Abstract
This developmental paper proposes an Information-Events-State (IES) framework for understanding how artificial intelligence can support project risk identification. The framework conceptualises project risk as emerging from a hierarchy of observable information, project events, and resulting compliance states. We apply this lens to benchmark results comparing four LLM-based risk identification approaches – a naive single-call method, an agentic iterative tool-use method, a standalone IES pipeline that operationalises the framework as a three-pass extraction architecture, and an IES-Agentic hybrid that combines agentic exploration with IES-structured reasoning. Evaluated across two language models on a synthetic project with 500 information pieces, the results reveal that while all IES-based approaches achieve perfect precision, the hybrid IES-Agentic mode achieves both perfect recall and substantial false-positive reduction for one model – suggesting that the IES framework can serve not only as an analytical lens but as a viable agent architecture. We outline a development plan toward a full paper incorporating additional models, multi-project evaluation, and human validation.