by david erlichConsulting Director sofrecom
Artificial intelligence (AI) is the foundation of every IT-driven industry. AI in telecommunications started long before 2022 (when ChatGPT was launched) and has expanded far beyond large-scale language models (LLMs). To look beyond the hype, you should focus on the most commonly used or popular forms of AI, such as machine learning, large-scale language models, and agent AI.
A look at the most common forms of AI
Predictive AI is the most mature AI family in communications. Use statistical and machine learning models to detect patterns and predict events based on historical data. Its power lies in the fact that it is not mandatory to discover explicit rules. Algorithms adapt to the information ingested. It then becomes a probabilistic oracle, detecting weak signals such as new failures prior to an outage or early weak behavior of future churning customers. AI-based predictive maintenance can help identify unusual patterns of events and report the probability of failure before it becomes a reality.
Customer value management (CVM) is one of the most profitable AI applications for maximizing customer lifetime value. CVM is highly evolved and uses more machine learning to determine the “next best action” or “next best offer” for the customer.
Generative AI (GenAI), such as LLM, can create new content such as text, code, and summaries by compressing vast libraries of information. The prerequisite is that the model can be created (through training) thanks to the computing power (usually an NVIDIA processor). With GenAI, computers give the complete illusion that they are speaking natural language. Operationally, they can respond to technical situations in easy-to-understand language (as opposed to strict traditional chatbots) or speak to customers in a personal way.
Analyzing new forms of AI
The new technology is agent AI. It is a system of autonomous agents that perceive, decide, and act toward defined goals. This integrated project manager aims to be the orchestration layer over the process. Agentic AI often uses LLM as its inference engine, augmented with other modules (e.g. memory/action/feedback loops). This opens up many opportunities for advanced automation.
To realize the full potential of AI, we need to overcome some of the barriers behind the hype and benefits.
AI is a transformative technology that needs to be introduced into traditional organizations. For example, CVM projects are primarily about organizational change and are broader in scope than IT projects, requiring a cross-cutting approach and executive support. The reality is that your organization may be limiting your ability to reach your full potential (siloed channels, outsourcing, etc.).
management, ethics
Fairly measuring the benefits of AI is also a challenge. Orange has developed a proprietary methodology (Data Value Measurement) to compare the value of different projects. But the end result is a mix of loss avoidance (customer retention), cost avoidance (operations optimization), and new revenue (upselling) with different budgets, nature, and schedules.
Data quality and labeling still needs to be improved. Models trained with old network configurations can lose relevance. This is a phenomenon known as model drift. LLM hallucinations are also a concern. You can never trust GenAI's output with 100% confidence. In these examples, the model must be periodically retrained or manually checked, reducing the benefits of automation.
With the latent space of neural networks and LLM, AI systems have become black boxes. Their strength in not having to construct complex laws to make AI work becomes a weakness when it comes to explaining AI-driven behavior. This can be a question of control (will you let the AI make the decision?), quality (was it the best decision?), and ethics (will you control for bias?).
In conclusion, beyond the hype surrounding LLM, AI is already reshaping the telecommunications business and is now a fundamental pillar for the emergence of automated interactions and autonomous networks.
