15,000 Actions an Hour: Why Telecom AI in 2026 Is About RAN Automation, Not Grand Claims

15,000 Actions an Hour: Why Telecom AI in 2026 Is About RAN Automation, Not Grand Claims

In telecom, the most convincing AI story in 2026 is not a keynote vision. It is what happens when a live network fixes problems fast enough that customers never notice and operators avoid sending engineers into the field. Nokia says stc’s commercial network executed 15,000 autonomous corrective actions per hour in 2025, alongside a 60% reduction in cell outages and a 10% increase in peak throughput. Those are the kinds of numbers that make sense to a finance committee because they point to fewer truck rolls, fewer degraded cells, and better use of spectrum during busy hours. (nokia.com)

That is the real shift underway as operators enter the 5G-Advanced phase. AI budgets are moving toward closed-loop automation: systems that sense network conditions, decide on an action, execute it, and verify the outcome without waiting for manual intervention. In practice, that means RAN parameter tuning, energy control, workload orchestration, and maintenance triage inside existing OSS environments, not broad “self-healing network” claims with little operational proof. GSMA Intelligence’s January 2026 market update said cost-saving deployments still account for the bulk of operator AI rollouts, even as some newer deployments aim at revenue. (gsmaintelligence.com)

This article looks at the narrower, more operational phase of telecom AI: where operators are getting measurable infrastructure outcomes in network optimization, predictive maintenance, churn reduction tied to network experience, and 5G-Advanced rollout efficiency. The core question is simple and unforgiving: which AI use cases survive contact with live OSS/BSS stacks, multi-vendor RANs, and operating margin targets?

Where spending is moving

The industry has spent years talking about autonomous networks. The measurable evidence now points to something more selective: AI is getting funded where it can trim OpEx, protect network KPIs, or accelerate 5G expansion with fewer human touches. GSMA Foundry summarized the sector’s position in 2025 by noting that 65% of operators had adopted an AI strategy and 41% were still in early-stage deployment, which helps explain why budgets are concentrating around a smaller set of use cases with clearer proof points. (gsma.com)

The shift is economic, not ideological. In mobile networks, the RAN, or Radio Access Network, often accounts for more than 80% of energy use, according to Ericsson’s March 2024 statement on its work with Chunghwa Telecom. If the most power-hungry and expensive part of the network can be tuned with AI to sleep, wake, or reconfigure more intelligently, the business case is immediate. (ericsson.com)

As a result, operators are increasingly treating AI in telecom as three separate investment classes: closed-loop RAN automation for capacity, coverage, and energy; predictive maintenance that reduces site visits, improves repair preparation, or catches degrading assets before outages; and customer-retention analytics only when the input data is tightly linked to network experience. The broader promise of end-to-end self-healing still appears in roadmaps, but the public evidence for scaled returns remains much thinner.

For operators building the 2026 business case, the cleanest filter is also the hardest: fund AI projects that sit inside an existing operational loop, change a cost line within twelve months, and can be measured against network KPIs rather than model accuracy alone. Anything else risks becoming an integration project disguised as innovation.

RAN automation is winning

Closed-loop automation has become the most credible AI story in telecom because it ties data, decision, and action to the network’s costliest constraints: spectrum, energy, and field operations. A closed loop is not just analytics. It is a control system in which software monitors live conditions, recommends or executes a change, and checks whether the intended result happened. In RAN operations, that can mean adjusting tilt, power, sleep states, handover settings, or congestion parameters at machine speed.

The stc case stands out for exactly that reason. In Nokia’s 2026 solution brief, the operator’s live network is presented not as a lab environment but as a commercial setting where MantaRay AutoPilot carried out 15,000 autonomous corrective actions per hour, cut cell outages by 60%, and lifted peak throughput by 10%. Nokia also says traffic during Hajj in 2024 grew 40%, while the automation system adapted on a 15-minute learning interval. Those are infrastructure numbers: outages, throughput, traffic load, and corrective actions, not vanity metrics. (nokia.com)

MasOrange offers a second signal, though the evidence still comes from a vendor-issued deployment announcement and should be treated with some caution. Ericsson said in December 2025 that MasOrange deployed its Intelligent Automation Platform, including SMO, or Service Management and Orchestration, and a mix of rApps for automated RAN optimization and energy efficiency in the commercial network. The important detail is not the press-release language but the architecture. Operators are increasingly using SMO and rApps, specialized radio applications designed to run on top of O-RAN-style management layers, because they let automation logic plug into operational workflows without replacing the whole OSS stack. Direct public performance data from MasOrange remains limited, but the deployment pattern fits a wider push toward modular automation rather than end-to-end rip-and-replace. (ericsson.com)

AT&T’s collaboration with Ericsson adds another piece of evidence. Ericsson says the operators’ network modernization program, tied to a USD 14 billion infrastructure investment, delivered up to a 20% annual gain in network energy efficiency while handling nearly double the network traffic through Open RAN, Cloud RAN, software changes, and AI-powered rApps. Because this is a vendor case study rather than a regulatory filing, it should not be read as an audited savings statement. Still, it is one of the clearer public examples of AI being folded into a broader infrastructure program instead of sold as a standalone miracle. (ericsson.com) (about.att.com)

The strongest 2026 use case, then, is not “AI-RAN” as a slogan. It is RAN automation attached to a defined control plane, a narrow set of high-frequency decisions, and a rollback path. If an organization cannot say which network parameter will be adjusted, how often, by which policy engine, and through which OSS interface, the project is not deployment-ready.

Energy has the clearest ROI

If one area already has enough evidence to shape board-level decisions, it is AI-assisted energy management. Operators do not need abstract sustainability narratives to justify it. Electricity is a direct operating cost, and 5G densification increases pressure on site power, cooling, and radio efficiency even before 5G-Advanced adds more complexity.

Telefónica’s work on energy optimization made that case early. The operator said in 2022 that testing Ericsson’s Radio Deep Sleep Mode in a 5G deployment in Madrid showed network energy consumption optimization of between 8% and 26%. The concept is straightforward: during low-traffic periods, software can shut down parts of radio equipment, carriers, sectors, or cells, then restore them when demand returns. AI and machine learning improve the timing so the network saves energy without visibly hurting customer experience. (telefonica.com)

Chunghwa Telecom reported even stronger results. Ericsson said in March 2024 that its AI-powered Service Continuity solution delivered up to 34% energy savings on Chunghwa’s network by predicting load and automating energy actions such as deep sleep mode. Again, this is vendor-supplied evidence, but it is operationally meaningful because it ties AI output directly to a network cost line. (ericsson.com)

Vodafone UK’s 2025 trial suggests where this class of use case is heading during 5G-Advanced rollout. Ericsson said the trial combined 5G Deep Sleep, 4G Cell Sleep Mode orchestration, and a Radio Power Efficiency Heatmap to predict and optimize energy consumption, with deep sleep saving up to 70% energy consumption during low-traffic hours for radios in the tested scenarios. Trial language matters: this is evidence of capability, not yet proof of scaled commercial savings across a national footprint. But it shows why operators keep funding energy apps before flashier AI ideas. The input data is abundant, the action set is clear, and the payback model is easier to defend. (ericsson.com)

The practical lesson is to treat energy automation as a sequencing problem, not a science project. Start with low-risk sleep and wake controls at sites with predictable traffic troughs, then connect them to guardrails such as dropped-call rate, accessibility, and congestion thresholds. The winner will not be the operator with the most AI models. It will be the one that can prove kilowatt-hour savings without dispatching engineers to clean up after a bad policy decision.

Maintenance must cut dispatches

Predictive maintenance has been marketed in telecom for years, often too loosely. In 2026, the useful distinction is between systems that produce interesting alerts and systems that actually reduce field work, improve repair quality, or prevent service degradation. The second group is where operators are finally showing evidence.

O2 Telefónica Germany offered one of the clearest public examples in June 2024. The operator said its intelligent analysis software helped avoid more than 10,000 technician trips in twelve months and saved an estimated 1 million kilometers of travel. The system pulls information from network databases, identifies the likely on-site fault, tells technicians what hardware and knowledge they need before leaving, and increasingly detects faults in advance as part of predictive maintenance. This is not glamorous AI. It is dispatch optimization tied to maintenance prediction, with obvious implications for labor utilization and site restoration times. (telefonica.de)

The operator’s TM Forum case study adds the broader operational frame: Telefónica Germany described the effort as a move from reactive maintenance toward predictive operations with closed-loop control, improving mean time to repair, or MTTR, and network availability. Publicly available detail on the exact magnitude of MTTR improvement is limited, but the direction is consistent with the dispatch result above. (info.tmforum.org)

Türk Telekom’s GSMA Foundry case provides another concrete example. According to the January 2026 case study summary, an AI and computer-vision system deployed across an initial 200 sites reduced daily manual inspection time from 14 hours to 15 minutes, a 98% reduction, while improving inventory accuracy and feeding predictive maintenance models with visual and operational data. Because the underlying page is not fully accessible through open fetch in this environment, practitioners should treat this as a summarized GSMA case rather than a fully inspectable primary document. Even so, it is notable because it links predictive maintenance to an operational bottleneck many operators still ignore: site-room visibility. (gsma.com)

Predictive maintenance should not be funded on the promise of “finding anomalies.” It should be funded when it measurably changes the economics of repair: fewer truck rolls, better first-time fix rates, shorter MTTR, or less unplanned outage time. If maintenance AI does not alter dispatch logic or spare-parts preparation, it is probably still an analytics pilot.

Churn only counts with network data

Churn reduction belongs in this discussion, but only in a narrow form. Too much telecom AI storytelling still treats churn as a marketing-automation problem, when in a mature mobile market it is often a service-quality problem in commercial clothing. The operators getting value here are not simply predicting who might leave; they are separating customers whose dissatisfaction can be fixed by an offer from those whose dissatisfaction is rooted in radio performance, repeated faults, or chronic local congestion.

Public evidence is thinner here than in RAN optimization or maintenance, and that thinness is revealing. TM Forum’s case study on Taiwan Mobile’s OSS transformation reported customer churn down 30% after the operator combined AI analytics, service automation, and OSS modernization. But the case dates to 2020, before the current 5G-Advanced investment cycle, and the result bundles multiple interventions together rather than isolating which share of the churn improvement came specifically from network-derived AI insight. It shows that operations data can influence retention, but it does not yet provide a clean contemporary benchmark for European or North American operators navigating multi-brand 5G portfolios. (inform.tmforum.org)

More recent operator disclosures are informative in a different way. Orange’s March 2026 strategic update said it aims to improve churn by up to 3 points in European countries by 2028, supported by AI digital assistants, loyalty mechanics, and a broader customer-experience program. That is a meaningful board-level target, but it remains a strategic ambition rather than attributable evidence. It also highlights a common reporting problem: once churn initiatives mix contact-center AI, offer management, digital servicing, and network-quality improvements, the causal chain becomes blurry. Investors may accept the aggregate objective; operators trying to allocate AI budgets should not. (orange.com)

For practitioners, the real dividing line is whether churn analytics merge BSS and OSS at the decision level rather than the dashboard level. BSS, or Business Support Systems, contain customer profile, tenure, billing, tariff, device, and offer history. OSS, or Operations Support Systems, contain alarms, topology, performance counters, ticket history, service degradation records, and quality-of-experience data by cell, cluster, or geography. A model built mostly on tenure, ARPU, and promotion response will usually optimize discounting. A model that also sees repeated failed handovers, dropped sessions, chronic low SINR, neighborhood outage exposure, or unresolved repair tickets can identify cases where the right intervention is not a retention credit but a network fix. That distinction matters because saving a customer with a discount while leaving the radio problem untouched simply moves cost from churn to margin erosion.

For operators that own both network and commercial budgets, churn reduction should be built around service-quality cohorts, not generic propensity scores. Separate customers into at least three buckets: those best addressed by an offer, those requiring a service intervention, and those in structurally weak coverage or capacity zones where retention spend should be contingent on a dated remediation plan. In many cases, the best retention action is not a discount. It is a RAN fix, a capacity add, or a maintenance dispatch targeted at the cluster causing the complaints.

5G-Advanced raises the bar

The arrival of 5G-Advanced is making AI less optional and less romantic. As operators introduce denser carrier configurations, more dynamic power management, uplink optimization, and increasingly software-defined RAN features, the decision count rises faster than headcount or engineering time. That changes the ROI test. The question is no longer whether AI can produce insight; it is whether it can absorb a growing volume of low-latency operational decisions cheaply enough to preserve 5G economics.

That is why serious operator spending is clustering around AI as a control-layer enhancement, not a moonshot architecture. Nokia’s March 2025 AI-RAN announcement with KDDI, SoftBank, T-Mobile US and other partners framed the initiative around practical goals including reducing 5G network-related costs and power consumption through automation. The point is not the consortium branding. It is that even the public language from major operators and suppliers is now anchored in cost, power, and operational automation rather than the older promise of fully autonomous networks arriving all at once. (nokia.com)

The friction is operational, not conceptual. In live networks, counters drift after software upgrades, topology changes during modernization, vendors expose different management interfaces, and every automated action must survive change-control policy and customer-impact guardrails. A model that looks impressive in a controlled trial can become economically unattractive once OSS integration, observability, rollback logic, and cross-vendor testing are added. That is why narrow closed loops are advancing faster than broad anomaly-detection and “self-healing” claims. The former can be tied to a bounded action set and a measurable KPI delta; the latter often collapse into expensive exception handling.

The likely outcome over the next 18 months is not a dramatic leap to generalized autonomy but a widening gap between operators that operationalize small loops well and operators that keep buying broad AI narratives. By late 2027, the strongest tier-one deployments will likely show routine automation in energy management, parameter optimization, and maintenance triage across commercial footprints, with adoption measured in recurring KPI movement and OpEx impact rather than demo count. Generalized self-healing will still exist mostly in bounded domains: a limited fault library, specific vendor environments, or tightly defined service classes. That is not a disappointment. It is what industrialization usually looks like when hype gives way to infrastructure economics.

Under 5G-Advanced, every AI network proposal should face the same capital-allocation test as any other infrastructure feature, with mandatory reporting on KPI movement, OSS/BSS integration cost, rollback design, operational ownership, and twelve-month OpEx impact. If a proposal cannot specify the decision loop, the target domain, the human override path, and the unit economics, it should not survive procurement. Back the systems that keep cells on air, technicians off the road, and 5G economics under control.