[論文レビュー] PMAx: An Agentic Framework for AI-Driven Process Mining

Anton Antonov, Humam Kourani+3 more|arXiv (Cornell University)|Mar 16, 2026

Business Process Modeling and Analysis被引用数 0

ひとこと要約

PMAx is an autonomous, privacy-preserving framework that uses a two-agent (Engineer and Analyst) setup to generate and interpret process mining artifacts locally, avoiding LLM hallucinations and data leakage.

ABSTRACT

Process mining provides powerful insights into organizational workflows, but extracting these insights typically requires expertise in specialized query languages and data science tools. Large Language Models (LLMs) offer the potential to democratize process mining by enabling business users to interact with process data through natural language. However, using LLMs as direct analytical engines over raw event logs introduces fundamental challenges: LLMs struggle with deterministic reasoning and may hallucinate metrics, while sending large, sensitive logs to external AI services raises serious data-privacy concerns. To address these limitations, we present PMAx, an autonomous agentic framework that functions as a virtual process analyst. Rather than relying on LLMs to generate process models or compute analytical results, PMAx employs a privacy-preserving multi-agent architecture. An Engineer agent analyzes event-log metadata and autonomously generates local scripts to run established process mining algorithms, compute exact metrics, and produce artifacts such as process models, summary tables, and visualizations. An Analyst agent then interprets these insights and artifacts to compile comprehensive reports. By separating computation from interpretation and executing analysis locally, PMAx ensures mathematical accuracy and data privacy while enabling non-technical users to transform high-level business questions into reliable process insights.

研究の動機と目的

• Democratize process mining by enabling non-technical users to obtain process insights through natural language interactions.
• Eliminate reliance on external AI services for raw data by performing computation locally with privacy-preserving metadata.
• Provide a reliable, self-correcting execution loop to ensure accurate outputs without human intervention.
• Offer an extensible, open-source architecture based on standard Python tools for easy deployment and customization.

提案手法

• Introduce a privacy-preserving metadata abstraction of event logs to keep raw data in the local environment.
• Deploy a divide-and-conquer multi-agent workflow with an Engineer agent that synthesizes executable Python code for PM4Py-based analysis and an Analyst agent that interprets artifacts and generates reports.
• Apply a constrained, whitelisted toolset (PM4Py, pandas, numpy, matplotlib, plotly, seaborn) with static verification and a self-correcting loop for runtime errors.
• Maintain a collaborative memory state to separate code-generation context from interpretation context and ensure efficient information exchange.
• Provide an open-source extension within ProMoAI, enabling local deployment and extensibility for new process mining tasks.
• Describe a code-synthesis workflow that includes role-based context injection, data access via a local object, and strict API constraints to prevent unsafe execution.

実験結果

リサーチクエスチョン

• RQ1 How can AI assist process mining while ensuring data privacy by keeping raw logs local?
• RQ2 Can autonomous agents accurately synthesize and execute process mining analyses (models, metrics, visuals) from natural language prompts?
• RQ3 To what extent can a dual-agent (Engineer and Analyst) setup improve reliability and interpretability of process mining results compared to direct LLM interrogation?
• RQ4 What is the practicality and effectiveness of PMAx on real-world datasets for answering business questions such as workflow discovery, throughput, bottlenecks, and offer-related outcomes?

主な発見

• The framework enables data-grounded insights from natural language cues without sharing raw event data externally.
• A verified code-generation loop produces executable Python scripts that run locally and are self-correcting on errors.
• The Analyst component generates comprehensive reports that combine narrative interpretation with visual evidence, grounded in empirical results.
• PMAx is open-source and built on standard Python libraries, designed to be extensible for additional process mining tasks.
• A live demonstration on a real dataset shows autonomous handling of several business questions through artifact generation and reporting.