Sunday, November 15, 2026 Independent reporting on artificial intelligence Newsletter
AI News Beats
AI for Business

KT uses read-only AI agents to help engineers diagnose network problems

KT and Microsoft describe a network AI system that gathers evidence and recommends next steps while leaving operational decisions and actions to people.

Computer components and wires are illuminated inside an electronic system
Photo: Florian Krumm / Unsplash

South Korean telecom company KT is using AI agents to support network diagnosis without giving them authority to change the network. A November 11 Microsoft account describes AIO-Net, a platform developed with KT experts and Microsoft engineers that helps assemble evidence when equipment behaves unexpectedly.

According to the company account, the system runs on Azure and uses models available through Microsoft Foundry. KT’s specialist engineers supply operational knowledge and review the agents’ findings. The agents analyze problems and suggest next steps, while people retain decisions and actions.

Read-only access is the important boundary

Microsoft says a permitted list of read-only tools is enforced in application code, rather than relying only on instructions given to the model. That is the most consequential design detail in the account: an incorrect recommendation is not automatically an instruction executed against live infrastructure.

The report describes deployment across 70 critical equipment units, with expansion planned. KT’s reported improvements in diagnosis speed are company claims, not an independently reproduced performance benchmark. This deployment account does not establish the same results for other networks.

For infrastructure managers considering similar systems, the architectural lesson is narrower and more useful than a promise of autonomous operations. Separate the right to inspect from the right to modify. An assistant that can collect useful evidence may deliver value before it is trusted with potentially disruptive changes.

A longer engineering relationship

The project fits into a relationship announced in September 2024, when KT and Microsoft outlined a five-year, multibillion-dollar partnership. That agreement covered AI, cloud technology, research, and services for Korean customers, alongside KT’s own transformation.

The earlier announcement included customized AI solutions, sovereign cloud work, and plans to develop agents using Microsoft tools. It provides context for the collaboration, but it should not be read as proof that every promised initiative is now deployed or that the network system is a generally available product.

Operational teams evaluating this approach should ask for examples from their own incident history. A useful test would compare the assistant’s proposed explanation with the eventual confirmed cause, while tracking both missed problems and plausible but incorrect suggestions.

What to measure before widening access

Speed is only one part of that evaluation. A faster initial diagnosis can still create extra work if the evidence is incomplete or the recommendation points engineers in the wrong direction. Teams should measure the quality of the handoff, not just the time until an answer appears.

Another practical question is whether a reviewer can reconstruct why the system reached a conclusion. Logs, source observations, and an explicit distinction between observed facts and inferred causes can make an assistant easier to challenge. Human oversight is more meaningful when the person has enough information to disagree.

KT’s deployment is worth following as an example of bounded assistance in a consequential environment. The next evidence to watch is how performance holds up across a broader range of equipment and incidents, and whether the read-only boundary remains explicit as the system expands.

Stay on top of AI

Get the stories that matter in artificial intelligence, straight to your inbox.

Free. One email with the stories that matter. Unsubscribe any time. See our privacy policy.