Automatic AI-Based Extraction of Processes from Administrative Texts: The Case of Waste Management

Public administrations produce large volumes of unstructured administrative documents that regulate operational procedures, responsibilities, and regulatory constraints. These documents are characterized by high information density and fragmented procedural knowledge, making both human interpretation and automated extraction challenging. This issue is particularly critical in the waste management domain, where regulatory complexity and local variability significantly impact operational processes. This paper presents a hybrid pipeline for the automatic extraction and structuring of operational procedures from unstructured administrative texts. The proposed approach combines LLMs with RAG to ensure faithful, traceable extraction constrained to the source document. The pipeline identifies procedural sections, extracts relevant information from each section, and generates visual workflows. The system is evaluated on real-world administrative documents for waste management. Results show that the integration of generative AI within a structured pipeline improves the accuracy and reliability of procedural knowledge extraction, while mitigating hallucination risks.