Telecom networks have used artificial intelligence for years. Operators already apply machine learning to traffic forecasting, predictive maintenance, fault detection, capacity planning, and radio access network (RAN) optimization. However, most of these applications operate as separate intelligence layers sitting on top of existing network infrastructure.
AI-native telecom networks represent a different architectural approach. Instead of adding AI tools to individual network functions, intelligence becomes part of how the network is designed, operated, optimized, and evolved.
This shift has important implications for 5G-Advanced, future 6G networks, cloud-native infrastructure, network automation and telecom operations. It also changes the role of network engineers, operations teams and traditional OSS platforms.
The ITU-T established its Focus Group on AI-Native Networks in 2024 to examine the architectural requirements, opportunities and challenges associated with embedding AI into telecommunications networks.
For operators, the question is no longer simply how AI can improve network performance. It’s how the network itself should change when intelligence becomes a fundamental architectural capability.
What Makes a Telecom Network AI-Native?
There’s an important difference between an AI-assisted network and an AI-native network. In an AI-assisted environment, conventional network functions continue operating largely through predefined logic, configuration rules, and management procedures. AI applications provide recommendations or perform specific optimization tasks, but the underlying architecture remains mostly unchanged.
An AI-native network integrates intelligence more deeply into network functions and operational processes. AI models can analyze network conditions, understand service requirements, predict changes in demand, and coordinate decisions across different network domains.
For example, a traditional network management system might generate an alarm when a cell site experiences congestion. An AI-assisted system could recommend a parameter adjustment. An AI-native architecture could correlate the congestion with transport capacity, neighboring cells, subscriber experience, and service priorities before coordinating an appropriate response. The difference is the level of integration between intelligence, decision-making and network execution.
How Network Architecture Changes
Moving toward AI-native networking requires changes across several architectural layers.
1. Intelligence Becomes Part of Network Functions
AI is no longer restricted to a separate analytics platform. Network functions can use AI models for specific operational decisions, including resource allocation, traffic prediction, anomaly detection, and performance optimization.
This requires operators to consider where models execute, how they receive network information, and how their decisions interact with existing control mechanisms. Some intelligence may operate centrally, while time-sensitive decisions may need to happen closer to the RAN, core network or edge infrastructure.
2. Network Data Becomes a Shared Operational Asset
AI-native networks depend on continuous access to reliable data. Network telemetry, signaling information, performance counters, subscriber experience measurements, service policies and infrastructure information need to be available in a consistent form.
This is more complicated than collecting data in a central analytics platform. Operators need to address data quality, ownership, synchronization, security and access across multi-vendor environments. Without that foundation, AI models may optimize one network domain while creating problems elsewhere.
3. Closed-Loop Automation Becomes More Important
Traditional network automation generally follows predefined workflows. AI-native architectures can introduce more adaptive decision-making. An AI system can identify a performance issue, determine an appropriate action, request execution through an authorized interface and evaluate whether the change produced the expected result.
TM Forum’s Autonomous Networks architecture identifies self-healing, self-optimization and closed-loop management as important capabilities in the evolution toward more autonomous network operations. However, closed-loop automation requires clear policies and operational boundaries. Not every network decision should be delegated to an AI model.
AI-Native RAN: Moving Beyond Traditional Radio Optimization
The Radio Access Network is one of the most promising areas for AI-native architecture because radio conditions change continuously.
Traffic distribution, interference, mobility, signal quality, and user density can vary significantly throughout the day. Traditional RAN optimization relies heavily on configured parameters, thresholds, and engineering interventions.
AI-native RAN introduces more adaptive resource management. Models can analyze traffic patterns, predict congestion, optimize radio resources, and coordinate decisions across neighboring cells.
Open RAN also creates opportunities for this approach through its RAN Intelligent Controller (RIC) architecture.
The Near-Real-Time RIC supports applications such as xApps for near-real-time RAN optimization, while the Non-Real-Time RIC supports longer-term intelligence and policy management through rApps.
The challenge is coordinating these applications. Multiple AI models may attempt to optimize the same network resources, potentially producing conflicting decisions. Research published by IEEE in 2026 highlights conflict-aware control and intent-driven orchestration as important considerations for AI-native O-RAN architectures.
AI-Native Core Networks and Intelligent Resource Management
The mobile core is another area where embedded intelligence could change network operations. In 5G Standalone environments, operators manage network functions, subscriber sessions, policy decisions, authentication, charging and network slicing. These functions generate large volumes of operational information that can support predictive optimization.
An AI-native core could use this information to anticipate resource demand, identify abnormal session behavior and improve service allocation.
For example, an enterprise network slice supporting industrial applications may have different performance requirements from a consumer broadband service. Intelligence could help predict demand and coordinate resource allocation according to service policies.
This doesn’t mean AI replaces the 5G Core control plane. Existing 3GPP procedures and network-function responsibilities remain important. AI provides additional intelligence that helps operators make better decisions within the established architecture.
The longer-term direction is to make network functions increasingly capable of adapting to changing service requirements without requiring constant manual intervention.
Digital Twins and Predictive Network Operations
Digital twins could become an important component of AI-native telecom architecture. A network digital twin provides a virtual representation of network infrastructure and its behavior. Operators can use it to simulate traffic conditions, evaluate configuration changes and understand the possible consequences of network modifications before applying them to production.
Imagine an operator planning a major sporting event in a city. Instead of relying entirely on historical traffic estimates, a digital twin could simulate different demand scenarios, including increased mobile data consumption, temporary congestion and changes in subscriber distribution.
AI models could analyze the simulation results and recommend capacity adjustments or configuration changes. This approach is particularly useful when network modifications carry operational risk. Testing a decision in a controlled environment can reduce the possibility of introducing unexpected problems into a live network.
What Happens to Traditional OSS and Network Operations?
AI-native networking doesn’t make OSS platforms irrelevant. In fact, it increases the importance of integration between network management, service orchestration, inventory, assurance and automation systems.
Traditional OSS environments often operate through separate workflows for fault management, performance monitoring, configuration and service assurance. AI-native architecture requires these capabilities to exchange information more effectively.
For example, a network performance problem may affect several services simultaneously. An intelligent operations platform should be able to understand the relationship between the infrastructure issue, affected customers, service dependencies and business impact.
TM Forum’s June 2026 AI-Native ODA roadmap addresses this challenge by proposing an open architectural foundation for interoperable AI agents operating across telecom IT and network environments. This is important because operators don’t want separate AI platforms making disconnected decisions across every network domain.
Where Telecom Technology Vendors Fit In
The transition toward AI-native networks will involve both network infrastructure providers and BSS/OSS technology vendors. Companies such as Amdocs, Optiva, Telgoo5, TelcoEdge Inc. and MATRIXX Software operate in different parts of the telecom software and monetization ecosystem. Their relevance depends on how operators connect network intelligence with subscriber management, charging, billing, policy and operational workflows.
The broader architectural opportunity is connecting network intelligence with commercial intelligence. If network conditions change, operators should be able to understand the impact on service quality, customer experience, resource consumption and monetization. AI-native networks will therefore require stronger coordination between traditionally separate network and business systems.
The Challenges Operators Need to Address
Embedding AI into network architecture introduces challenges that go beyond model accuracy. Data quality and interoperability: AI models need consistent information across multi-vendor infrastructure. Incompatible data formats and fragmented telemetry can limit the effectiveness of cross-domain intelligence.
Security and governance: AI models need controlled access to network functions. Operators must prevent unauthorized configuration changes, data exposure and unsafe automated decisions.
Explainability: Network engineers need to understand why an AI system recommended or executed a particular action, especially when service availability is affected. Operational stability: Two independent AI models can make conflicting optimization decisions. Coordination, policy enforcement and validation mechanisms are essential.
Legacy integration: Most operators cannot replace their entire network architecture at once. AI-native capabilities must coexist with existing network functions, OSS platforms and operational processes. These challenges make incremental implementation more realistic than attempting to build a completely autonomous network in one transformation program.
How Operators Should Prepare for AI-Native Networks
Operators should start by identifying network domains where AI can deliver measurable operational improvements. RAN optimization, predictive maintenance, fault correlation and energy management are reasonable starting points because they have identifiable performance indicators and established operational workflows.
The next step is building the data and integration foundation. Network telemetry, service information, inventory and operational events should be accessible through consistent interfaces. Operators should then introduce closed-loop automation gradually. Initially, AI can recommend actions for engineering teams. Once the models demonstrate reliability, selected low-risk decisions can be automated under defined policies.
TM Forum’s Autonomous Networks framework provides a useful maturity model for this journey. Its Level 4 capability includes intent-driven, predictive decision-making and closed-loop management, while higher levels move toward greater network autonomy.
The Network Will Become More Than a Connectivity Platform
AI-native telecom networks represent a deeper architectural change than adding machine learning to network monitoring. Intelligence becomes part of how network resources are managed, how services are delivered and how operational decisions are made. RAN, core, transport, edge computing and OSS environments increasingly need to work together through coordinated intelligence.
For operators, the transition will take time. Legacy infrastructure, multi-vendor integration, governance and operational trust remain significant challenges. But the direction is becoming clearer: future telecom networks will not simply carry traffic more efficiently; they will increasingly understand, predict and adapt to the demands placed on them.
The real question for network architects is how to build that intelligence into the architecture without compromising reliability, security or operational control.