KT ‘Meisters’ and Microsoft engineers build AI agents to spot network issues before customers do

Microsoft's forward-deployed engineers worked from KT's offices alongside the development team, writing code and jointly building the processes for collecting, applying, and rolling back feedback.

"AI development is constantly evolving, which makes it difficult to approach through traditional outsourcing," says Hosong Lee, head of the KT AIO-Net development team. "The Microsoft team understood what we were trying to achieve and worked alongside us through implementation. They also brought deep experience in how best to use Azure."

Microsoft Frontier Company builds on the Forward Deployed Engineering (FDE) model-embedding engineering experts directly into the customer's environment to co-design, deploy, and continuously improve AI systems at scale. Outcome-driven by design, every system is built to deliver measurable business results.

Learning from the Meister

For AIO-Net to analyze a network incident like an experienced operator, it was trained on the company's standard operating procedures and by the Meisters themselves. Experts like Oh, for example, have intrinsic knowledge gained from over two decades in the field.

"Network operations involve a wide range of manuals and standards, but simply having access to more data does not necessarily make it easy to reach an accurate decision," Jeong explains. "We believed that allowing AI to learn from the expertise of KT's existing Meisters would help improve confidence in the system."

The system, powered by Microsoft Azure, is not just a repository for KT's institutional knowledge. It's also designed to create a continuous learning loop, a process known as hill-climbing in AI, improving processes for both human and machine.

Every minute, its AI agents-built using multiple large language models in Microsoft Foundry -analyze anomalies in large volumes of real-time data generated by KT's network equipment across the country. The agents' conclusions and processes are all documented, enabling the network experts to evaluate them and hone their accuracy. The experts' feedback informs how the system handles future cases, and new types of network incidents.

"Building a perfect AI system from the outset is extremely difficult," says development team head Lee. "The system therefore needs to be designed to improve continuously."

Meanwhile, operators who use AIO-Net are contributing to a feedback loop, just by using it.

"Previously, we had to request feedback, and typically one or two regional representatives would respond," says Seonghyun Lee, KT's development team manager. "The current system captures feedback naturally as people use it, even if we do not explicitly ask them to provide it. They simply need to continue using the system."

The results so far? According to KT, it now takes 2 to 3 minutes for the system to detect a network issue, from over 10 minutes previously. On average, agents can investigate an incident and produce a report in five minutes, about half the time they used to take manually.