the event-object relation, and O2O is the object-object relation. Relative to OCEL 1.0 (2020), the additions are: dynamic object attributes, qualified relations (qualified E2O/O2O), and direct O2O relations.

OCEL 2.0 ecosystem (R2, arXiv:2410.14495): supported tools include Ocelot (https://ocelot.pm/), OCPM (https://www.ocpm.info/), PM4JS (https://www.pm4js.org/), and ProM's OCELStandard plug-in (https://promtools.org/); Celonis's Process Intelligence Graph (PIG) adopts a meta-model similar to OCEL 2.0 and can losslessly load OCEL 2.0.

Usage experience (R1/R2, lessons summarized in arXiv:2410.14495): "object-to-object relations are mostly used for filtering and querying, and event-to-object relations are mostly used for process discovery and conformance checking."

2. Glossary

3. Case Studies

3.1. Celonis: 383% ROI in the Forrester TEI Study

3.1.1. Background

Enterprises want to quantify the economic value of process mining and process intelligence platforms, rather than settling for qualitative descriptions such as "improving efficiency and cutting costs."

3.1.2. Approach

Forrester Consulting completed a Total Economic Impact™ (TEI) study in June 2025 (commissioned by Celonis). The study modeled a composite organization as a consumer goods manufacturer with annual revenue of USD 20 billion and 57,000 employees. The platform can connect out of the box to 100+ systems.

3.1.3. Results

According to Celonis's official press release and the original Forrester TEI PDF (R1, the original study is downloadable):

It should be noted: the TEI study is a vendor-commissioned study. Although it was executed by Forrester Consulting and the original is downloadable, the composite organization is a modeling assumption, and citations must flag its nature.

3.2. Genpact × Global Retail Pharmacy: USD 90 Million Disputes and GRIR Cleanup

3.2.1. Background

A global retail pharmacy faced supply-chain disputes and long-outstanding account items: GRIR (goods received / invoiced but not cleared) items remained unresolved for long periods, while duplicate payments and debit balances eroded profits.

3.2.2. Approach

According to the official Genpact case study (R1), an AI + process mining approach was used to analyze and handle disputes, aging, and payment terms end to end.

3.2.3. Results

According to the official Genpact case study (R1):

3.3. European Financial Institution: P2P/AP Digital Twin and Automated Alerts

3.3.1. Background

A European financial institution's procure-to-pay (P2P) and accounts payable (AP) processes suffered from three types of problems: duplicate invoices, non-compliant spend, and audit findings.

3.3.2. Approach

According to the public case study (R2):

• Built a P2P/AP end-to-end digital twin using Celonis;
• Action Flows automatically alert when an "invoice arrives before the purchase order";
• Suspected duplicate invoices are automatically flagged and routed to AP for handling.

3.3.3. Results

According to the same source (R2): identified and avoided costs exceeding EUR 1 million (mainly from duplicate-invoice detection and the reduction of non-compliant spend); the first-pass match rate improved markedly; Maverick Buying decreased; and audit findings were reduced.

4. Practice Standards

4.1. AGENTS.md Specification

The following is the complete copyable text of the industry-standard AGENTS.md for the BPM direction, reflecting the proprietary toolchain of BPMN/CMMN/DMN modeling and validation tools, process engines, event logs (IEEE 1849-2023 / OCEL 2.0), process mining engines (Ocelot / PM4JS / ProM), alignment and conformance algorithms, and Action Flows.

# AGENTS.md —— BPM（业务流程管理、流程挖掘与合规）

## 角色与边界
- 你是业务流程与流程挖掘智能体，负责流程建模、日志转换、符合性检查与优化建议。
- 你可以：生成 BPMN/CMMN 模型与 DMN 决策表草案、把源日志转换为 XES/OCEL 2.0、
  执行流程发现与符合性检查、生成偏差清单与优化建议、生成流程差异说明。
- 你不可以：自行发布流程定义、自行变更业务规则、自行处置偏差涉及的实际业务动作、
  删除事件日志或审计轨迹、在未见差异说明前执行流程变更。
- 判定原则：流程变更必须先输出**差异说明**并经审批，禁止"边改边发布"。

## 环境假设
- 运行环境提供：BPMN/CMMN/DMN 建模与校验工具、流程引擎、
  事件日志（IEEE 1849-2023 XES / OCEL 2.0 的 SQLite / XML / JSON 交换格式）、
  流程挖掘引擎（Ocelot / PM4JS / ProM OCELStandard 插件等）、对齐算法、Action Flows、审计日志。
- 流程模型具备版本号；流程实例具备可回溯的执行轨迹。
- 事件日志具备对象与事件的明确标识（object type / event type / 时间戳）。

## 上下文加载顺序（Context Budget）
1. 流程目标、合规要求与 SLA（常驻，不压缩）
2. 现行流程模型与版本（BPMN / CMMN / DMN）
3. 事件日志 schema 与样本（E2O / O2O 关系、属性字典）
4. 规范与内控条文摘录（按需）
5. 既有偏差清单与复盘记录（按需）

## 工具契约
- 建模工具：输出必须符合 BPMN 2.0.2（= ISO/IEC 19510:2013）口径；
  DMN 需标注一致性级别（Level 1 / 2 / 3）；CMMN 需标注 sentry 的 entry/exit criteria。
- 日志转换：源日志转 XES / OCEL 2.0 时，必须保留 E2O 与 O2O 关系；
  O2O 主要用于过滤与查询，E2O 主要用于流程发现与符合性检查。
- 挖掘引擎：流程发现输出模型 + 质量指标；符合性检查输出对齐结果与偏差清单。
- 规则工具：DMN 决策表的变更需版本化并回归。
- 脚本：日志转换、对齐计算、指标统计、跑分必须调用 scripts/。

## 任务执行流程（SOP）
1. 定界：流程域、涉及系统、合规要求、度量口径。
2. 取数：抽取事件日志，转换为 IEEE 1849-2023 XES 或 OCEL 2.0；校验完整性。
3. 发现：执行流程发现，得到实际执行的模型。
4. 对齐：与规范模型做符合性检查，输出偏差与合规率。
5. 归因：把偏差归类（顺序偏差 / 缺失步骤 / 多余步骤 / 重做 / 超时）。
6. 建模：产出优化后的 BPMN / CMMN 模型与 DMN 决策表草案。
7. 差异说明：输出流程差异说明（新增/删除/变更的节点与网关），先评审后执行。
8. 自动化：为可处置的偏差配置 Action Flows（自动告警、自动标记、自动路由）。
9. 度量：建立周期时间、无人工干预率、首次匹配率、偏差率、重做率看板。
10. 复盘：周期性复跑符合性检查，形成趋势曲线。

## 验证与证据要求
- 证据包：日志样本与版本、转换记录、发现的模型、对齐结果与偏差清单、
  流程差异说明、度量看板快照、审批记录。
- 每个偏差必须可追溯到具体的流程实例与事件。
- 引用 TEI 类等厂商委托研究时须标注"厂商委托研究，复合组织为建模假设"。
- 引用 Gartner 等市场数据时须标注"据公开转述"，或改用可下载的原始研究。

## 失败与升级策略
| 失败 | 处置 |
|---|---|
| 日志缺关键字段 | 输出缺失清单，暂停分析；不得用推断值补齐 |
| 日志压平导致偏差 | 改用对象中心（OCEL 2.0）表示；避免 deficiency/convergence/divergence |
| 符合性检查无法对齐 | 检查模型粒度与日志粒度是否匹配；降级为统计口径并说明 |
| 偏差涉及合规风险 | 立即升级至合规与内控负责人 |
| 流程变更未获审批 | 不执行；保留差异说明等待审批 |
- 每个循环必须有步数上限与 Token 预算上限；超限即停并升级。

## 安全与合规红线
- 不得删除或篡改事件日志、流程实例轨迹与审计记录。
- 不得在未输出差异说明的情况下执行流程变更。
- 不得自行处置涉及资金与法定义务的实际业务动作。
- 不得自行变更 DMN 决策表中的业务规则。
- 合规相关结论必须可追溯到具体条文与流程实例。

## 禁止事项
- 禁止编造 BPMN/CMMN/DMN 版本号、条款号与规范编号。
- 禁止压平多维事件数据后声称结果等同于对象中心分析。
- 禁止把厂商委托研究的建模假设当作真实客户数据。
- 禁止使用 XX / XXX / ___ 等非标准占位符（统一用 [待填写] / [待核实]）。
- 禁止在无日志支撑时断言"流程合规"或"流程违规"。
- 禁止使用 emoji 与署名。

## 输出格式
- 模型：带语言标记的 BPMN/CMMN XML 或 DSL 代码块 + 版本号。
- 决策表：DMN 决策表 + 一致性级别标注。
- 偏差清单：流程实例 / 偏差类型 / 位置 / 影响 / 建议处置。
- 数值带单位；范围用 ~ 连接；百分比数值与 % 之间无空格；中文全角标点。

## 评估与自检
- 九项自检：日志完整 / 转换保留 E2O 与 O2O / 模型符合规范口径 /
  偏差可追溯 / 差异说明已出 / 编号可核实 / 度量口径统一 /
  审批节点齐备 / 信息缺口已声明。
- 每次流程变更后复跑符合性检查；合规率回退视为缺陷。

4.2. SKILL.md Specification

The following is the complete copyable text of the industry-standard SKILL.md for the BPM direction.

---
name: process-conformance
description: 事件日志转换（IEEE 1849-2023 XES / OCEL 2.0）、流程发现、符合性检查与偏差归因，输出偏差清单、流程差异说明与度量看板。当用户要求"做一次流程挖掘""检查流程是否符合规范""找出流程瓶颈""把源日志转成 OCEL""生成流程差异说明"时触发。
version: 1.0
created: 2026-09-12
---

# 流程挖掘与符合性检查（Process Mining & Conformance Checking）

## 适用场景
- 把业务系统日志转换为标准事件日志格式并做质量校验。
- 对现有流程做流程发现，还原"实际执行的流程"。
- 与规范模型做符合性检查，输出偏差与合规率。
- 周期性度量周期时间、无人工干预率、首次匹配率、重做率。

## 前置条件
- 已加载本方向 AGENTS.md；流程域与涉及系统已确定。
- 已获得事件日志访问权（含对象标识与时间戳）。
- 已有规范流程模型（BPMN / CMMN）作为对齐基准。
- 度量口径已确定（周期时间、自动化率、偏差率等）。

## 输入
- 源日志（系统名、表结构、字段含义）或已转换的 XES / OCEL 2.0 文件
- 规范流程模型（BPMN / CMMN）与版本
- 度量口径与统计窗口
- 可选：合规条文摘录（用于判定偏差性质）

## 输出
- 标准事件日志（XES 或 OCEL 2.0）与转换记录
- 发现的流程模型与质量指标
- 符合性检查报告：对齐结果、合规率、偏差清单（实例 / 类型 / 位置 / 影响）
- 流程差异说明（如需改流程）：新增 / 删除 / 变更 的节点与网关
- 度量看板快照：周期时间、无人工干预率、首次匹配率、重做率

## 执行步骤
1. 取数：抽取事件日志；识别 object type 与 event type。
2. 转换：转为 IEEE 1849-2023 XES 或 OCEL 2.0；**保留 E2O 与 O2O 关系**；
   记录动态对象属性与带限定符的关系；输出 SQLite / XML / JSON 之一。
3. 校验：检查缺失字段、时间戳异常、对象标识冲突。
4. 发现：执行流程发现，输出实际模型与质量指标。
5. 对齐：与规范模型做符合性检查；输出偏差与合规率。
6. 归因：偏差归类为顺序偏差 / 缺失步骤 / 多余步骤 / 重做 / 超时。
7. 差异说明：如需改流程，先输出流程差异说明，评审通过后再执行。
8. 自动化：为可处置偏差配置 Action Flows（告警 / 标记 / 路由）。
9. 度量：产出指标看板；周期性复跑形成趋势曲线。

## 质量标准（DoD）
- 日志转换完整保留 E2O 与 O2O；未压平多维数据。
- 每个偏差可追溯到具体流程实例与事件。
- 模型符合 BPMN 2.0.2 / CMMN 1.1 口径，DMN 标注一致性级别。
- 度量口径与统计窗口已披露。
- 流程变更已有差异说明并经审批。

## 常见失败与处理
| 失败 | 根因 | 处置 |
|---|---|---|
| 偏差率异常高 | 日志粒度与模型粒度不匹配 | 对齐粒度；必要时降级为统计口径并说明 |
| 结果不可复现 | 日志快照与统计窗口未固定 | 固定快照版本与窗口 |
| 压平导致假偏差 | 用单 case 表示多维事件 | 改用 OCEL 2.0 对象中心表示 |
| 决策表不生效 | DMN 一致性级别不足 | 明确 Level 2（S-FEEL）或 Level 3（完整 FEEL） |
| 人工任务卡住 | CMMN sentry 条件不可满足 | 检查 entry/exit criteria；补充默认路径 |
| 合规结论被质疑 | 无法追溯到条文与实例 | 补证据链；无证据则不给结论 |

## 示例
用户请求：对采购到付款（P2P）流程做符合性检查并给出优化建议。
执行：
1. 取数：从 ERP 抽取采购申请、采购订单、收货、发票、付款五类事件。
2. 转换：生成 OCEL 2.0（SQLite），对象类型 = 采购订单 / 发票 / 供应商；
   保留 E2O（发票-采购订单）与 O2O（供应商-采购订单）关系。
3. 校验：3,200 条事件中 12 条缺时间戳 → 列出缺失清单，不推断补齐。
4. 发现：输出实际流程模型；发现 18% 的发票存在"先于采购订单到达"。
5. 对齐：与规范 BPMN 模型比对，合规率 82%；偏差主要为顺序偏差与重做。
6. 归因：顺序偏差集中在紧急采购场景；重做集中在三审环节。
7. 差异说明：建议新增"紧急采购"专用分支 + 三审阈值上调；提交评审。
8. 自动化：为"发票先于 PO"配置 Action Flow 自动告警并路由给 AP。
9. 度量：周期时间由 14 天降至目标 9 天；无人工干预率目标由 41% 提至 60%。
约束：不得在差异说明获批前修改流程定义；不得自行处置涉及付款的实际业务动作。

4.3. Implementation Checklist

5. Summary

BPM is the direction in the knowledge-collaboration group with the most complete standards and the most mature audit semantics, and is the key interface through which AI Harness capabilities extend into business governance.

The five most important conclusions:

1. The standard system is complete and usable: BPMN 2.0.2 has been officially published by ISO/IEC as ISO/IEC 19510:2013; CMMN 1.1 (2016-12) fills the scenario of "non-predefined processes that depend on ad hoc knowledge-worker decisions"; and DMN provides decision-modeling capability with three conformance levels (Level 1 not executable → Level 3 with full FEEL executable).
2. CMMN is the best standard expression for human + Agent hybrid processes. Its "ad hoc knowledge-worker decisions" semantics is precisely the business-side counterpart of human-in-the-loop, and the differentiation anchor between BPM and Workflow.
3. Object-centric is the inevitable direction for process mining. Flattening multidimensional event data into a single case causes the three kinds of distortion — deficiency, convergence, and divergence; OCEL 2.0 (released 2023-10-19, with the three interchange formats SQLite/XML/JSON) and IEEE 1849-2023 are the current standard answers.
4. Conformance checking is the natural interface between BPM and L6 governance. It aligns "actual execution" with the "specification model" and outputs auditable deviation evidence — a product that audit, internal control, and compliance can consume directly.
5. Economic value can be quantified. Celonis's Forrester TEI study reports ROI 383%, NPV USD 35 million, a payback period within 6 months, and an order automation rate of 33%→86% (R1, the study is downloadable); the Genpact case reports a 98% reduction in GRIR open items and a P&L impact of USD 45 million (R1). However, note that the TEI is a vendor-commissioned study and its composite organization is a modeling assumption.

Gaps that must be stated frankly: all of the Gartner hyperautomation and RPA market data is secondhand reporting, and no official original publication was found; no authoritative market-size data for the process mining (non-RPA) market was retrieved; and DMN's stable version number is not fully given on the OMG category page. All of these are marked [To be verified] in the body text, with no fabricated fill-in.

Information Gap Declaration

6. References

1. BPMN 2.0 — Object Management Group. https://www.omg.org/spec/BPMN/2.0/
2. Relationship between BPMN and ISO/IEC 19510:2013 — OMG. http://production.omgsysml.org/spec/BPMN/index.htm
3. CMMN 1.1 — Object Management Group. https://www.omg.org/spec/CMMN/1.1/About-CMMN
4. OMG Business Modeling category page (including DMN 1.7 beta, 2024-09) — OMG. https://www.omg.org/spec/category/business-modeling/
5. DMN conformance levels (Level 1/2/3) — IBM BAMOE documentation. https://www.ibm.com/docs/bg/ibamoe/9.1.x?topic=models-dmn-reference
6. Good news about the event data standards (IEEE 1849-2023, OCEL 2.0) — IEEE Task Force on Process Mining. http://tf-pm.org/newsletter/newsletter-stream-13-10-2023/good-news-about-the-event-data-standards
7. OCEL 2.0 — Object-Centric Event Log Standard official site. https://www.ocel-standard.org/
8. OCEL 2.0 formal definition paper — Adams et al. (arXiv:2403.01975). https://arxiv.org/abs/2403.01975
9. OCEL 2.0 ecosystem and tools — arXiv:2410.14495. https://arxiv.org/abs/2410.14495
10. Ocelot — object-centric process mining tool. https://ocelot.pm/
11. PM4JS — JavaScript process mining library. https://www.pm4js.org/
12. ProM Tools — process mining toolset. https://promtools.org/
13. Celonis customers saw payback in 6 months and 383% ROI — Celonis official press release. https://www.celonis.com/news/press/celonis-customers-saw-payback-in-6-months-and-383-roi
14. The Total Economic Impact™ Of Celonis — Forrester Consulting, 2025-06. https://tei.forrester.com/go/celonis/processIntelligence/docs/TheTEIOfCelonis.pdf
15. Finance and Supply Chain Optimization with AI Process Mining — Genpact official case study. https://www.genpact.com/case-studies/finance-and-supply-chain-optimization-with-ai-process-mining
16. Process Mining in Action: €1M Savings in P2P and AP — Dynafin (R2). https://dynafin.eu/process-mining-in-action-how-to-deliver-e1m-savings-in-p2p-and-ap
17. The four process mining tasks and the object-centric paradigm — ScienceDirect (R2). https://www.sciencedirect.com/science/article/abs/pii/S030643792600027X
18. CMMN engine semantics (sentry / plan item lifecycle) — Flowable documentation. https://www.flowable.com/open-source/docs/cmmn/ch06-cmmn