Enhancing Risk Analysis in Drilling Operations Using an Agentic LLM Workflow

This paper presents a novel, modular agentic workflow leveraging large language models (LLMs)—a technology increasingly validated for its ability to handle complex textual data in drilling operations (e.g., Ferrigno et al., 2024)—and deterministic rule-based heuristics to automate high-fidelity extraction and classification of drilling risk events from daily reports. The system translates unstructured text into structured risk objects, enabling seamless integration with enterprise databases and significantly reducing manual risk analysis workload. The solution was developed as a rapid proof of concept (POC) within eight weeks, designed to pragmatically validate the approach with open-source datasets and scalable architecture. Validation on open-source datasets demonstrated nearly 90% recall for risk event detection and 75% categorization accuracy. The workflow surfaced more than 30 hours of previously unrecognized diminished capacity within productive time (PT) intervals for a single well, while reducing manual review time from 3 hours to less than 5 minutes per well. By surfacing these hidden risks and enabling fine-grained classification, the workflow supports more proactive, data-driven decision-making for drilling engineers and operational teams. This method establishes a new benchmark for automated, consistent, and scalable risk analysis across a broad spectrum of drilling operations.