The RAN has always been the most expensive and operationally demanding part of the telco network. For decades, the strategy has been to manage that cost through vendor relationships, rigid planning cycles, and reactive maintenance. Automation, when it arrived, was largely applied to the same paradigm – doing the same things faster and with fewer people.
That framing is increasingly inadequate. The more consequential question today is what becomes possible when AI is embedded into the RAN as a foundational capability. That is a different conversation – about new services, new business models, and a network architecture that is genuinely responsive to demand rather than one that simply tries to keep up with it.
Most operators have made meaningful progress on RAN automation in recent years. Self-optimizing networks, automated interference management, and AI-assisted capacity planning have all delivered measurable efficiency gains. But there is a ceiling to what closed-loop, task-level automation can achieve, and the industry is approaching it.
The ceiling is structural. When automation is applied within a fixed operational model, optimizing the same parameters, managing the same KPIs, responding to the same triggers, it improves execution without changing outcomes. Costs come down incrementally. Quality improves marginally. But the network itself does not become more capable.
Operators looking to AI for step-change improvement will not find it at the task level. They will find it when AI is applied to the architecture itself – for dynamic resource allocation, real-time service assurance, and network behavior that responds to user and business needs.
“The shift is not from manual to automated. It is from reactive to intelligent – a network that anticipates demand and configures itself accordingly.”
The conversation about AI in the RAN often defaults to the future tense. However, the capability is here today. Not in a lab or on a roadmap, but running in production on existing infrastructure. Rakuten Mobile's collaboration with Intel is a direct illustration of this. Built on a long-standing vRAN deployment using Intel Xeon processors, the two companies are already validating AI use cases across the full RAN stack, targeting real-time AI inference with hardware that operators already have in their networks. The outcomes being delivered today include enhanced spectral efficiency, smarter resource allocation, automated operations, and measurable energy savings – all within the existing cost structure, with no change to the underlying supply chain. The business case does not require a new infrastructure bet. It requires operators to start.
What makes agentic AI architecturally significant is the ability to pursue multi-step goals, act across systems, and adapt based on feedback without human instruction at each stage. Applied to the RAN, that capability affects the nature of network management in transformative ways.
Leading operators working on technology strategy and architecture have increasingly focused on exactly this kind of capability – systems that do not simply respond to defined conditions but reason about network state and make forward-looking decisions. The distinction matters: an automated system executes a policy; an agentic system determines what the policy should be in context.
Experience on the AI product management side reflects a parallel shift in how operators are thinking about RAN intelligence. The question has moved from ‘can we automate this process’ to ‘what does the network need to know, and how do we build systems that act on that knowledge reliably and at scale’.
One of the more significant strategic choices operators face is whether agentic RAN capabilities remain internal operational tools or whether they are exposed externally as programmable services. The two paths lead to very different business models.
Closed-loop automation, however sophisticated, captures value only in operational savings. Exposing agentic workflows as APIs changes the revenue equation. Enterprises, application developers, and solution providers gain the ability to negotiate network behavior programmatically – requesting specific QoS profiles, triggering resource allocation in response to application events, or building service guarantees directly into their platforms.
This is not hypothetical. The emergence of network-as-a-platform thinking across the industry reflects a growing recognition that the RAN’s real commercial potential lies in what it can enable for others, not only in what it costs to run. The technical challenge is making network intelligence accessible without making it brittle – APIs that expose capability without requiring ecosystem partners to understand the underlying complexity.
The phrase ‘demand-led’ gets used loosely, but it points to something specific in the context of AI-native RAN design: a network that allocates resources in response to actual service requirements rather than pre-configured capacity plans. That capability, reliably delivered, enables monetization models that are structurally different from the traditional connectivity tariff.
The monetizable services that become viable in an AI-driven RAN environment include:
Each of these depends on the same underlying capability: a RAN that can reason about its own state, predict demand shifts, and reallocate resources faster than any human-driven process could manage. The technical foundation and the commercial opportunity are not separate conversations.
The broader implication of AI-native RAN design is not just operational efficiency or incremental service innovation. It is a fundamental shift in where and how value is created in the telecom stack.
Traditional telco value chains are vertically integrated and slow to adapt – designed for a world where network capacity was scarce, planning cycles were long, and customer requirements were relatively homogeneous. AI changes all three of those assumptions simultaneously. Capacity can be dynamically allocated. Cycles compress from months to milliseconds. And enterprise customers increasingly need networks that adapt to application behavior rather than the reverse.
Re-architecting for that reality means disaggregating the RAN not just physically – as Open RAN has begun to do – but functionally, so that intelligence, control, and service exposure can be separated and recombined according to customer and market requirements.
This architectural logic extends beyond terrestrial networks. Rakuten's work on NTN-TN convergence – presented at the O-RAN F2F in Rome – demonstrates how RIC and AI agent-driven optimization can manage load balancing and traffic steering between terrestrial and non-terrestrial networks, NTN beam and cell management, UE geo-location estimation, and satellite trajectory prediction, all from a single orchestration layer. Validating these capabilities through a geo-spatial NTN digital twin (stress-testing QoS, coverage, cell switching, and link failures before service launch) is what makes this more than architecture on paper. The boundary of the intelligent RAN is not the cell tower. It means building operational models that treat the network as a platform rather than a managed asset.
The operators that get this right will occupy a different position in the value chain – one that captures a share of the application and enterprise AI growth that is currently flowing past them.
My upcoming session at FutureNet World brings together practitioners who are working through these questions operationally. Join me to discuss the boundary between automation and agency in the RAN context and take a first-hand look at a genuinely demand-led network architecture.
