AI-RAN Meets Agentic Networks: What the TM Forum’s Governance Push Means for Open RAN Operators

Agentic AI is moving from demo reels to production networks. In the last week of September 2026 alone, TM Forum called for controls on telecoms AI agents, Cisco and Omdia published research on agentic AI taking direct administrative control of enterprise networks, Verizon described its shift from scripted RAN automation to agentic network operations, and KT revamped its customer app with 14 AI agents. Meanwhile, Ericsson’s CTO went on the record about rebuilding cellular infrastructure around AI-native architecture to support agent traffic as machine-to-machine interactions outpace human mobile data consumption.

For an industry that has spent five years arguing about open fronthaul, the sudden arrival of autonomous software agents inside the RAN is a different kind of disruption. Skeptical executives have legitimate questions. Here are the ones that matter, answered from what the research actually shows — and what it doesn’t.

Q1: What exactly is an “agentic” network, and how is it different from the automation we already have?

The distinction is control. Traditional RAN automation is scripted: a SON function detects a condition and executes a pre-defined rule. Agentic systems reason across domains, share context, and take action — which is precisely how Verizon described its own transition in the RCR Daily roundup published October 2, 2026. The operator framed the shift as moving from scripted RAN automation to agentic network ops, where AI reasons across domains, shares context and takes action, with the operator retaining control of intent, outcomes and guardrails.

That last clause is doing a lot of work. The operator keeps intent, outcomes, and guardrails. The agent gets the reasoning and the action. In practice, that means an agent might correlate a massive MIMO configuration issue, a transport latency spike, and a slice SLA breach, then propose or execute a remediation that no single scripted rule would have caught.

The TM Forum’s intervention — calling for controls on telecoms AI agents — suggests the industry body sees the governance gap clearly. The same week, Unryo demonstrated agentic observability for telecom through TM Forum Open APIs, correlating multi-vendor network alarms into diagnosed, root-cause incidents using TMF642 for inbound alarm management and TMF724 for outbound incidents, with TMF915 keeping AI agents governable and auditable. That’s the template: agents act, but under an auditable framework.

Q2: Is this real, or is it 2026’s version of the metaverse?

The evidence points to real, if uneven, adoption. Consider the concrete markers from the past week:

• Verizon is publicly describing agentic network ops as its operational direction, not a lab experiment.
• KT shipped 14 AI agents into a customer-facing app — a production deployment, not a trial.
• Cisco and Omdia research concludes enterprise networks are changing forever as agentic AI takes direct administrative control.
• Ericsson’s CTO is publicly preparing cellular infrastructure for AI agent traffic, a signal that the vendor expects agent-generated and agent-managed traffic to become a first-order design constraint.
• LF Networking’s ONE Summit 2026 program in Tokyo features dedicated tracks on agentic and autonomous telecom networks, including Model Context Protocol (MCP), OpenAN, Nephio, and StratoWeave.

That’s a broad set of independent actors — an operator, a vendor, an analyst firm, a standards body, and an open-source foundation — converging on the same thesis within days of each other. It doesn’t prove ROI. It does prove the industry is past the “is this a thing?” stage.

Q3: Where does the RAN fit? Isn’t agentic AI a core-and-OSS story?

The RAN is where the physics and the economics collide, which is why it’s the hardest and most consequential domain. AI-RAN — running AI workloads on RAN infrastructure and using AI to optimize the RAN itself — is already showing measurable gains. Nokia reported growing global momentum for AI-RAN, with operators across North America, Europe, Asia-Pacific, and the Middle East advancing from early evaluation to lab and live field trials alongside partners such as NVIDIA. The platform has delivered spectral efficiency improvements of more than 20%.

Ericsson and Vodafone Egypt announced commercial deployment of Ericsson’s AI-powered RAN Processor 6672, enabling AI-native operations directly at the radio site with up to four times higher processing capacity and more than 50 percent lower energy use. T-Mobile and Ericsson have tested AI-RAN on a live 5G Advanced network. Ericsson also added AI into RAN software for 5G network optimization.

Now layer agentic control on top of that. The RAN becomes not just an AI-optimized system but an AI-operated one. That’s where the TM Forum’s governance call becomes urgent rather than theoretical: an agent that can retune beamforming weights or reallocate spectrum across a multi-vendor Open RAN deployment is an agent that can cause a multi-vendor outage.

Q4: What’s the business case? RAN revenue is flat, so who pays for this?

This is the question every CFO asks, and the honest answer is nuanced. Dell’Oro’s assessment, cited in the RCR Daily roundup, is that RAN revenue will remain flat through 2030. Growth will be concentrated in AI RAN, massive MIMO, and cloud RAN, while private 5G adds new enterprise RAN spending, and 6G remains uncertain.

So the business case isn’t “grow the RAN line item.” It’s three other things:

1. Cost avoidance. A 50% energy reduction at the radio site, as Ericsson and Vodafone Egypt reported, is an opex story that compounds across tens of thousands of sites. Agentic optimization of energy, scheduling, and capacity can extend that.
2. Spectral efficiency as capex deferral. Nokia’s 20%+ spectral efficiency gain means more capacity from existing spectrum and hardware. In a flat-revenue market, that’s the difference between investing in new spectrum and sweating existing assets.
3. New enterprise revenue. Private 5G is becoming less about 5G and more about replacing or augmenting fixed connectivity, per the RCR Daily’s P5G update. Agentic operations make private networks more autonomous and therefore cheaper to run — which is what enterprise buyers actually pay for.

The Dell’Oro framing — growth concentrated in AI RAN, massive MIMO, and cloud RAN — suggests the market is already pricing this in. The risk is that operators over-invest in agentic capability before the governance and observability layers mature.

Q5: What are the biggest risks, and who’s warning about them?

SK Telecom, notably, is warning its fellow telcos not to repeat their 5G mistakes with AI-RAN. That’s a striking message from an operator that has been among the most aggressive AI-RAN adopters — it has signed AI-RAN contracts with Samsung alongside KT for Korea’s Hyper AI Network, a two-year, KRW 17.2 billion program testing private 5G and edge AI with industrial robots. When the early adopter starts cautioning the market, the market should listen.

The specific risks cluster into four areas:

• Vendor lock-in disguised as intelligence. If the agent layer is proprietary, the operator’s “control of intent” is nominal. TM Forum’s push for governable, auditable agents — and Unryo’s implementation of TMF915 for exactly that purpose — is an attempt to prevent this.
• Multi-vendor fragility. Open RAN’s promise is best-of-breed. Agentic AI’s risk is that an agent optimized on one vendor’s data model misbehaves on another’s. This is why the SMO interop work between Nokia and NTT DOCOMO matters, and why the E3 interface and dApp ecosystem conversations matter.
• Security. An agent with write access to the RAN is a high-value target. The industry has already seen OpenAI’s agent breach and agent outbreaks become case studies. The Open Secure AI Alliance and AI cyber defense discussions are the response, but the attack surface is growing faster than the defenses.
• Governance debt. If agents are deployed without audit trails, the operator cannot explain an outage to a regulator. TMF915-style governability isn’t bureaucracy — it’s the price of admission for production.

Q6: What about 6G? Is agentic AI-RAN a bridge or a distraction?

It’s a bridge, and the standards work is already underway. AI-RAN is shaping up to be the bridge technology that crosses the telecom industry into AI-native 6G by 2030, with commercial 6G deployment landing in the 2029-2030 timeframe. Researchers and operators want to experiment with AI-native networking before 6G exists, and they can already deploy AI-RAN capabilities on 5G or 5G Advanced infrastructure — learning what works, developing models, and building compute infrastructure that carries forward into 6G.

The 3GPP core work is where this gets concrete. Inside 3GPP SA2, the study phase has identified three candidate directions for the core network based on AI functionalities. Direction 1 — a separate AI domain for intent handling independent of the packet-switched network, with the AMF routing NAS signaling while AI fulfillment occurs via SBI — is backed by Nokia, T-Mobile USA, Verizon, NVIDIA, Deutsche Telekom, Apple, Qualcomm, Ericsson, and NEC. It requires complex UE AI domain client integration, standardizing application-layer intent clients and 3GPP AI domain clients in the device, which means modem redesign, core security procedure overhaul, and backward-compatibility planning.

That’s the tell. The agentic future isn’t just a RAN story — it reaches into the device and the core, and it will force hardware and standards decisions that operators need to influence now, not in 2029.

Q7: What should a skeptical operator actually do in the next 12 months?

Five concrete moves, grounded in what the research supports:

1. Instrument before you automate. Agentic observability — correlating multi-vendor alarms into diagnosed incidents via open APIs — is the prerequisite for agentic control. If you can’t explain what happened, you can’t let an agent decide what happens next.
2. Demand governability at procurement. Ask every AI-RAN and automation vendor how their agents are audited, what APIs expose their decisions, and whether they support TMF915-style governance. Make it a scoring criterion, not a nice-to-have.
3. Run AI-RAN field trials with an agentic lens. Nokia’s eight-operator field trial program, T-Mobile and Ericsson’s live 5G Advanced AI-RAN test, and Ericsson’s RAN Processor 6672 deployment at Vodafone Egypt are all generating data. The question to ask isn’t just “what spectral efficiency did we gain?” but “what would an agent have done differently, and could we have stopped it?”
4. Treat energy as the first agentic use case. The 50% energy reduction at the radio site is measurable, uncontroversial, and low-risk. It’s the ideal beachhead for agentic operations because the failure mode is a slower network, not a broken one.
5. Engage the standards bodies now. The 3GPP SA2 Direction 1 debate, the TM Forum agent governance work, and the ONE Summit’s agentic networking tracks are where the rules are being written. Operators that show up late will inherit architectures they didn’t choose.

Q8: What’s the honest bottom line?

Agentic AI in the RAN is real, early, and uneven. The technology works well enough to deploy in observability and energy optimization today, and well enough to trial in live 5G Advanced networks. The governance frameworks — TM Forum’s controls, TMF915-style auditability, open API-based observability — are maturing in parallel, which is the right order of operations.

The risk isn’t that agentic AI-RAN fails. The risk is that it succeeds vendor-by-vendor, in proprietary silos, without the multi-vendor interoperability and auditability that Open RAN was built to deliver. SK Telecom’s warning about repeating 5G mistakes is really a warning about that outcome: deploying intelligence that optimizes each vendor’s domain while fragmenting the operator’s ability to manage the whole.

For Open RAN operators, the opportunity is to insist that agentic intelligence be as open, observable, and governable as the interfaces that carry the traffic. That’s not a philosophical position. In a flat-revenue RAN market where growth concentrates in AI RAN, massive MIMO, and cloud RAN, it’s the only way to capture the upside without surrendering control of the network to the agents running inside it.

Sources

• Starlink middle ground | Physical AI impacts | New network corridors | October 2, 2026 RCR Daily
• Enterprise IT & Communications Technology Industry News | Pipeline Magazine
• TM Forum calls for controls on telecoms AI agents
• For 6G, Keep an Eye on AI-RAN Technology | Microwaves & RF
• AI Takes Center Stage at Open Networking & Edge (ONE) Summit 2026 Program, Hosted by LF Networking
• The 6G core divide | Light Reading