Telecom has reached Level 4 autonomy in live networks. The operational implications are significant

It is time to move telecom’s autonomy narrative out of strategy decks and into production networks. 

TM Forum has formally validated Rakuten Mobile at Level 4 autonomy for RAN energy efficiency – on a live, at-scale Open RAN network in Japan. It’s no longer in a lab. Neither is it a proof-of-concept. We’re currently looking at real-world production traffic, with real customers, running continuously. The validation benchmarks against TMF GB1059H are in and the results are concrete: 

• Approximately 20% RAN energy conservation delivered through AI-driven closed-loop control.
• Zero impact to customer experience KPIs.

This is a world-first. And for anyone running a telecom network, it deserves more than a passing glance.

Why Level 4 and why now

TM Forum's autonomy framework has five levels. Level 4 is where intent replaces instruction. These involve networks that don't wait for a human-authored script to fire on a predefined condition; they sense, decide, and act within guardrails set by operations teams. They continuously optimize operations toward preset business outcomes. 

To put it simply: below Level 4, automation is typically reactive. It handles what you anticipated. Level 4 handles what you didn't.

The human moves up the loop from firefighter to architect.

That distinction is not semantic. For years, operators have layered scripts and condition-based triggers onto networks and called it automation. Some of it is genuinely useful. But it's also brittle – every new band, every new slice, every new site adds to the maintenance burden rather than being absorbed by the system itself. That is the operational ceiling the industry has been bumping against.

Energy is the business case that makes this urgent

Energy has quietly become one of the most consequential line items on a network P&L. GSMA analysis places it at 20-40% of network OPEX for most operators, and significantly higher in markets with diesel dependency or volatile grid pricing. The RAN alone accounts for upwards of 70% of mobile network energy consumption.

That concentration is actually good news. It means a material energy lever exists and it's located in a single domain. A validated ~20% RAN energy saving is not an engineering footnote. Run the arithmetic against your network's energy bill. It translates directly into OPEX reduction and Scope 2 emissions improvement, both of which now live on the CEO and board agenda, not just in the engineering org.

ESG commitments increasingly require operators to demonstrate measurable, auditable progress. A TM Forum-validated outcome on a live network is exactly the kind of evidence that moves from the sustainability report into the investor relations conversation.

What actually changes in operations

The most important thing to understand about Level 4 is what it does not mean. It does not mean removing engineers from the loop. It means changing what those engineers do.

Look at this scenario: In a traditional NOC, a quiet urban cluster at 2 a.m. is largely unused. In a Level 4 environment, the network anticipated that lull before it arrived, right-sized capacity in real time, and began ramping again ahead of morning traffic – without a single ticket raised, without a human-written conditional script, and without a degraded KPI report to review at shift change.

The shift in workload is significant. Operations teams move from:

• Alarm chasing to intent-setting: defining energy and performance policies rather than scripting individual responses to individual conditions.
• Fixed workflows to adaptive execution: the autonomy layer weighs energy, experience, resiliency, and performance simultaneously, in real time.
• Linear headcount scaling to non-linear complexity absorption: network density, spectrum proliferation, and service slicing grow without a proportional increase in operational load.

The architecture this requires – and why shortcuts are not an option

The Rakuten Mobile validation also makes an architectural argument the industry needs to hear plainly: you cannot achieve Level 4 by bolting an AI layer onto a legacy, vendor-locked network.

This is not a theoretical concern. The operational requirements of closed-loop autonomy at scale demand specific foundations:

• Cloud-native, disaggregated network functions that expose granular telemetry and real control points to an external autonomy layer. Black-box network elements cannot be autonomously managed; they can only be manually reconfigured.
• Open RAN architecture with standardized interfaces, enabling AI/ML logic through rApps and xApps on a RAN Intelligent Controller without locking into a single vendor's optimization logic.
• A high-fidelity digital twin – not a dashboard, but a real-time simulation environment where every autonomous action is validated before hitting the live network. Fixing a RAN issue that inadvertently creates a Transport bottleneck is not a theoretical edge case. It is what happens without this guardrail.
• A semantic data layer that harmonizes data across RAN, Core, and Transport. An xApp and a Core NWDAF that speak different data languages will break autonomy at the domain boundary. A Common Information Model is the connective tissue that makes true cross-domain closed loops possible.
• Explainability and deterministic fallback baked into the architecture. The AI must output its reasoning. The network must know safe states and revert to them instantly if model behavior drifts.

Without this foundation, what operators are left with is a collection of automation projects – valuable in isolation, but nowhere near Level 4 in aggregate.

Case in point: What is already running in production

The Rakuten Mobile implementation is worth understanding at a functional level, because it is not just a reference architecture.

Cloud-native Open RAN is the substrate, with disaggregated functions providing granular cell-level telemetry across the national network. The Rakuten Symphony platform, including an advanced OSS and RAN Intelligent Controller, operates as the decision and execution layer. AI/ML-driven rApps continuously analyze KPIs, forecast cell-level traffic patterns, and drive real-time energy decisions: dynamic capacity adjustment, safe resource idling during low-traffic windows, and automatic restoration as demand returns.

The closed loop runs end-to-end against live traffic. There is no human in the middle of each decision cycle. The humans defined the intent and set the guardrails. The system executes.

What operators should take from TM Forum’s validation

The industry's autonomy argument has shifted. It is no longer whether this is achievable; Rakuten Mobile has answered that. The question is now how fast operators can position themselves to replicate it.

That requires honest answers to three internal questions. 

• Do you have architectural visibility and control across your network domains, or have you ceded that to vendors? 
• Is your data coherent enough (across RAN, Core, and Transport) to support cross-domain AI decisions? 
• And does your operations organization have the skills and governance model to work alongside an intent-driven autonomy fabric?

Technology is rarely the constraint. Mindset and architecture are.

For Rakuten Mobile, Level 4 autonomy is an operational reality delivering energy savings, OPEX reduction, and a new model for how networks run themselves. For everyone else, speak to us at Rakuten Symphony to achieve the same for you.

If you want to discuss what this means for your network architecture, connect with us at MWC 2026.