Let’s Talk
Continue to explore
Oops! Something went wrong while submitting the form.
|---Module:text|Size:Small---| As with almost everything in life, the debate around artificial intelligence in telecommunications is often divided into two opposing camps, and neither seems to be looking at the right problem.
On one side are the sceptics. They have already seen expert systems, self-organising networks, and predictive analytics come and go. Each wave promised autonomous operations and ended up as just another dashboard panel that no one opens. Their conclusions are often understandable: new vocabulary for an old disappointment.
On the other side are the believers. Their thesis is appealing in its simplicity: pour everything into a data lake, alarms, metrics, network inventory, incident history, among many other sources, connect it to a sufficiently powerful model, write a carefully crafted prompt, and intelligence will emerge. If it does not emerge, then there was not enough data, or the model “is not quite there yet, we need larger and/or more specialised models”.
I disagree with both. And I disagree with them for exactly the same reason. Here too, as in everything in life, virtue lies in the middle.
Two symmetrical mistakes
The sceptic evaluates the technology based on examples of failed implementations and concludes that the problem is ambition. It is not. Those projects failed because they were trying to reason for a network that no system, or ordinary person, could fully describe.
The believer, meanwhile, assumes and trusts that scale and model capability can replace any need to create structure and context. They cannot. An agent that reads forty thousand events overnight can tell us which are the most frequent; it cannot tell us that dozens or hundreds of them are temporally, topologically and factually correlated and are, in fact, the same incident or problem. Frequency or temporal correlation are not synonymous with causality, and no amount of raw data combined with a model containing billions of parameters can transform one into the other. A data lake without a relational model is not context; it is a large archive. And even if that strategy were to succeed, at what cost in processing tokens? The purpose of an autonomous system is not to convert people-related operational costs into token-related costs.
The common mistake is the same: both sides focus the discussion on technology (agents and LLMs) and on the model that powers it. The sceptic doubts it; the believer trusts it, and both treat network representation as an engineering detail to be solved later. The reality is exactly the opposite.
What is missing is not intelligence, but context
An alarm, in isolation, says “something has failed at this point”. It does not say what depends on that point. That information exists, in inventory, in each vendor’s systems, and in the minds of three engineers who have spent fifteen years in the company. What does not exist is a living, unique, and queryable model of the network: what resources exist, how they relate to each other, and what depends on what.
That model has come to be known as a Digital Twin, and it is worth dispelling a misunderstanding: I do not see a Digital Twin as a three-dimensional visualisation for the operations centre, nor as a solution that simulates every aspect of a system when we do not need to infer from them. It is a graph, nodes and relationships, rebuilt regularly from authoritative sources, where the question “which services fail if this site or device loses connectivity or power?” is answered through traversal rather than through a war-room-style meeting (yes, that is the name given to such meetings, and it is revealing in itself).
In a medium-sized national network, a few thousand radio sites (RAN) and transmission nodes produce a graph with hundreds of thousands of nodes and half a million relationships. Technically, this is trivial. The difficult part is keeping it true, and that is where almost every programme runs aground.
An example every operations director will recognise
Power fails at a site. What reaches the operations centre are three apparently distinct things: an environmental alarm, a dozen alarms from the radio domain cells, and signal loss in all transmission nodes that depended on that site, which may be twenty kilometres away and belong to a different vendor.
Three domains. Three teams. Three incident queues. And, with any luck, three different incident records opened for a single event.
Traditional correlation, based on time windows and text similarity, groups this poorly: it joins what is simultaneous and separates what is causal. Correlation through topological traversal returns to a single incident and a root cause. The agent did not become more intelligent; it simply gained a contextual map to navigate.
So, what is the role of an agent?
It is worth being precise: “agent” has become a word that means everything and therefore nothing.
A rules engine handles known cases well. What it cannot do is investigation: deciding which question to ask next when the first hypothesis proves wrong. That is the specific contribution of an agent, not merely generating text, but querying the context within the Digital Twin and the historical record in sequence, formulating a causal hypothesis, testing it, discarding it, proposing the next action and explaining its reasoning.
Notice the order of the words: the agent queries the context within the Digital Twin. It does not replace it, nor does it guess it. An agent without context is an eloquent amnesiac, capable of delivering the most impressive hallucinations to teams that expect precision and speed to act. Conversely, a Digital Twin without intelligence or agency layered on top of it is merely an engineering achievement and an excellent archive that nobody consults at three in the morning, and therefore an operational cost.
The uncomfortable implications
First: roughly four-fifths of the effort is data engineering. Lossless ingestion, harmonisation of incompatible models, collection of temporal metrics at tight intervals, handling peaks, and flapping alarms. None of this appears in a potential autonomous-system demonstration, yet each of these elements, when implemented properly, determines whether the autonomous system works in production.
Second: the real problem is identity, not algorithms. The same physical element has different names in different systems. Reconciling and mapping these identities, ontologies, and the laws governing relationships between concepts is the project. Those who treat this as preparatory work discover, too late, that it was in fact the main task.
Third: without an explicit causal trail, there is no adoption. An agent that states “the root cause is site X” without showing how it reached that conclusion will not be trusted by an on-duty engineer and should not be. Explainability is not an abstract requirement here; it is the condition that allows someone to act on the result.
Autonomy is not bought, it is earned
The industry adopted the TM Forum autonomy levels as vocabulary and forgot that they are levels. Jumping directly to fully autonomous operation in networks that support emergency services is a category error.
The defensible path is slower: the system proposes, the human decides, and every decision is recorded. What believers interpret as a temporary limitation is the training regime itself. Every accepted or rejected diagnosis becomes a labelled example that no public dataset can provide.
The way success is measured also changes. The first honest indicator is not “percentage of automation”; it is mean time to detect multi-domain incidents, incidents with automatically identified root causes and points of failure, and reductions in field interventions, metrics that cannot be inflated through a polished presentation.
The question that actually matters
The next time someone shows you an agent for network operations, do not ask which LLM it uses. Instead, ask four things: what should this agent know about my network, where does that information come from, how often is it updated, and how can it show me the causal path it followed?
If the answers are vague, side with the sceptic. If they are precise, the believer was right; they were simply looking at the wrong side of the system.
The autonomous network does not begin with the agent. It begins at the least photogenic moment of the entire process: when the network finally comes to know itself.
Beyond telecommunications
None of these are specific to telecommunications. The power grid shares the same problem: it is also a graph, and the sector has been producing contingency analyses for decades. The difference is that those models were built for planning, not for real-time operations, and distributed generation has reversed the direction of flow and rendered part of the topology essential to operations invisible. The Digital Twin is therefore the prerequisite for any operation that seeks to automate decision-making.
It was in telecommunications that these challenges became most visible. Yet the principles are universal: whenever there are interdependent assets, dispersed data and critical decisions, from Energy and Utilities to Banking, the path to autonomy begins with the creation of context.
