Communications Service Providers (CSPs) face the dual challenge of managing escalating network complexity driven by 5G and edge computing, while simultaneously controlling operational expenditures (OpEx) and creating new revenue streams. Traditional, manual approaches to network management are no longer sustainable. In response, the industry is accelerating its adoption of AIOps and autonomous networks, leveraging artificial intelligence and machine learning to automate operations, predict and prevent failures, and optimize performance in real-time. This transition promises not only to reduce the cost-to-serve but also to fundamentally enhance network reliability and deliver superior customer outcomes through proactive and intelligent management.
This paper presents a holistic analysis of the strategic, technical, and forward-looking dimensions of AIOps and network automation in telecommunications. It addresses the impact across all primary network domains: the Radio Access Network (RAN), Core, Transport, and Enterprise Services. The objective is to synthesize evidence from industry frameworks, technical trials, and real-world deployments to build a comprehensive understanding of the business cases, enabling architectures, and implementation challenges. By examining the trajectory toward higher levels of autonomy, specifically TM Forum Levels 4 and 5, this paper clarifies the role of emerging technologies like Generative AI (GenAI) and digital twins in realizing the vision of a fully autonomous, zero-touch network.
The foundation for autonomous networking is built upon several key industry frameworks that define architectures and maturity levels. The TM Forum’s Autonomous Networks (AN) Framework is central, providing a maturity model that guides CSPs on their journey. Level 4, or Highly Autonomous Networks, is characterized by predictive closed-loop management across multiple domains to prevent risks, while Level 5, or Full Autonomous Networks, envisions complete closed-loop automation across the entire service lifecycle with minimal human intervention.5 Generative AI is identified as a critical enabler for this transition, functioning as the network’s ‘decision brain’ to translate business intent into network actions.5
Standardization bodies have developed specific architectural components to support this vision. The 3rd Generation Partnership Project (3GPP) has evolved its network automation architecture significantly, notably with the Network Data Analytics Function (NWDAF). Release 17 decomposed the NWDAF into the Model Training Logical Function (MTLF) for training ML models and the Analytics Logical Function (AnLF) for inference, supported by new functions for efficient data collection and storage.14 Complementing this core-centric view, the O-RAN ALLIANCE provides a framework for RAN-level intelligence through its Service and Management Orchestration (SMO) Framework and the RAN Intelligent Controller (RIC).
The RIC is designed to host AI/ML-driven applications (xApps and rApps) for fine-grained radio resource optimization and slice assurance, a domain not explicitly covered by the NWDAF.15 Furthermore, the ETSI Generic Autonomic Networking Architecture (GANA) framework positions functions like NWDAF as data sources for a higher-level Knowledge Plane (KP) Platform, which houses autonomic decision-making elements.21
This paper employs a structured analysis and synthesis of contemporary industry reports, technical specifications, proof-of-concept results, and academic research to construct a comprehensive overview of AIOps and autonomous networks in the telecommunications sector. The methodology is qualitative, focusing on the collation and interpretation of evidence from diverse, authoritative sources to address the multifaceted nature of network automation. The analysis is structured to align with the strategic, technical, and forward-looking perspectives requested. It systematically examines documented business cases and ROI metrics, deconstructs key technical architectures and their interplay, and evaluates prioritized use cases across different network domains. Finally, it synthesizes the identified challenges and enablers to map a realistic trajectory toward advanced network autonomy as defined by established industry frameworks.
The findings reveal how AIOps and autonomous network architectures translate into measurable improvements across cost-to-serve, reliability, and customer experience. This section synthesizes financial, technical, and operational evidence to illustrate the practical impact of automation across key telecom domains.
The financial incentives for pursuing network autonomy are substantial. An average CSP can realize approximately US$794 million in annual value from implementing autonomous networks, a figure comprising US$650 million in capital and operational expenditure savings and US$144 million in revenue uplift.6 Crucially, about 30% of this financial benefit is unlocked only upon reaching the higher maturity levels of TM Forum Levels 4 and 5.6
Specific deployments have yielded quantifiable ROI. Netcracker reported that its Agentic AI solutions delivered a 40% reduction in time to resolve call center calls, a 70% reduction in time to resolve complex B2B billing issues, a twelve-fold acceleration in making catalog changes, and a seven-fold improvement in service design creation speed.3 Similarly, another operator improved efficiency by over 20% by deploying a GenAI-enabled chatbot that provided technicians with instant access to digitized network equipment manuals.1 These results underscore the direct impact of AI-driven automation on operational efficiency and cost reduction.
Achieving network autonomy relies on a sophisticated and evolving set of architectures. Agentic AI is emerging as a powerful paradigm, with two primary architectural models being considered: hierarchical, which centralizes decision-making, and distributed, which pushes intelligence to the network edge for local optimization. Future systems are expected to be hybrid, blending these approaches.3 However, it is argued that AI agents should be viewed not as a standalone architecture but as a realization technique for functions within existing standardized frameworks, such as TM Forum’s Intent Management Functions or O-RAN’s rApps, using established interfaces to ensure interoperability.13
The 3GPP NWDAF and O-RAN RIC are foundational pillars of this new architecture. While the NWDAF provides centralized analytics for the 5G Core,14 the RIC enables fine-tuned control and optimization within the RAN.15 This has led to a debate on implementation, with some proposing that a separate, containerized service assurance solution is needed to provide true end-to-end visibility, potentially integrating with or enhancing the NWDAF.16 AIOps platforms are positioned as a flexible alternative for implementing NWDAF use cases, offering a more pliable architecture under direct operator control.20
Bridging these systems, digital twins and GenAI are proving to be powerful integrators. The TM Forum BIND Catalyst project, a collaboration involving Amdocs, Vodafone, Telstra, and Google Cloud, demonstrates this approach.4,18 Its architecture ingests siloed data using TM Forum Open APIs (TMF638, TMF639, TMF642, TMF628) to create real-time digital twins of the network.4 Upon this foundation, a layer of GenAI-powered agents collaborates using Google’s open Agent-to-Agent (A2A) Protocol to proactively identify, simulate, and resolve issues.4,17
The application of AIOps varies across network domains, with operators typically prioritizing use cases that offer immediate OpEx reduction before shifting focus to revenue generation.
In the Radio Access Network (RAN), energy efficiency is a primary focus. Ericsson’s Energy Cockpit rApp analyzes the network at the cell level to activate power-saving features without impacting traffic.12 A multi-vendor O-RAN trial coordinating an energy-saving rApp with a traffic-steering xApp demonstrated a 25% reduction in energy consumption while maintaining 99.999% accessibility.10,25 Another critical use case is RAN slice SLA assurance. A proof of concept involving Juniper Networks and Vodafone used AI/ML models to forecast and proactively prevent 15% to 30% more SLA violations than traditional methods.11 Furthermore, in a real-world offshore wind farm deployment, LSTM models integrated into an O-RAN platform reduced connectivity issues by over 90%.27
For Core and Transport Networks, closed-loop service assurance is paramount. Nokia’s Assurance Center has been shown to automatically resolve 40% of incidents without human intervention and reduce network events by 95-98% by proactively managing SLAs and triggering actions across domains.9 As 5G matures, dynamic resource optimization for network slicing becomes essential for monetization, with the global market projected to be worth US$300 billion by 2025.8 AI is critical here, enabling real-time monitoring, predictive analytics, and automated resource allocation to enforce dynamic SLAs for demanding applications like autonomous vehicles and remote surgery.7 This is often achieved by integrating assurance with the 3GPP NWDAF to trigger closed-loop corrective actions.8
The discussion interprets the findings within the broader context of telecom transformation, highlighting the strategic implications of progressing toward higher levels of network autonomy. It also examines the organizational, technical, and operational considerations that shape the path forward.
The findings illustrate a clear, albeit challenging, path toward higher levels of network autonomy. The BIND Catalyst project serves as a compelling blueprint for achieving Level 4+ autonomy, demonstrating that the combination of digital twins, federated data, and interoperable AI agents can achieve up to 90% fault prediction accuracy, 95% end-to-end process automation, a 40% reduction in operational costs, and a 50% cut in mean time to repair (MTTR).4 This model aligns with the vision of Generative AI as the network’s ‘decision brain,’ translating high-level business goals into concrete, automated actions required for L4 and L5 operations.5
Despite the promise, significant barriers remain. Key challenges preventing CSPs from reaching higher autonomy levels include difficulties in accessing and correlating data from disparate systems, ensuring personnel trust in automated decisions, and managing the unique risks of AI in mission-critical environments.6
Technical hurdles are prominent, particularly in multi-vendor environments like O-RAN. The O-RAN Alliance categorizes conflicts between xApps into three types: direct (updating the same parameter), indirect (updating dependent parameters), and implicit (unobvious adverse effects), which require sophisticated conflict mitigation frameworks to resolve.23,24 A comparative analysis of different Deep Reinforcement Learning-based xApps also found that design choices can lead to competitive, rather than cooperative, behavior.26
Trust and security are overarching concerns. The ‘black-box’ nature of many AI models is a major barrier to adoption, as operators cannot fully understand their decision-making processes. This has led to calls for eXplainable AI (XAI) to ensure transparency, fairness, and accuracy, especially when an AI decision leads to an SLA violation.28 For network slicing, AI systems are vulnerable to security threats like adversarial attacks, posing risks for sensitive applications.19 AI models for slicing also face inherent limitations, including high dependency on large datasets, significant computational overhead, and challenges in adapting to unforeseen network conditions.19 Finally, not all business models are sound; the concept of AI-RAN, which involves selling idle GPU capacity at cell sites, has been criticized as flawed due to the correlation of network demand and AI demand, effectively devaluing the available capacity.2
The integration of AIOps and autonomous systems represents a critical evolutionary step for the telecommunications industry. The evidence demonstrates clear and significant benefits in reducing cost-to-serve, enhancing network reliability, and improving customer outcomes through intelligent automation. The strategic path forward involves a phased approach, beginning with use cases that deliver immediate operational efficiencies, such as closed-loop assurance and predictive maintenance, and progressing toward revenue-generating capabilities like dynamic network slicing.
Achieving the industry’s ultimate goal of full, zero-touch autonomy—as defined by TM Forum Levels 4 and 5—is contingent upon the successful integration of advanced technologies like Generative AI, agentic AI, and digital twins within standardized architectural frameworks. However, the journey is fraught with challenges. Overcoming issues of data fragmentation, ensuring trust through explainability, mitigating security vulnerabilities, and resolving technical complexities like application conflicts in O-RAN are paramount. Future research and development must prioritize the creation of robust, transparent, and secure AI systems to unlock the full transformative potential of the autonomous network.
