Agentic system for construction safety risk assessments using large language models and knowledge graphs

• Integrates domain-specific adaptations of AI technologies into construction safety. • Uses LLM-based agents to generate structured risk assessment drafts. • Transforms safety knowledge into ontology-aligned SPARQL for knowledge graph integration. • Human safety experts evaluate the generated risk assessments and SPARQL queries. • Demonstrates that AI produces actionable drafts, while expert oversight remains essential. Construction safety risk assessments are labor-intensive and disconnected from dynamic project conditions and digital systems. This research investigates how domain-specific multi-agent generative artificial intelligence systems can automate construction safety risk assessments and subsequently share them with knowledge graphs (KGs) to enable downstream reuse of the extracted knowledge and overcome information isolation. The research employs a Design Science Research methodology to develop a novel multi-agent system leveraging Large Language Models (LLMs) and semantic-web techniques. Individual agents within the system perform specific tasks, including hazard identification, risk mitigation, and ontology-aligned SPARQL query production. An evaluation with eight construction safety experts validates that the system successfully automates the production of complete and relevant risk assessment documents. A semantic, structural, and syntactical correctness assessment through SHACL shapes, and manual completeness evaluation, furthermore validates that the system reliably converts the risk information into accurate SPARQL insertions compatible with existing ontology-based KGs. Unlike prior research, this agentic pipeline automates task-based assessment generation and semantic integration, enabling downstream reasoning and interoperability within digital twin environments. Safety professionals benefit from faster and machine-readable availability of safety information that supports downstream reasoning processes (e.g., hazard detection or simulation), which are not evaluated in this study. Future research should identify whether the provided knowledge is sufficient to streamline such downstream processes and outline clear information requirements to enhance the system further. Future research should investigate automated and scalable semantic validation methods, reasoning over concurrent tasks, and integration with digital twin systems.