From weak signals to risk indicators: An LLM-based information extraction architecture for IT project risk management

Abstract

Effective risk management is crucial yet challenging in modern projects because early warning signals often surface in informal, unstructured data sources and evolve as project contexts change. Since prevailing project risk management (PRM) practices largely depend on structured inputs and episodic control routines, they can miss these subtle signals, creating a temporal detection gap between risk emergence and formal identification. This paper introduces an LLM-based architecture for information extraction in PRM. Following a Design Science Research approach, we design, instantiate, and evaluate it, governed by two complementary flows: an Information Flow that continuously transforms heterogeneous project communication into structured, evidence-linked risk indicators through domain-specialized extraction and cross-source reconciliation, and an Extraction Control Flow that versions extraction behavior to remain adaptable, transparent, and auditable. We abstract four nascent design principles capturing transferable design knowledge for LLM-based information extraction in dynamic, context-rich domains.