Cascading Multi-Agent Architectures for Multilingual IT Support: Integrating Explainability and Dynamic Workload Optimisation

Abstract

Modern IT helpdesks increasingly need to support multiple languages while staying transparent and resistant to hallucination, yet most deployed chatbots still rely on single-model architectures that struggle on both counts. This paper presents NexaServe v2, a four-tier cascading multi-agent IT helpdesk system combining retrieval-based response, LLM validation, LLM fallback, and human escalation, alongside a hybrid multilingual classification pipeline, confidence-gated routing, and workload-aware ticket assignment. The system is evaluated on a public Kaggle multilingual helpdesk dataset (n = 28,587 tickets), 89 logged chatbot interactions, and operational logs collected during development and testing. Fine-tuning mBERT on combined English-German data reached 89.84% category accuracy, while an English-trained classical model applied to translated German priority tickets dropped from 77.70% to 57.23% accuracy, a much steeper decline than the equivalent category-classification result. The chatbot cascade maintained a low factual hallucination rate (2.7% on a 75-message held-out batch) despite rejecting 93.3-100% of retrieval-layer responses as insufficiently confident. An AI-based workload supervisor, despite generating fluent and specific justifications for its decisions, performed no better than a simple deterministic heuristic and violated its own explicit numerical rule in every recorded overload flag. These results demonstrate that a transparent, near-zero-marginal-cost multilingual helpdesk is achievable using local LLMs for classification and cascaded response generation, while showing that LLM-based supervision does not reliably outperform deterministic alternatives for tasks with strict, checkable correctness criteria.