Telcos harness digital twin progress for network autonomy

A recent TM Forum survey of 128 communication service providers (CSPs) about their business and technology strategies to scale AI adoption highlighted the importance of digital twins in network automation: Fully 72% of respondents said they plan to use digital twins to enable autonomous networks and agentic AI, while 13% said they already do so.

Bertrand Decocq, Smarter Networks Program Manager, Orange, explained in an interview with TM Forum’s Insight what many CSPs are setting out to do: “We try to study all of this together: artificial intelligence, autonomous networks and digital twins. We want to solve the cross-domain situations [and address silos in] organizational and network domains.”

Traditionally digital twins have relied on periodically updated, siloed network data to create replications of physical network systems that CSPs can use for testing. Now, however, operators are working to build twins that keep up with the reality of the network.

Verizon, for example, is combining graph neural networks (GNNs), which are deep learning models that process data structured in graphs, and digital twins to provide a real-time network model. The aim is to use the digital twin to achieve Level 4 of network autonomy, as Abhitabh Kushwaha, AVP AI Networks, Verizon explained in a discussion with TM Forum’s CEO, Nik Willetts.

“Our static dashboards are not good enough. So, we are building world's largest digital twin in partnership with Google,” said Kuswaha. “In fact, we are creating a replica of our network, which is a network map connecting our RAN, transport and core network.”

Verizon worked with CSPs and suppliers as part of a TM Forum Catalyst to deploy “business-aware GNN healing networks”. The closed-loop architecture “detects silent degradation from a single service layer trigger, traces the contributing factors across the RAN core and a transport network, and then sequences remediation based on the business criticality, not just the technical severity,” according to Kushawa. GNNs, he explained, make sense because the “network is a graph. Cell connects to the transport. Transport connects to the core. Core connects to the cloud. When a fault propagates, it follows those edges. Message passing between nodes mirrors how a fault actually propagates through topology.”

Verizon’s use of GNN reflects an industry-wide acceleration in digital twin progress.

“It's evolving very fast – faster than I was expecting,” said Orange’s Decocq, pointing to the example of advances hyperscalers have made in graph databases both for real-time management and for analyzing large amounts of data in non-real time.

He also highlighted the rapid progress being made in the semantic analysis and querying of knowledge graphs, allowing CSPs to leverage the contextual data for decision making.

“We are really trying to leverage all the contextual data, all the semantic information in these [knowledge] graphs.”

Telia, for example, is building a semantic digital twin in partnership with NumoData, as part of its evolution towards an autonomous network, as Vincenzo Procopio, Head of IT Production & Cybersecurity, Telia, explained in a presentation at DTW Ignite 2026.

“AI needs to understand what that information actually means, so the semantic digital twin really is a mix of the relationships of the data, an understanding of this data, and also rules based on the knowledge of engineers embedded in that model as well,” explained Robin Osagie, Business Development Manager, NumoData, speaking alongside Telia’s Vincenzo.

It’s not only a question of enabling AI agents’ understanding. One of the challenges CSPs face is that probabilistic AI agents operating within deterministic network systems create new demands for verification and validation.

“A capable agent on a poorly observed, hard-to-reverse substrate is not an autonomous system. It is a faster incident factory,” according to Philippe Ensarguet, speaking at Telecom TV’s AI-Native Forum event. “It’s impossible if we are not setting the right boundaries. Without context, without memory, without digital twins, without ontology, without guardrails ... how we can trust the system we build?”

Amid rapid progress in technology, Decocq also emphasized the importance of ensuring digital twins deliver business value, including for network autonomy, which he described as a high-value scenario for digital twins.

“When it comes to the automation use case, we analyze the added value of AI or network digital twins. The idea is to use them only if it is relevant, especially AI, because it's not only costly but also energy consuming,” he explained.

Examples of where digital twins bring measurable value include impact analysis to prevent concomitant operations that result in service outages, explained Decocq. “For example, if we have an operation at the mobile core network and another in the infrastructure layer on two different routes, we use the digital twin to take into consideration the different layers, and to measure and to share this information about operations ... and to take into consideration alarms coming from the other layer or backup route.”

He added: “So, that's really bringing value because around 20% of the outages ... in the network are due to concomitant operations.”

Orange also weighs up network scale, operational complexity and cost-benefit considerations to decide whether to converge network domains in a single digital twin, or to have a system in which multiple digital twins interoperate.

For smaller OpCos it can make most business sense to combine IP and transmission networks within a single digital twin to support alarm correlation and cross-domain root-cause analysis, according to Decocq. For larger networks, interconnected twins for individual domains may provide a better return on investment. In either case, what matters is the business outcome. “It's the same value because they solve the cross-domain ... silos in terms of organization and network domains.”