Discussions about artificial intelligence in telecommunications typically pit two groups against each other, yet both are addressing the wrong underlying issue.
The first camp consists of sceptics who have witnessed the rise and fall of expert systems, self-organising networks, and predictive analytics. Each promised autonomous operations, only to become another unused dashboard. Their verdict: fresh terminology masking old failures.
The opposing camp—believers—offers an appealingly straightforward vision: aggregate network data into a lake, feed it to a sufficiently capable model with a well-written prompt, and intelligence will materialise. When it does not, they attribute the gap to insufficient data or models that "are not quite there yet."
Both perspectives contain a symmetrical flaw, rooted in the same misunderstanding.
Two symmetrical mistakes
Sceptics judge technology through failed deployments and conclude the problem was excessive ambition. It was not. Those initiatives collapsed because they attempted to reason about networks that neither any system nor any person could fully characterise.
Believers, conversely, presume that scale and model power can eliminate the need for structured knowledge and context. They cannot. An agent processing forty thousand overnight events can identify the most frequent ones; it cannot recognise that dozens or hundreds are temporally, topologically and factually linked as a single underlying incident.
Frequency and temporal correlation do not equal causality. No volume of raw data paired with a model containing billions of parameters can convert one into the other. A data lake without relational structure is merely a vast repository, not context. Even if such an approach succeeded, what would be the cost in processing tokens? Autonomous systems should not simply shift operational expenses from people to token consumption.
The shared error: both sides concentrate on technology, agents and large language models, treating network representation as a secondary engineering concern to address later. The actual priority is inverted.
What is missing is not intelligence, but context
A single alarm indicates "something has failed at this location." It reveals nothing about what depends on that location. That knowledge exists scattered across inventory systems, vendor platforms, and the experience of three engineers with fifteen years at the company. What does not exist is a unified, current, and queryable representation of the network: which resources are present, how they connect, and what relies on what.
This model is termed a Digital Twin. A clarification is needed: a Digital Twin is not a three-dimensional visualisation for an operations centre, nor a system simulating every aspect when inference is unnecessary. It is a graph—nodes and relationships—maintained current from authoritative sources, where questions like "which services fail if this site or device loses connectivity or power?" are answered through graph traversal rather than through a war-room meeting.
A medium-sized national network with a few thousand radio sites and transmission nodes generates a graph containing hundreds of thousands of nodes and half a million relationships. Technically straightforward. The challenge lies in keeping it accurate—where most programmes falter.
An example every Operations Director will recognise
Power fails at a site. The operations centre receives three seemingly separate alerts: an environmental alarm, roughly a dozen alarms from radio domain cells, and signal loss across all transmission nodes dependent on that site—potentially twenty kilometres away, possibly from another vendor.
Three domains. Three teams. Three incident queues. Ideally, three separate incident records for one event.
Conventional correlation using time windows and text matching performs poorly: it groups simultaneous events and separates causal ones. Correlation via topological traversal identifies a single incident with a root cause. The agent did not become smarter; it gained a contextual map.
So, what is the role of an Agent?
Precision matters: "agent" has become a term encompassing everything and therefore nothing.
A rules engine manages known scenarios effectively. What it cannot perform is investigation—determining which question to pose when an initial hypothesis fails. That is where an agent's specific value lies: not merely generating text, but querying the Digital Twin context and historical data sequentially, forming a causal hypothesis, testing it, rejecting it, suggesting the next action and explaining the reasoning.
Notice the sequence: the agent queries context within the Digital Twin. It does not replace it or speculate about it. An agent without context becomes an eloquent amnesiac, capable of producing impressive hallucinations for teams expecting precision and rapid response. Conversely, a Digital Twin without intelligence or agency layered atop it remains merely an engineering accomplishment and an excellent archive that nobody accesses at three in the morning—and therefore an operational burden.
The uncomfortable implications
First: roughly four-fifths of the work is data engineering. Lossless data ingestion, reconciliation of incompatible schemas, collection of temporal metrics at tight intervals, handling traffic peaks, and managing flapping alarms. None of this features in an autonomous-system demonstration, yet each element, when done correctly, determines whether the system functions in production.
Second: the core challenge is identity, not algorithms. The same physical resource carries different names across different systems. Reconciling and mapping these identities, ontologies, and the rules governing relationships between concepts constitutes the actual project. Those treating this as preliminary work discover too late it was the central task.
Third: without an explicit causal chain, adoption fails. An agent declaring "the root cause is site X" without demonstrating how it reached that conclusion will not earn an on-duty engineer's trust—and should not. Explainability is not an abstract requirement; it is what permits action on the result.
Autonomy is not bought, it is earned
The industry adopted the TM Forum autonomy levels as terminology and overlooked that they are levels. Leaping directly to fully autonomous operation in networks serving emergency services is a logical error.
The credible route moves slower: the system proposes, the human decides, and every decision is logged. What believers frame as a temporary constraint is actually the training process itself. Every accepted or rejected diagnosis becomes a labelled example no public dataset can supply.
Success metrics shift as well. The first honest measure is not "percentage of automation"; it is mean time to detect multi-domain incidents, incidents with automatically identified root causes and failure points, and reductions in field work—metrics that cannot be inflated through polished presentations.
The question that actually matters
When someone demonstrates an agent for network operations, do not ask which large language model powers it. Instead, ask four questions: what should this agent understand about my network, where does that knowledge originate, how frequently is it refreshed, and how can it show me the causal reasoning it followed?
If answers are unclear, align with the sceptic. If they are specific, the believer was correct—they were simply examining the wrong part of the system.
The autonomous network does not start with the agent. It begins at the least visually compelling moment: when the network finally understands itself.
Beyond telecommunications
These principles extend beyond telecom. The electrical grid faces identical challenges: it too is a graph, and the sector has produced contingency analyses for decades. The distinction is that those models were designed for planning, not real-time operations, and distributed generation has reversed flow direction and made portions of the topology essential to operations invisible. The Digital Twin becomes a prerequisite for any operation seeking to automate decisions.
Telecommunications made these obstacles most apparent. Yet the underlying principles are universal: wherever interdependent assets, scattered data and critical decisions exist—whether in Energy, Utilities, or Banking—the path to autonomy begins with establishing context.
