The telecoms debate around AI and autonomous networks is becoming polarised. Yet both camps may be missing the central issue: context.
Sceptics remember earlier waves of automation. Expert systems, self-organising networks, and predictive analytics promised operational change. Many became unused dashboard features. Their caution feels understandable.
Meanwhile, believers argue that scale will solve the problem. They expect data lakes, large models, and clever prompts to deliver intelligence. That sounds attractive, especially for complex telecom operations.
As João Miguel Antunes of Celfocus wrote, “I disagree with both. And I disagree with them for exactly the same reason.” He added that “virtue lies in the middle.”
The core issue is not model power alone. Networks need a living view of themselves. Operators need to know which resources exist. They also need to know how those resources depend on each other.
This is where the Digital Twin becomes important. It should not mean a flashy 3D control room visual. Instead, it should act as a live map. It links sites, devices, services, alarms, and customer impact.
Consider a power failure at one mobile site. The operations centre may see several separate events. Environmental alarms appear first. Then radio cell alarms arrive. Transmission losses may follow from another vendor domain.
Without context, teams may open several incident records. With topology, the system can trace one cause. It can connect distant symptoms to one failed site. That changes response speed and field activity.
This also clarifies the role of an agent. The agent should not guess the network. It should query the Digital Twin and service history. Then it should test likely causes and explain its reasoning.
That explanation matters. Engineers will not trust a black-box answer during an outage. A system must show its causal path. Otherwise, it risks producing confident but wrong conclusions.
However, this approach needs hard groundwork. Operators must clean and align data from many sources. They must reconcile different names for the same asset. They must also keep the network model current.
That work rarely shines in product demos. Still, it decides whether autonomy works in production. Large models cannot replace poor identity mapping. More data cannot create causality by itself.
The path to autonomy also needs discipline. Networks supporting critical services cannot jump straight to full automation. A safer model lets systems recommend actions. Humans decide, and every decision becomes training data.
Therefore, success should not start with automation percentages. Better measures include faster incident detection and clearer root-cause identification. Reduced field interventions also show real value.
The lesson extends beyond telecoms. Energy, utilities, and banking face similar dependency challenges. Wherever assets interconnect, automation first needs context.
For network teams, the key question is simple. Do not ask only which model powers the agent. Ask what it knows about the network. Then ask where that knowledge comes from, and how often it updates.
