Telcos Talk AI. O2 Shipped One.

O2 Telefónica Germany has deployed an AI agent to help its network engineers do the job that used to require flipping through decades of tribal knowledge and hoping the senior tech was available.

The system, called NOA (Network Operations Agent), is a generative AI assistant that analyzes network data and delivers structured recommendations for troubleshooting and fault resolution. O2 Telefónica announced it March 19, 2026. Unlike the steady drum of telco AI announcements that amount to a demo and a press release, this one has a named executive, a deployment date, and a place in a documented autonomy roadmap.

"Artificial intelligence significantly reduces the time between detection and action for our network engineers," said Mallik Rao, chief technology and business customer officer at O2 Telefónica, in the original German-language announcement. "NOA shows how humans and artificial intelligence can successfully collaborate to ensure stable network operations for consumers and businesses."

NOA was developed in-house. That matters. Most telco AI deployments lean on a vendor's name and call it a partnership. O2 built this one itself, using GenAI technology, and is running it as a daily tool for engineers and service managers. The system answers natural-language queries about network events, suggests diagnostic steps based on historical cases, and can be used to train new hires in multiple languages. It integrates into the operator's broader ticket-handling workflow, with the explicit goal of increasing the share of incidents processed automatically.

The architecture is worth examining. NOA sits inside what O2 calls its Autonomous Network Journey — a program dating to 2021 that targets TM Forum Level 4 autonomy by 2030. Level 4, per the Telefónica press release, means systems that act autonomously based on human-defined intentions, with engineers stepping in only for complex or novel cases. The operator finished 2025 with 12 Level 4 use cases already operational, including operations analytics and diagnostics automation in Germany, Brazil's AI-driven 5G core self-healing, and AI-based energy optimization that cut electricity costs in the antenna network by roughly 10 percent. Fiber planning time dropped from 60 days to under a week. ATEA, an internal analytics platform O2 has been building since 2011, now handles thousands of manual tasks automatically.

None of this is science fiction. None of it is autonomous either — at least not in the way the roadmap implies.

NOA recommends; a human decides. That distinction — between AI-assisted and AI-autonomous — is where most telco AI deployments actually live today, even when the roadmap language says otherwise. The gap between what NOA does now and what Level 4 promises is the structural story of this entire category.

The vendor layer underneath doesn't simplify the picture. O2's parent company Telefónica has a separate initiative with Tech Mahindra and NVIDIA: a Large Telco Model, trained on network data using NVIDIA AI Enterprise, NeMo, and NIM microservices, announced at GTC in March 2025. Microsoft has its own NOA Framework — distinct from O2's NOA — which uses an open-source, MCP-based multi-agent architecture with supervisor and specialist agents, TM Forum-standard trouble ticket APIs, and governance controls. Taiwan's Far EasTone, using that Microsoft stack, is running roughly 60 percent of its network operations center as AI-assisted with about 10,500 tasks executed per month. Vodafone, using Google Cloud's Autonomous Network Operations framework, cut repair times by 25 percent — the clearest verified benchmark in this comparison.

Those numbers — from comparable operators, not projections — are the benchmarks that matter when evaluating whether a deployment is real. O2 hasn't published equivalent metrics for NOA specifically. What it has is a shipped product, a named customer (its own engineers), and a place in a documented roadmap with quantifiable results in adjacent areas.

Telefónica and Mavenir, a network software provider, also announced a joint AI innovation hub for core network orchestration, positioned as a real-world testbed for AI-driven autonomous operations. Location is unconfirmed; presumed Spain. A different initiative from NOA, but part of the same directional bet.

Andrea Folgueiras, Global CTIO at Telefónica, presented the autonomous network use cases at MWC on March 3, 2026 — two weeks before the NOA announcement went out. The sequencing suggests the company wanted the MWC visibility to frame the deployment.

The telco AI agent space is thick with announcements and thin on operational detail. O2's NOA is a genuine deployment, not a pilot with a landing page. The question that remains open — as it does for nearly every operator in this space — is whether the metrics will follow. Energy savings, ticket automation rates, and time-to-resolution figures for NOA specifically haven't been published. The ATEA platform's track record and the Level 4 use cases are the best proxy evidence that this infrastructure works. Until NOA has its own numbers, the story is the architecture and the intent — not the outcome.