At FutureNet World 2026, Nilmar Seccomandi David, Director of Autonomous Network & Infrastructure at Telefónica, gave an interesting update on the operator's journey towards Level 4 autonomous networks.
The presentation, “Autonomous Operations: Why the Journey to AN L4 Depends on It?”, went beyond the increasingly familiar discussion about using AI in telecom networks. The more important message was that Level 4 is not something an operator achieves simply by deploying enough AI use cases. It requires a much broader transformation of network technology, operational platforms, processes and people, together with a way of measuring autonomy across the whole organisation.
Telefónica launched its Autonomous Network Journey (ANJ) programme in 2021 with its three main operating companies in Spain, Germany and Brazil. David explained that it adopted a holistic approach from the beginning, covering the main network lifecycle processes of planning, testing, deployment and operations, as well as domains including IP, transport, radio, fixed access, core and cloud.
This is important because Telefónica does not view the autonomous network programme as a collection of isolated automation projects.
The programme has four main workstreams. The first is technology transformation, including more open, softwarised and disaggregated networks as well as the removal of legacy technologies. The second is what Telefónica calls the brain, its automation platform, covering data integrity, data lakes, OSS modernisation and the platforms required to implement automation and AI.
The third is the heart, involving process redesign. David's point here was that simply automating an existing process does not necessarily capture the full benefits of automation. Processes themselves may need to be redesigned as new capabilities are introduced.
The fourth is people, including new ways of working, reskilling and creating an AI culture.
This framework is not new. In his FutureNet World presentation a year earlier, David used essentially the same four dimensions, describing them as The Network, The Brain, The Heart and People. The 2025 material already included open and softwarised networks, data and automation platforms, process redesign and organisational change.
That earlier presentation also showed Telefónica moving from conventional automation towards what it called Hyper Automation, with GenAI, agentic AI, cognitive cross-domain AI and digital twins becoming increasingly important between 2025 and 2030.
The difference in 2026 is that the discussion has become much more focused on how the operator measures progress and what autonomous operations actually look like in production.
Telefónica estimates that when the programme started five years ago, its overall autonomy level was around 1.1. It has progressed through the maturity levels year by year and closed 2025 at 3.42.
The company has now made its longer-term objectives public. It is targeting an average autonomy level of 3.75 by 2028 and Level 4 by 2030 across Spain, Brazil and Germany. Those targets formed part of Telefónica's November 2025 Capital Markets Day and its wider Transform & Grow strategy.
David made an interesting observation about the importance Telefónica now attaches to the number: autonomy level has become a KPI followed by the company's board, alongside conventional indicators such as revenue and Net Promoter Score.
That raises the obvious question: what exactly does a figure such as 3.42 mean?
This was perhaps the most useful part of the presentation.
Telefónica uses the TM Forum Autonomous Networks framework as its basis, but David stressed that operators have to adapt the methodology to their own networks.
At the top level, Telefónica's Group autonomy figure consolidates the results from Spain, Brazil and Germany.
Within each operating company, the network is divided into domains such as IP, transport, radio, fixed access, core and cloud. Those can then be divided into sub-domains. Transport, for example, may include optical and microwave, while the core can include voice and packet-core functions.
Each domain is then assessed across processes such as planning, testing, deployment and operations. Those can be broken down again into individual activities. Operations, for example, includes areas such as fault and performance management.
The result is a large hierarchy of measurements which eventually rolls up into the Telefónica autonomy figure.
David's argument was that this complexity is necessary. Knowing that an operator has one, ten or even twenty impressive autonomous scenarios tells you relatively little about the autonomy of the network as a whole.
The same point was already visible in Telefónica's 2025 reporting. Its autonomy index showed progress occurring at different speeds across six network domains and multiple processes. IP and RAN were among the more advanced areas, while Core and Telco Cloud were further behind.
The 2026 results make that variation particularly clear.
The slide shows the results in two different ways.
On the left are Telefónica's three operating companies. The countries are deliberately anonymised, but the data is real. IP is the most autonomous domain in all three operating companies, while Cloud is the least autonomous.
The right-hand view breaks the assessment down by process. Even inside IP, which is the most mature overall domain, testing is less autonomous than the other main processes.
This illustrates one of the difficulties with discussions about “reaching Level 4”. An operator does not suddenly become Level 4 everywhere. Different countries, technologies and processes progress at different rates, and the overall figure can hide significant differences underneath.
Telefónica closed 2025 with 12 individual Level 4 use cases, spread across Spain, Germany and Brazil. Two were associated with planning, five with deployment and five with operations. Telefónica describes Level 4 cases as those capable of acting autonomously based on an intention provided by a human.
These include autonomous network-capacity creation, transport digital twins, autonomous IP fault resolution, 5G Core self-healing, fibre planning, software changes and multi-domain correlation.
The first Level 4 operations example David discussed was NetOptimizer, a digital twin of the O2 Germany transport network.
NetOptimizer maintains an end-to-end representation of the German transport network and is used for analysis, simulation and bottleneck detection.
One particularly interesting capability is resilience testing. David explained that the system can periodically simulate the failure of individual links across the thousands of links in the transport network and analyse what would happen elsewhere.
Potential bottlenecks can therefore be identified before an actual failure exposes them.
The benefits shown on the presentation slide include 80% less time required for analysis, 40% fewer transport capacity issues, more than 90% fewer sites with very high capacity loading and a 5% latency improvement.
Telefónica has separately published the same figures for NetOptimizer, confirming the 80% reduction in time spent on planning, operations and network-optimisation analysis.
This is also worth noting because the automatic transcript of the presentation renders the first figure as 8%; the slide itself and Telefónica's published material make clear that the figure is 80%.
The second Level 4 example addressed IP interface flapping, where an interface repeatedly transitions between active and inactive states because of problems such as fibre attenuation, hardware faults, temperature or other instability.
Telefónica's system detects the condition, attempts to establish its cause and applies corrective action automatically. Where the problem cannot be resolved automatically, the affected port can be blocked and the appropriate field process triggered.
The company says the solution has reduced the impact of flapping on services by 70%, while removing the need for manual intervention in the closed-loop resolution process.
David then presented several other examples which are not yet Level 4 but are interesting because they demonstrate that Telefónica is not relying on a single flavour of AI.
One is Correlax in Brazil.
The starting problem was surprisingly mundane: NOC staff complained about gaps in the network inventory, which made it difficult to determine relationships between apparently separate incidents.
Rather than waiting for a perfect inventory, Telefónica applied graph techniques similar in principle to those used to establish relationships in social networks. It inferred relationships between network elements from available information and then used these relationships to correlate trouble tickets.
David said Correlax is reducing the relevant tickets by around 44%.
Another example is ATEA, an AI Factory deployed in Germany in collaboration with Google and using Gemini. One of the scenarios David described was cell-health analysis.
What had previously been a manual process involving the correlation of KPIs from multiple databases can increasingly be automated, bringing together information from different sources to analyse the condition of the network.
David also discussed rApps as a Service. Telefónica is testing an Ericsson solution using AWS which, according to the presentation, could be deployed in only a few days. The initial application involved anomaly detection, with additional agents being developed.
These examples are useful because the FutureNet World agenda specifically asked about the roles of predictive, generative and agentic AI, but the presentation suggests that this may be the wrong way to think about autonomous networks.
Telefónica is using whatever technique is appropriate to the problem: conventional automation, traditional algorithms, machine learning, graph analytics, digital twins, GenAI, rApps and increasingly AI agents.
The 2025 FutureNet World presentation made the same point in a different way. Telefónica already had more than 400 use cases spanning AI-driven network design, digital twins, capacity forecasting, GenAI document analysis, predictive maintenance, incident correlation and automated optimisation.
Its main case study that year was Vivo's Fractal system, which automated network creation using a mixture of DBSCAN clustering, Coral Reef algorithms, Telefónica's internally developed House of Cats algorithm and Dijkstra's shortest-path algorithm. The accompanying TelcoTitans report said Fractal had helped reduce mobile-site deployment time from around three months to one week.
The progression between the two FutureNet World presentations is therefore quite revealing. In 2025, the centrepiece was an advanced individual planning and network-creation use case. In 2026, the discussion is much more about spreading autonomy through the wider operating model and measuring that transformation systematically.
One of the most interesting parts of the 2026 presentation was David's reminder that AI by itself is not always enough.
He used RAN energy consumption as an example.
Traditional network power-saving features can already deliver substantial reductions. David suggested savings of around 20–30% compared with a network that does not use them.
AI can then improve when and how those features are activated. Depending on the implementation, David suggested that another 5–10% might be achievable.
Eventually, however, optimisation runs into the physical characteristics of the installed equipment.
At that point, modernising the hardware can potentially produce another significant improvement. David suggested that combining power-saving features and AI with newer hardware could yield an additional 20–30% and allow the energy-consumption curve to track the traffic curve much more closely.
The precise savings will clearly vary between networks, configurations and equipment generations, but the underlying message is important:
software intelligence cannot indefinitely compensate for inefficient hardware.
This is a useful counterweight to the idea that every network problem will eventually be solved by more sophisticated AI.
The final example in the presentation brought service observability into the autonomous-operations picture.
Traditional network operations tend to concentrate heavily on network KPIs. David argued that improving the network view is not sufficient: operators also need to understand the service view.
The example combines network status with service information and external factors such as weather conditions. Problems can then be analysed not only according to what is happening inside the network, but according to which services and customers are actually affected.
David said this approach is reducing Mean Time To Repair (MTTR) by around 30% and helping Telefónica prioritise which field tickets need attention first.
This may ultimately be one of the most important requirements for Level 4. A genuinely autonomous network should not merely know that a KPI has crossed a threshold or that a network element has failed. It needs enough context to understand what that failure means for the service, decide its relative importance and take the appropriate action.
There has already been further progress since the FutureNet World presentation.
Recently Telefónica provided a newer update saying it now has more than 500 AI use cases in production, with 15 already at Level 4. The targets remain Level 3.75 in 2028 and Level 4 in 2030.
That update reinforces the main message I took from David's FutureNet World presentation.
Level 4 is not something an operator installs.
It is the result of progressively changing how the network is designed, observed, deployed and operated; improving the data and OSS platforms underneath it; redesigning processes around closed-loop operation; and changing how people interact with increasingly autonomous systems.
AI is becoming increasingly important to that journey, particularly as GenAI and agentic approaches mature. But Telefónica's experience also shows that autonomous networking is much broader than AI.
It includes digital twins, conventional algorithms, closed-loop automation, observability, modern network architectures, accurate data, process redesign and, where necessary, replacing the physical infrastructure itself.
Perhaps that is why the title of the presentation was so appropriate.
The journey to Autonomous Network Level 4 ultimately depends on autonomous operations.
Related Posts:
- Operator Watch Blog: Telecom Argentina's View of Autonomous Networks and the Road to Level 4
- Operator Watch Blog: Why Operators Must Simplify Before They Automate
- Operator Watch Blog: Automation and Data Driven Network Optimization in Swisscom’s Mobile Strategy
- Operator Watch Blog: Telefonica’s Journey Towards End-to-End Autonomous Networks
- Operator Watch Blog: What Europe’s Five Greenest Telecom Groups Are Doing Differently
- Operator Watch Blog: Telefonica's Vision of Open RAN from MWC 2022
