Abstract

Public administration processes often suffer from hidden bottlenecks and complex procedural heterogeneity. While traditional process mining discovers workflow mappings, it struggles to provide contextual prescriptive solutions. We propose a Neuro-Symbolic dual-pipeline architecture. The first pipeline utilizes predictive machine learning, applying Strict Temporal Splitting, to classify delayed cases without data leakage. The second introduces a Multi-Agent Large Language Model (LLM) ensemble utilizing Bottleneck-Anchored Context Windowing (BACW). By standardizing semantic milestones and deterministically bounding contexts within event logs, the system autonomously synthesizes evidence-based root-cause hypotheses. Evaluated on the BPIC 2015 dataset, this architecture bridges the gap between predictive monitoring and actionable process redesign, significantly reducing LLM hallucinations compared to full-trace baselines.

Recommended Citation

Orlikowski, P., Kozina, A. & Pikus, M.(2026). AI-Augmented Process Mining: A Framework for Accelerating Digital Transformation in Public Administration. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.42

Paper Type

Poster

DOI

10.62036/ISD.2026.42

Share

COinS
 

AI-Augmented Process Mining: A Framework for Accelerating Digital Transformation in Public Administration

Public administration processes often suffer from hidden bottlenecks and complex procedural heterogeneity. While traditional process mining discovers workflow mappings, it struggles to provide contextual prescriptive solutions. We propose a Neuro-Symbolic dual-pipeline architecture. The first pipeline utilizes predictive machine learning, applying Strict Temporal Splitting, to classify delayed cases without data leakage. The second introduces a Multi-Agent Large Language Model (LLM) ensemble utilizing Bottleneck-Anchored Context Windowing (BACW). By standardizing semantic milestones and deterministically bounding contexts within event logs, the system autonomously synthesizes evidence-based root-cause hypotheses. Evaluated on the BPIC 2015 dataset, this architecture bridges the gap between predictive monitoring and actionable process redesign, significantly reducing LLM hallucinations compared to full-trace baselines.