From Classical Ensembles to Agentic Intelligence: Modernizing Automated Incident Ticket Routing in Telecommunications

Abstract

Automated incident ticket routing in telecommunications, historically addressed through TF-IDF vectorization paired with XGBoost and Random Forest classifiers, has reached a performance ceiling defined not by model capability but by the data engineering architectures that feed those models. This paper examines a prior machine learning framework for classifying and routing Remedy tickets within a major telecommunications carrier, identifies the specific pipeline failures responsible for its structural limitations, and proposes a three-layer modernization comprising domain-adaptive Transformer fine-tuning (ModernBERT with Low-Rank Adaptation), Retrieval-Augmented Generation for resolution recommendation, and a constrained agentic orchestration layer for priority incidents. Each layer is specified at the data pipeline level, not only the model level, addressing ingestion, feature store, vector store, and MLOps infrastructure. The central contention is that model architecture is the secondary constraint in ITSM automation systems, and the binding factor is the quality and adaptability of the data pipeline surrounding any model. The data engineering investment required to sustain these layers in production feature stores, schema contracts, event-driven ingestion, RAGOps monitoring, and continuous parameter-efficient adaptation is where the real implementation work lies, and where prior literature on ITSM automation has been systematically underweighted. This paper presents a critical retrospective of a deployed production system and a modernization roadmap grounded in published empirical results.