Skip to main content Skip to footer

AI-native networks: hype or reality in 2027?

Ai Native Networks

Artificial intelligence is becoming integral to how telecoms networks operate, promising to predict faults, optimise performance and enable greater autonomy. As operators invest in AI-RAN and intelligent orchestration, what will be commercially viable in 2027 – and can their BSS/OSS keep pace?

The telecoms industry has never been short of promises about next-generation networks.

Over the past decade, even before the recent artificial intelligence boom, AI and machine learning have become increasingly common in network operations, helping CSPs predict equipment failures and automate fault responses.

Self-organising networks, intent-based networking, zero-touch operations; each has brought us a little closer, but none has completely removed the need for human intervention.

Together, these developments are laying the groundwork for what the industry now calls AI-native networks, in which intelligence is integral to the architecture. Rather than simply following predefined instructions, these networks continuously interpret conditions, make decisions and adapt based on the results.

However, “AI-native” is becoming an umbrella term for an enormous range of capabilities, from relatively conventional predictive analytics to networks that could, theoretically, run themselves. It isn't always clear where one ends and the other begins.

TM Forum's autonomous networks framework offers a useful way to distinguish between them, defining six maturity levels from manual operations at Level 0 to full autonomy at Level 5.

Level 4 is regarded as a significant milestone at which networks can make predictive, intent-driven decisions and manage operations through closed-loop automation. But according to TM Forum research involving 80 operators, just 20% of those surveyed expect to reach Level 4 or above by 2027, although 81% are targeting that milestone by 2030. Even so, achieving autonomy in individual network functions is very different from coordinating decisions across complex multi-vendor networks.

The foundations of AI-native networks

Achieving network autonomy requires more than sophisticated AI models. Operators need common foundations for managing those models, sharing trustworthy data and securing interactions between autonomous systems.

In March 2026, TM Forum launched three projects within its AI-Native Blueprint, addressing fundamental barriers to deploying AI reliably at scale:

Model as a Service (MODaaS) aims to standardise how AI models are deployed, managed and consumed across cloud, edge and on-premises environments, reducing integration complexity and dependence on individual vendors.

Data Products Lifecycle Management (DPLM) aims to address fragmented and inconsistent data by establishing standards for making operational information discoverable, accessible and governed across network and business domains.

Agentic Interactions Security focuses on developing common security policies and machine-readable rules to govern interactions between autonomous AI agents, including what actions they are permitted to perform.

AI-RAN: breakthrough, or money pit?

One of the most prominent tests of these foundations is AI-RAN, which promises to integrate AI processing into radio infrastructure to improve spectrum utilisation, manage interference, reduce energy consumption and potentially support AI workloads at the network edge.

In October 2026, Nokia and NVIDIA announced progress towards commercial AI-RAN deployments. T-Mobile completed what Nokia described as the first AI-RAN call at a live commercial cell site, while BT announced plans for a live network trial in 2027. That same week, T-Mobile also reported that, during peak hours in a large-scale commercial trial with Ericsson, AI-managed beamforming improved downlink throughput by 16% and spectral efficiency by 12%, using an AI model running on existing RAN computing infrastructure.

Nokia and NVIDIA are pursuing a GPU-accelerated architecture capable of supporting AI workloads alongside radio functions. This could create new commercial opportunities, but it would also introduce additional computing, power and integration costs. Meanwhile, Ericsson’s results suggest operators may be able to realise AI benefits without adopting an entirely new GPU-accelerated architecture.

In November, VodafoneThree also announced trials with Nokia and NVIDIA, initially in a laboratory before progressing to a live network site. The operator intends to benchmark improvements in capacity, reliability and spectral efficiency, while exploring applications including energy optimisation and differentiated connectivity services. However, the operator has yet to demonstrate the additional demand or revenues such capabilities might generate.

BT is planning AI-RAN trials on its EE network in Birmingham during 2027, while Virgin Media O2 has already activated a trial site in London. All three are exploring the potential for more intelligent radio networks, although their deployments remain at different stages of maturity.

The growing number of trials suggests genuine industry momentum, but also illustrates the gap between technological ambition and commercial reality. Operators are increasingly demonstrating that AI-RAN can be deployed; 2027 will need to establish whether it delivers sufficient benefits to justify the investment.

These competing approaches underline that AI workloads must be matched to the right computing resources. At its September 2026 AI-Native Telco Forum, TM Forum emphasised that these choices should reflect cost, latency and the required outcome.

Operators could become increasingly dependent on proprietary computing platforms, potentially shifting value towards the companies supplying them. And whilst the technology may work, operators still risk investing heavily in AI-ready infrastructure before identifying how to generate a return.

Trust issues

With greater autonomy comes greater consequences if things go wrong.

AI agents might be able to make a network configuration change in milliseconds, but what happens if that decision turns out to be wrong?

An incorrect recommendation might inconvenience one customer; an autonomous decision affecting network routing, capacity or provisioning could disrupt thousands of customers. And speed creates another risk: an AI agent could execute changes in milliseconds, while operators may take considerably longer to verify the consequences.

Probabilistic AI can help diagnose faults and identify patterns, but network operations frequently demand deterministic outcomes, predictable behaviour and strict policy enforcement.

Making networks more autonomous may also, paradoxically, require greater investment in the systems governing that autonomy. TM Forum argues that full autonomy is neither immediately achievable nor desirable in every scenario. Instead, it advocates for “bounded, trusted autonomy”, where AI agents operate within clearly defined permissions, policies and responsibilities.

AI-native BSS/OSS

However, trusted autonomy is only part of the challenge. AI-native networks also depend on the business and operational systems that turn network decisions into commercial services.

Consider an enterprise customer requesting guaranteed connectivity for a temporary event. An AI-native network could assess capacity, configure resources and maintain the required service quality, but monetising that capability requires systems to create the product, provision the service, measure consumption and apply charges.

The network might be able to deliver the service autonomously, but can the operator generate revenue from it with the same speed and flexibility?

Operators have spent decades accumulating fragmented systems and complex integrations. AI-native networks cannot deliver autonomous customer experiences if the supporting BSS/OSS platforms still rely on manual workflows, rigid product configurations and disconnected data.

Adding autonomous agents to this fragmented environment would not remove the underlying complexity. Without common data models, open interfaces and consistent governance, CSPs risk replacing traditional application silos with agentic silos, creating further integration challenges and conflicting objectives.

Ultimately, the commercial benefits of AI-native networking will depend on connecting intelligent network decisions with equally flexible business processes.

What does 2027 hold?

We probably won't see networks running themselves from end to end in 2027. What we should see is a clearer picture of which AI-native technologies are worth investing in.

Progress should be incremental but increasingly measurable, with the greatest near-term benefits coming from targeted applications in energy optimisation, predictive maintenance, service assurance and capacity management.

For CSPs, this points to an evolutionary approach: introducing intelligence into existing infrastructure while improving coordination between network and business domains.

Can AI reduce operating expenditure after accounting for computing, integration and governance costs? Can it open new revenue opportunities, or will it simply introduce another expensive technology layer?

Achieving genuine business value will depend as much on flexible BSS/OSS as on sophisticated AI models. The priority in 2027 should be targeted, governed automation that delivers a measurable return.

AI-native networks need equally intelligent business systems. Explore the Cerillion AI Hub to learn how Cerillion’s embedded AI capabilities, intelligent agents and MCP-enabled integration work across its pre-integrated BSS/OSS Suite, helping CSPs introduce AI into real operational workflows with the governance and control needed to deliver measurable business value.

About the author

Adam Hughes

Content Specialist, Cerillion

Keep up with the latest company news and industry analysis