On September 4, 2026, South Korea's Ministry of Science and ICT, led by Deputy Prime Minister Bae Kyung-hoon, announced a free agentic AI service for every citizen. Kakao (with LG), KT and SK Telecom will each run a consortium offering it. Beta testing starts in October and full launch is due by the end of 2026. Users will reach the agent through KakaoTalk, phone calls and text messages. It will handle reservations, payments, authentication and public-service applications on their behalf, according to The Korea Times. That report contains no information on privacy or oversight safeguards. The design is ambitious, and the ministry should publish its safeguards before the beta opens.
The strongest case for the model
The case for building on private channels is strong. KakaoTalk, KT and SK Telecom already sit where citizens spend their time. A state-built app would have to win adoption from scratch, and many Koreans struggle with apps and websites. SK Telecom's consortium is explicitly targeting those users through phone and text. The ministry chose the three operators from six applicants, judging them on usability, user-acquisition strategy and contribution to the AI ecosystem, per JoongAng Daily. Competition among three consortiums also avoids the single-vendor lock-in that a monolithic government platform would create. For a country that wants homegrown models such as Kakao's Kanana and LG's EXAONE to gain scale, this is a pragmatic route.
What is different about an agent
An agent that makes payments, authenticates users and files public-service applications is not a chatbot. It acts for the user, and each action creates a record linking identity, money and government interaction. When that agent sits inside a messenger or a carrier network, one company can see conversations, payment intent and official paperwork together. The government will supply 512 Nvidia B200 GPUs this year and will fund nationwide service costs from 2027, per the same JoongAng report. Public money is therefore underwriting private infrastructure, and that raises the question of who answers for it.
The design questions are specific, and none of the coverage answers them:
- Can the agent act without a fresh confirmation for each payment or application?
- Is conversation data used to train the consortium's commercial models?
- Can a citizen use the service without the operator profiling them for its other products?
- Who is liable when the agent files the wrong form or authorises the wrong payment?
- Can a citizen move to a rival consortium without losing their history?
Korea already has the regulatory pieces
This is not a regulatory vacuum, and the gap is one of disclosure. The Personal Information Protection Commission (PIPC) has a record on AI services. On May 27, 2025 it issued a prior adequacy review resolution approving Naver's personalised AI search agent, as logged by Digital Policy Alert. That shows a workable pre-launch review path for agents that use personal data. PIPC also recommended on March 4, 2026 that generative AI companies improve the transparency of their privacy policies, and published a user guide on May 19, 2026, per Digital Policy Alert. The sensible step is to apply the same pre-launch scrutiny to all three consortiums and publish the results. I could not verify whether PIPC is already reviewing the programme, and I found no public statement either way.
Proportionate safeguards, not a brake
The argument here is not for delay or for a new statute. A rigid licensing regime would slow a promising service and favour the largest incumbents. Four light-touch commitments, published before or alongside the beta, would address most of the risk:
- Purpose limitation by default. Data from the public-service agent is kept apart from advertising and commercial profiling, and is not used for model training without opt-in consent.
- Human confirmation for consequential actions. Payments and official filings need an explicit user confirmation, with a plain-language log the user can read and export.
- Portability and interoperability. Users can switch consortiums and take their history with them. This preserves the competition the three-way structure is meant to create.
- Independent reporting. The ministry or PIPC publishes incident and error data during the beta, so Koreans can judge the service on evidence.
The Electronic Frontier Foundation argues that digital sovereignty should mean user control over data, interoperability and strong encryption, not just national control. That standard fits here. A service can be domestically owned and still leave users with little control over their own information. Messaging and carrier channels can carry public infrastructure only if the rules governing them are as visible as the service.
What to watch
Kakao says it aims for 5 million monthly active users by December and 10 million by the end of 2027, per The Korea Times. Those numbers would make this one of the largest agent deployments in the world. The beta is the point at which design choices still cost little to change. The ministry has its own press channel and PIPC runs an English-language portal, so there is no obstacle to publishing a short safeguards statement. If the beta opens without one, the lesson would be that private messaging platforms were treated as public infrastructure before anyone had defined what the public is owed.