KT uses AI to diagnose network faults, but keeps repairs in human hands

KT is using AI to investigate network faults while leaving people in charge of repairs. In a November 11 feature published by Microsoft, the Korean telecom says its jointly developed AIO-Net platform detects problems in two to three minutes. Operators have used it since August, with strict limits on what agents can do to equipment.

The task starts with a familiar operational problem: an alarm does not explain itself. KT network engineer Heeyoung Oh previously moved between separate systems to investigate a warning. Her aim was to establish the cause within the first 10 minutes, leaving time to respond before a fault became a major failure.

AIO-Net brings that information into one view and assigns agents to help identify the cause. The work supports a provider serving 29 million mobile subscribers, although the platform’s current deployment covers only 70 units of critical equipment.

Teaching agents how network experts investigate

Powered by Azure, AIO-Net uses multiple large language models through Microsoft Foundry. Every minute, its agents analyze unusual activity in live data generated by KT’s network equipment across the country. Their conclusions and the steps behind them are documented so network specialists can evaluate the work.

The training combines KT’s standard operating procedures with knowledge from experienced network specialists, whom the company calls Meisters. Oh brought more than two decades of field experience. The goal was not simply to make manuals searchable, but to help the system assess incidents using expertise that operators had built through years of work.

Operator feedback then informs how AIO-Net handles future cases, including new types of incidents. KT development manager Seonghyun Lee says the system captures feedback during normal use. Previously, the team had to request it, typically receiving responses from only one or two regional representatives.

Microsoft engineers worked inside KT’s offices alongside its developers rather than handing over a finished system. They wrote code together and built processes for collecting, applying and rolling back feedback. That joint development work included the machinery for revising the system as operators used it.

AI’s authority extends to analysis and the presentation of supporting evidence. Decisions and actions remain with people.

Seonil Jeong, KT’s head of network operations innovation

A read-only boundary, enforced in software

AIO-Net can flag an incident, query and inspect equipment, and suggest a solution. But agents may use only approved read-only commands: instructions that inspect equipment without changing it. They cannot carry out corrective action.

That boundary is enforced within the application, not merely through instructions in a prompt, according to development team head Hosong Lee. He says the software prevents a prohibited action even if the AI attempts it. His reason is concrete: an incorrect command sent to network equipment could itself cause an incident.

Faster investigation, a limited deployment

KT puts the previous detection time at more than 10 minutes. It also says agents now investigate an incident and produce a report in five minutes on average, about half the time the manual process took. These are KT’s performance figures, published by its development partner.

For Oh, the benefit is more time to decide how to respond and assess whether customers are affected. Jeong says reliability takes priority over operational efficiency, and expects time saved to let staff focus more on facility and quality management.

KT’s next planned step is to increase coverage to 280 units of critical equipment, then eventually extend AIO-Net into access networks closer to subscribers. That remains an expansion plan, rather than the scope of today’s deployment.