Self-Optimising Networks: How AI Is Changing Telecom Operations

A modern mobile network layers 2G, 3G, 4G and increasingly 5G technologies on top of one another, each with its own configuration parameters, interference patterns and failure modes. Every cell site has antennas pointed at specific tilts, power settings, neighbour lists and handover thresholds. Traffic shifts by hour, by event and by season. Historically, armies of engineers tuned these parameters with drive tests and rules of thumb. That approach does not scale to the density and dynamism of today's networks.

At the same time, operator economics are under pressure. Data consumption keeps rising while revenue per user, particularly in many African markets, does not grow at the same pace. The GSMA Mobile Economy reports consistently describe this squeeze and the push toward automation and efficiency. Energy is a particularly painful line item in Nigeria, where many tower sites rely on diesel generators because grid power is unreliable. AI that reduces energy use or prevents a site outage has direct, measurable value.

Fault prediction and AIOps

Networks generate vast streams of alarms, counters and logs. A single fibre cut can trigger thousands of downstream alarms, burying the root cause. AIOps systems use clustering, correlation and causal inference to group related alarms and identify the likely origin. Time-series models on key performance indicators (KPIs) such as call drop rate, throughput and latency detect degradation before customers complain. Increasingly, LLM-based assistants help operations engineers query logs and runbooks in natural language.

Graph neural networks for topology

A radio network is naturally a graph: cells are nodes, and neighbour relations, interference and handovers are edges. Graph convolutional networks and their descendants can learn how a change at one cell affects its neighbours, which makes them well suited to predicting interference, traffic spillover and the impact of configuration changes.

Radio optimisation with reinforcement learning

Self-organising network (SON) functions automate tasks such as coverage and capacity optimisation, load balancing and antenna tilt. Reinforcement learning agents can learn policies for these by trial and error, usually first in a digital twin or simulator. Safe exploration is the central challenge: an agent cannot be allowed to drop calls while it learns.

Energy saving

Traffic at many sites falls sharply at night. ML models forecast cell load and switch off carriers, put radios into sleep modes or reduce power when demand is low, waking them up before the morning peak. This is one of the most widely deployed and least controversial uses of AI in telecom.

Open, programmable architectures

The O-RAN Alliance has defined an architecture with RAN Intelligent Controllers (RICs) that host third-party applications, so-called xApps and rApps, which consume network data and push control decisions. This turns the radio network into a platform where AI models can be deployed in a standardised way rather than locked inside a single vendor's equipment. Standards work at 3GPP has also been adding data collection and AI/ML frameworks into the specifications.

Large equipment vendors and operators have publicly reported using AI for site energy optimisation, predictive maintenance of power systems and automated alarm correlation. Customer-facing applications are also widespread: churn prediction, next-best-offer models and AI-assisted customer care in multiple languages. Fraud detection, including SIM-box fraud and subscription fraud, relies heavily on anomaly detection over call detail records.

The ITU's AI for Good platform has run machine-learning challenges on network problems, which is a useful sign that the field is opening up beyond vendor labs. For African operators, the most visible gains tend to come from tower power management, site availability prediction and smarter planning of rural coverage, where satellite imagery and population data help decide where new sites will serve the most people.

• Cascading automation failures. Closed-loop systems that change configurations across many sites can amplify a bad decision network-wide in minutes. Guardrails, rate limits and rollback are essential.
• Data quality and vendor silos. Different vendors expose different counters with different definitions. Models trained on one vendor's data rarely transfer cleanly.
• Privacy. Network data reveals where people are and who they talk to. Using it for analytics, especially location intelligence, requires strict anonymisation, purpose limitation and regulatory compliance.
• Security. Opening the RAN to third-party apps expands the attack surface. A compromised xApp with control privileges is a serious threat, which is why security has been a major topic in O-RAN discussions.
• Uneven benefits. If optimisation targets revenue alone, rural and low-income areas may receive less investment and attention. Regulators and operators should track quality of service equitably.

1. Build a unified data layer. Normalise KPIs, alarms and configuration data across vendors into a common schema. This is the single biggest enabler.
2. Start with open-loop recommendations. Let models suggest changes that engineers approve. Move to closed-loop automation only for well-understood actions with clear rollback.
3. Prioritise energy and availability. In markets with diesel-dependent sites, forecasting generator fuel needs, battery health and grid availability delivers fast returns.
4. Use digital twins for training. Reinforcement learning and large configuration changes should be tested in simulation before reaching live traffic.
5. Invest in people. Network engineers who understand ML, and ML engineers who understand radio, are rare. Cross-training is a better investment than another vendor dashboard.
6. Treat security as part of the model lifecycle. Sign models, control which apps can act on which parameters and audit every automated change.

From my own work on backend systems at scale, the lesson that transfers most directly is observability: if you cannot trace why an automated system made a decision, you cannot operate it safely. That is doubly true when the system controls infrastructure millions of people depend on.

The operators that pull ahead over the next few years will likely be those that treat AI as an operating model rather than a set of pilots: shared data, clear automation policies and engineers empowered to question what the models recommend.

• Network complexity and cost pressure make AI-driven automation a necessity rather than a luxury for operators.
• Core techniques include KPI anomaly detection, alarm correlation, graph neural networks over cell topology and RL for radio optimisation.
• Energy saving and site availability are high-impact, low-controversy starting points, especially where towers run on diesel.
• Open RAN architectures turn networks into platforms for AI apps but raise new security concerns.
• Begin with human-approved recommendations and strong observability before closing the automation loop.

1. The Mobile Economy — GSMA
2. O-RAN Alliance — O-RAN Alliance
3. 3GPP — 3rd Generation Partnership Project
4. AI for Good — International Telecommunication Union
5. Semi-Supervised Classification with Graph Convolutional Networks — arXiv (Kipf and Welling)

Cover image: original illustration created for this article. Cited works remain the property of their respective authors and publishers.