AI in Telecoms: Lessons from 15 Years Across Africa's Networks

Before energy, there was telecoms. I spent years at Globacom — one of Africa's largest mobile operators — and what that environment taught me about data, complexity, and the limits of technology has shaped everything I have done since. Telecom networks generate more data per second than almost any other industry. The challenge was never having enough data. The challenge was turning any of it into decisions that actually changed how the network performed.

That experience is more relevant today than it was when I lived it. AI in telecoms is no longer a future conversation. The largest operators on the continent are deploying machine learning across network operations, customer experience, revenue assurance, and churn prediction. Some of it is working extraordinarily well. A lot of it is not. The difference between the two outcomes is almost never the algorithm.

What Telecoms Gets Right

Network Fault Prediction

The strongest AI use case in telecoms is predicting network faults before they affect customers. Base stations, fibre links, and transmission equipment all generate continuous performance telemetry — signal quality, traffic load, error rates, temperature readings. Models trained on this data can identify degradation patterns that precede failure with enough lead time to schedule preventive maintenance rather than emergency repair. The operational benefit is significant: reduced mean time to repair, lower truck roll costs, and fewer customer-impacting outages.

What makes this work is not the sophistication of the model. It is the quality of the telemetry data and the integration of model outputs into the existing network operations workflow. Operators who have succeeded here did not build a separate AI system. They embedded predictions into the tools their network operations centre teams were already using.

Revenue Assurance and Fraud Detection

Telecoms was one of the earliest industries to deploy AI for revenue assurance — detecting billing errors, SIM-box fraud, and interconnect bypass schemes that cost operators significant revenue every year. This is now a mature application area and most large operators have some form of automated anomaly detection running across their billing and interconnect data.

The operators that have moved beyond fraud detection to broader revenue intelligence — understanding which customer segments are undermonetised, which products are cannibalising each other, which pricing structures are leaking margin — are operating at a fundamentally different level from those still focused only on catching fraud.

What Telecoms Gets Wrong

Churn Prediction Without Retention Action

Almost every large African operator has a churn prediction model. Most of them have limited operational value. The model identifies customers likely to churn. The output goes to a retention team. The retention team contacts the customer — sometimes days later, sometimes not at all — with a generic offer that was not designed based on what the model knows about that customer's behaviour. The prediction was accurate. The intervention was not designed to match it.

Effective churn management requires connecting the prediction to a personalised, timely intervention — the right offer, through the right channel, at the right moment. This is a workflow and system integration problem, not a modelling problem. Most operators have solved the modelling problem and ignored the integration problem.

Customer Experience Analytics That Stop at Insight

Operators invest heavily in understanding customer experience — Net Promoter Score surveys, call centre analytics, app usage data, social media sentiment. The insight is often excellent. The operational response is frequently non-existent. Knowing that customers in a particular area are experiencing poor data speeds is valuable. Knowing it six weeks after it happened and presenting the finding in a PowerPoint to leadership is not.

The Lesson That Transferred

The single most important lesson I carried from telecoms into energy — and into every consulting engagement since — is this: the value of AI is entirely downstream of the decision it is meant to improve. If you cannot point to a specific decision, made by a specific person, at a specific moment, that the AI output is meant to inform — you do not have an AI deployment. You have an analytics exercise. The distinction matters enormously, and most organisations are still on the wrong side of it.