What is it about?
Most AI services treat you like a stranger every time you talk to them. You have to explain your situation, your preferences, and what you need again and again — and even then, the answers are often based on rough averages, such as what people your age typically want. This work asks a simple question: what if an AI could understand a customer without making them explain themselves each time? Telecom companies already hold a rich picture of how customers live and use their services. We built a framework that brings these different signals together into a single, structured "customer persona," and uses it to help a large language model give more relevant, personalized answers. Crucially, we designed the system for the real world of a telecom operator: customer data never leaves the company's own servers, and the system stays stable while handling very large volumes of data every day. In our evaluation, personalization based on this combined persona produced noticeably more relevant and satisfying responses than using no persona, or a single type of data alone.
Featured Image
Photo by Google DeepMind on Unsplash
Why is it important?
Personalized AI is often studied using tidy public benchmarks or profiles that users fill in by hand. Real companies rarely have that luxury. Their data is messy, scattered across different sources, sensitive, and bound by strict privacy and security rules that prevent it from being sent to outside AI services. This work is one of the few studies to tackle personalization under exactly those real-world constraints. It shows that combining several everyday signals into a single customer persona — rather than relying on one source — leads to clearly better answers, and that this can be done entirely within a company's own infrastructure, at the scale a telecom operator actually runs. For customers, it points toward services that feel genuinely understood and consistent across different touch points, without asking people to repeat themselves. For the industry, it offers a practical blueprint for deploying trustworthy, privacy-respecting personalized AI in production.
Perspectives
What made this project meaningful to me was that it started from a very human frustration rather than a technical one — the feeling of having to explain yourself over and over to a service that should already know you. Turning that everyday annoyance into a system that runs reliably inside a telecom operator's real constraints was harder, and more rewarding, than any single modeling result. The most surprising lesson was that no single piece of information was enough on its own; it was only when we brought different signals together that the AI truly began to understand the customer. I hope this work encourages more researchers to build personalization that respects privacy and works in the messy conditions of the real world, not just on clean benchmarks.
Jinmo Kang
LG Uplus Corp.
Read the Original
This page is a summary of: PersonaPlugin: A Multi-Source Persona Framework for LLM Personalization in Telecommunications, July 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3805712.3808416.
You can read the full text:
Contributors
The following have contributed to this page
