Artificial Intelligence in Telecoms Beyond the Hype

Artificial intelligence (AI) has become a cornerstone of every IT-driven industry. AI in telecommunications started way before 2022 (when ChatGPT went public) and extends far beyond large language models (LLMs). To look beyond the hype, we need to focus on the most-used or popular forms of AI: Machine learning, large language models, and agentic AI.

A Look at the Most Popular Forms of AI

Predictive AI is the most mature family of AI in telecoms. It uses statistical and machine-learning models to detect patterns and forecast events based on historical data. Its power resides in the fact that there is no imperative to discover an explicit rule; the algorithm adapts to the ingested information. It then becomes a stochastic oracle, detecting weak signals, such as emerging faults before outages or early fragile behaviors of future churning customers. Predictive maintenance based on AI helps to identify abnormal pattern events and report the probability of failures before they become reality.

Customer value management (CVM) is one of the most profitable AI applications maximizing customer lifetime value. CVM has evolved in sophistication, using more machine learning to determine the ‘Next Best Action’ or ‘Next Best Offer’ to the customer.

Generative AI (GenAI), such as LLMs, can create new content—like text, code, or summaries—through the compression of vast libraries of information. The prerequisite is to be able to create (through training) a model thanks to computing power (typically NVIDIA processors). With GenAI, computers give a full illusion of speaking natural language. On the operational side, it can interact with technical situations in intelligible terms or speak to customers in a personal manner (as opposed to rigid legacy chatbots).

Analyzing Emerging Forms of AI

The emerging technology is agentic AI, a system of autonomous agents that perceive, decide, and act toward defined goals. This synthetic project manager aims to be an orchestration layer above processes. Agentic AI often uses LLMs as a reasoning engine and is enhanced with other modules (memory/action/feedback loop, etc.). This opens many opportunities for high-level automation.

Behind the hype and benefits, several barriers need to be overcome to unleash the full AI potential.

AI is a transformative technology that must be deployed into a legacy organization. CVM projects, for instance, are primarily about organizational change, which is broader than an IT project, requiring a transversal approach and executive support. In real life, the full potential may also be constrained by the organization (siloed channels, outsourcing, etc.).

Control, Ethics

The fair measurement of AI benefits is also a challenge. Orange has developed its own methodology (Data Value Measurement) to compare the value of different projects. But in the end, there is a mixture of avoided loss (customer retention), avoided cost (operation optimization), and new revenues (upsell), which differ in budget, nature, and in timeline.

The quality and labeling of data still need to be improved. Models trained on old network configurations may lose relevance—a phenomenon known as model drift. Hallucinations of LLMs are also a concern; GenAI output can never be 100% trusted. These examples require regularly retraining models or manual checks, reducing automation benefits.

With neural networks and latent spaces of LLMs, AI systems have become black boxes. Their strength, which lies in the fact that no one must build complex laws to make them work, becomes a weakness when the time comes to explain an AI-led action. This can be a problem of control (Do we let AI make decisions?), quality (Was it the best decision?), and ethics (Do we control bias?).

In conclusion, beyond the hype surrounding LLMs, AI is already reshaping telco operations and is now a foundational pillar for automated interaction and the emergence of autonomous networks.