Human-AI Collaborative Frameworks for Intelligent Fault Detection, Proactive Quality Management, and Augmented Decision Support in Fixed Wireless Access Networks

Abstract

Introduction: The explosive growth of Fixed Wireless Access (FWA) as a primary broadband delivery mechanism has created unprecedented operational challenges for network management teams serving millions of subscriber devices across complex 5G Radio Access Network environments. Traditional reactive operational paradigms, characterized by manual fault response, fragmented monitoring systems, and engineering teams overwhelmed by non-actionable alarms, are fundamentally unable to scale to the quality management demands of modern FWA deployment.Objectives: This article aims to design, deploy, and evaluate a human-AI collaborative framework that unifies fragmented cross-domain telemetry, predicts subscriber-level quality degradation before it becomes customer-visible, closes the loop on the most common class of 5G connectivity fault, and supports engineering judgement through a constrained generative AI interface, while explicitly allocating decision authority between human and AI agents according to risk. Methods: The framework integrates four complementary components: a unified Single Source of Truth cloud-native data platform consolidating consumer profile, in-home device telemetry, and RAN performance data into a coherent real-time analytical substrate; a Network Quality Experience Score (nQES) aggregating over thirty technical features to identify subscribers at elevated churn risk; a closed-loop Missing Anchor anomaly detection and automated remediation system integrated with Self-Organizing Network infrastructure; and a Generative AI natural language interface constrained to Feature Store-grounded inference. The framework was deployed and evaluated across a production FWA network supporting 1.7 million active subscriber devices over a 180-day post-deployment period. Results: The deployed framework achieved an 80% reduction in Mean Time to Resolution (from a median of 4.2 hours to 51 minutes), a 5 percentage point reduction in voluntary churn for nQES-identified intervention cohorts, an 80% reduction in non-actionable alarm volume (from 1,200 to approximately 240 alarms per engineer per shift), and 93.67% classification accuracy on prospective service qualification assessments, a 21 percentage point improvement over the prior rules-based system. Conclusions: These results demonstrate that explicit architectural attention to the division of responsibilities between human and AI agents, operationalized through a tiered automation architecture, produces operational outcomes that neither human teams nor fully automated systems achieve independently, with the greatest gains attributable to unifying the underlying data infrastructure rather than to any single AI model.