Government ministries are also showing strong interest in AI transformation (AX). But adopting AI to boost staff productivity is not easy because budgets are set in advance each year and procurement procedures are complex. Park Jong-hyun (박종현), an official in the Broadcasting Media and Communications Commission spokesperson’s office policy and publicity team, solved the problem with AI coding. The only cost was a 30,000 won Claude subscription he paid for personally. If it had gone through a typical procurement process, it would have required a service contract costing at least 40 million won and about 6 months.

Park said, "Because agency budgets are allocated based on priorities, it is hard to reflect budgets related to improving employees’ work convenience, and the fact that a program of this level can be made with AI is meaningful in itself."

He joined the Korea Communications Commission in 2011 and is now in his 15th year as a public servant. He currently oversees the commission’s communications evaluation work. The role includes informing the media about key agency tasks and policies and managing performance to meet management evaluation standards.

He first encountered AI tools through a friend who works at a private company. He said, "I heard we live in an era where the gap between people who use AI and those who do not is widening, so I thought I should try applying it to my work."

He targeted news monitoring. It is repetitive and simple, but runs around the clock. To produce reports twice a day, he has to collect articles in the morning and evening, and even on public holidays he must cover content from the day before through the next morning.

After comparing several AI tools, he chose Claude. He said, "Gemini only wrote code in the results window and it was not code you could use right away, but Claude gave me an executable package file when I described the functions I wanted."

Although he studied information and communications engineering, he said AI coding is not solely a developer’s domain. He said, "I learned the C language more than 20 years ago, but I do not know Python at all, which is used in this program," and added, "If you ask AI, it tells you the details, so you just have to follow along." When errors occurred, he pasted the error messages as they were and asked for fixes. He also asked Claude again about unfamiliar concepts such as APIs, servers and domain connections.

The first version ran on a local PC, which limited shared use. That was because each team member had to install Python and packages individually. He obtained a spare PC within the agency, operated it as a server running 24 hours a day, and uploaded the program as a web app accessed via a URL so it can be used anywhere on PC or mobile.

It handles public information, so the security threat is not large, but he built safety measures. External access goes through the network security service Cloudflare. He said, "Because it is accessed through one step in between, the server IP cannot be known from outside and the ports are not open, so it is relatively safe."

The program’s structure is simple. It links to the Naver News API. When users set keywords and a time range and press the collection button, about 900 items are uploaded per run. If the staff member selects relevant articles, the AI summarizes them and automatically outputs them in a report format. The staff member only needs to paste it into a Hangul file.

A task that took 30 minutes to 1 hour now finishes in less than 5 minutes. It has made it possible to use the lunch break fully between the morning and afternoon reports. Park said, "In the past, the person in charge of news clipping usually skipped lunch or ate at their desk, but now we can eat together."

The perceived effect was bigger on weekends. Because it can also be accessed on mobile, there is no longer a need to match outing plans to the monitoring schedule. He said, "One day, a colleague next to me said that thanks to this, they are no longer afraid of weekends," and added, "I also gained the freedom to go out on weekends."

He is now called the commission’s "AI officer." Maintenance has naturally become his responsibility. Recently, at the team’s request, he added a video search feature. Searching is tricky because videos, unlike articles, have less exposed information such as titles. He solved it by including script content in the search scope. He said, "The obstacle was IP blocking. YouTube recognised periodic crawling as an external attack and blocked access," and added, "When I asked Claude, it told me a workaround, so I solved it."

His next goal is a system to respond to false reports. If a wrong article URL is registered, it automatically stores the original and then periodically checks the article to notify whether it has been corrected. The structure also automatically generates a report comparing the before-and-after content.

The program can be used anywhere in government ministries. He said, "If you only change the API key and keywords, any spokesperson’s office could use it," and added, "You can upload the code to Claude and modify it to fit each agency’s format."

Although it is an AI system, judgement is up to people. He explained that it is still difficult to implement with AI the sophisticated filtering needed to weed out unnecessary articles from results collected by keywords. He said, "If you strengthen filtering, there is a risk important articles will be omitted," and added, "You cannot rely on AI, where responsibility is unclear, when errors occur."

He said it is unlikely that the spread of AI coding within government ministries will significantly weaken private-sector software procurement. He said, "Programs requiring high-level security are not suitable to be replaced by vibe coding," and added, "The ripple effects of AI coding on the software industry are more likely to be felt in areas such as reduced hiring of junior developers than in reduced public procurement."