AI in South Korea: what is changing and how to approach it

South Korea is doing more than introducing new chatbots. It is trying to connect three goals: strengthening domestic technology companies, expanding access to computing power and making some AI services available to citizens without a subscription. This is an industrial and social policy choice. Describing it as a universal public service is, for now, an ambition rather than an outcome already achieved.

The scale of investment is significant. The government has allocated 10.1 trillion won to AI in 2026, more than three times the previous year’s level. The plan includes research, infrastructure and adoption across industries. In October 2025, Nvidia announced an agreement covering more than 260,000 GPUs for South Korean government and corporate initiatives. That is an announced supply commitment, not evidence that every GPU is already installed or operational. The budget plan ↗ and Nvidia’s announcement ↗ show both the scale of the strategy and one of its dependencies: sovereignty over models is not the same as independence from foreign hardware and suppliers.

The third direction is the development of domestic models. In the government’s sovereign foundation-model project, SK Telecom, LG AI Research and Upstage advanced to the next selection stage in August 2026. The competition is still evolving; the shortlist is not a final assessment of the models’ relative quality. Yonhap’s report ↗ says three teams advanced from a group of four.

The common thread is room to choose: investing in domestic capabilities and models while retaining access to a global supply chain for chips, cloud infrastructure and services. In this context, sovereignty is a matter of degree and managed dependencies, not an all-or-nothing condition.

AI for All is the Ministry of Science and ICT’s programme to offer citizens general-purpose AI services and agents for selected public tasks. The ministry says it aims to provide a free chatbot without usage limits, alongside an agent that can help people find services and handle requests. As of 16 October 2026, however, the programme is still being rolled out: the three selected consortia are demonstrating and testing their services, and public launch is planned for December. That date is an announced plan, not a service already available to the entire population.

The consortia are led by SK Telecom, KT and Kakao. Their proposals reflect different strengths: SK Telecom is working on agents that can turn a request into an action; KT aims to connect several services in a single conversational experience; Kakao can distribute features through a messaging and partner ecosystem already familiar to many users. The ministry ↗ requires at least 50 per cent of each service to use sovereign models developed in Korea, and at least 30 per cent to come from models developed by other Korean companies. The requirement also makes the programme a way to create demand for the domestic industry.

The initiative addresses a real divide, but does not automatically close it. The ministry reports that around two-thirds of citizens have used AI services and 23 million have used generative AI; about one-third of the population has not yet used AI. Free access, skills, accessibility, trust and service quality are separate questions. A service open to everyone may lower a financial barrier, but its impact will also depend on language, ease of use, support and its ability to meet people’s needs. The ministry’s figures and programme aims ↗ are published by MSIT.

The programme also raises a governance question. The ministry says operators may develop revenue models using data collected while providing the service, for example users’ prompts. This does not mean every conversation will be used to train models: the methods, purposes, retention and controls will depend on each service and its terms. It does make it essential to understand what is collected, for how long, who can access it and for what purposes.

South Korea’s AI Basic Act came into force on 22 January 2026. The framework combines industrial promotion with development and trust obligations. Its provisions do not apply identically in every case: the organisation’s role, the type of service and the way AI is used all matter.

The Ministry’s transparency guidelines distinguish, among other things, advance notice when generative or high-impact AI is used and the identification of AI-generated content. Labelling may work differently for content exported or shared outside a service than for content shown within its interface. For outputs that are difficult to distinguish from reality, such as some deepfakes, the guidelines call for recognisable indications. The Act also sets specific requirements for high-impact systems, including human oversight. It would be inaccurate to infer that every business using a generative assistant must follow the same set of obligations.

The ministry announced a grace period of at least one year, during which investigations and penalties concerning the transparency requirements would be deferred to help businesses adjust. That transition does not suspend the entire Act. As the law was amended in 2026 and obligations depend on role and use case, a company offering services in Korea should check the current text and guidance before launching or changing a service. The Act ↗, the amendment effective in July 2026 ↗ and the ministry’s guidelines ↗ are the place to start.

For an international brand, the practical question is not just “are we covered by the law?” It is also: which entity provides the experience, which systems and partners are involved, what data crosses borders and which local requirements apply? The answer may differ from one use case to another and warrants advice from local counsel.

The development closest to fashion and retail is the arrival of agents inside platforms people already use. Naver began the beta of its Shopping AI Agent in February 2026. It summarises information, analyses products and suggests alternatives in response to natural-language requests. In June, the company described an evolution in which the agent can start a conversation and suggest next steps. This is a direction announced by the platform, not proof that purchases are already being made autonomously. Naver describes the beta launch ↗ and the June update ↗.

Kakao has expanded Kakao Tools inside ChatGPT for Kakao, integrating partner services such as Olive Young and Musinsa. Users can move from a conversation to product discovery or partner services within a chat interface. The strategic point is not that every purchase has become “agentic”; it is that platforms, conversations and commercial catalogues are beginning to meet in the same journey. Kakao announced the integration in March 2026 ↗.

For brands, this creates another layer of discoverability. An agent can only work with information it can retrieve and interpret: product title and description, materials, sizes, availability, price, delivery, returns and service details. In fashion and luxury, these are not merely technical attributes; they also explain provenance, care, craftsmanship and differences between products.

“Optimising for agents” does not mean filling a catalogue with keywords. It means making information reliable, current, consistent and available in Korean, then checking how platforms present it. Visibility cannot be guaranteed: it depends on each service’s index, partnerships, rules and users’ requests.

Entering the Korean market does not require replicating the national programme or launching a proprietary chatbot. It is more useful to understand where customers discover, evaluate and buy products, and what information they need to understand the brand’s offer accurately.

**Build a considered local presence.** Naver and Kakao are not just media channels: they are building experiences that connect search, conversation and services. A brand should map relevant entry points, its product presence, local partners and the data each platform can use.

**Prepare the catalogue for conversational discovery.** Natural Korean descriptions, consistent attributes, clear imagery and materials, current availability, and information on care and returns help both people and systems distinguish products correctly. Localisation should respect the brand’s voice rather than rely on literal translation.

**Choose a measurable use case.** A customer-service assistant, a product-discovery journey or an internal tool for preparing appointments may make more sense than an agent that immediately buys on a customer’s behalf. Start by asking whether the experience improves: response relevance, errors, resolved queries, conversion or service quality.

**Govern data and partners.** Before connecting an assistant to a CRM, purchase history or clienteling data, define purpose, permissions, retention, traceability and human intervention. Sensitive or unnecessary information should not enter a workflow simply because it is technically accessible.

**Train teams on real work.** Retail, merchandising, marketing, legal and technology teams need to know what a system can see and do, how to correct an error and when to hand over to a person. Training works best when it is tied to everyday processes, not just a list of approved tools.

A first experiment can remain small and produce useful evidence before a larger investment.

Table 1. A proposed initial work plan; timing and metrics should be adapted to the use case.

Testing should include ordinary and difficult requests: an ambiguous question about size, two similar products, an unavailable item, a question about materials or care, and a returns issue. For a luxury brand, tone matters too: an answer can be formally correct and still fall short of the intended experience.

The effects of these changes do not stop at South Korea’s borders. A customer who becomes used to asking questions in natural language, comparing products in a conversation and receiving immediate information about sizes, materials and availability may carry those expectations in two directions.

In South Korea, they apply to European brands operating in the market. If an agent cannot find complete or up-to-date information in Korean, customers are unlikely to see this as a limitation of the platform. They may instead perceive the brand as less clear, less present or less attentive than its competitors.

In Europe, the same expectation reaches the boutique. A Korean customer entering a store in Milan or Paris may already have searched, compared and selected a product before arriving. Will the experience take those previous steps into account, or will the customer have to start again? Will staff have access to the same information the customer has seen, or will the journey break off at the door?

This is not about recreating an agent’s interface in the store. In luxury, value still lies in the relationship, service and care. The risk is that the gap between what customers find online in their home market and what they encounter in a boutique becomes more visible, and that they judge the experience against expectations formed elsewhere.

South Korea’s strategy involves both opportunities and trade-offs. A public programme may broaden access, but its success depends on adoption, quality, inclusion and long-term funding. Domestic models may build local capabilities, while chips and computing remain tied to global suppliers. Agents may reduce steps in product discovery, while attribution, accuracy and commercial accountability become harder to see.

The idea of “presence in agents” also needs care. Each platform controls data, ranking, integrations and user access according to its own rules. A strategy dependent on a single intermediary may improve discovery in the short term while increasing dependency over time. Proprietary information, direct channels and customer relationships therefore remain valuable.

Finally, public adoption does not automatically mean shared benefits. Free access and wider distribution can lower some barriers; outcomes for different groups, conversation handling, privacy and service sustainability still need to be understood. These questions are part of the programme’s design, not details to address only after launch.

South Korea is building an ecosystem in which support for domestic models, AI services for citizens and commercial platforms reinforce one another. AI for All is still being rolled out, but it signals the government’s direction; Naver and Kakao show how AI can become another layer between consumers and products.

For fashion and luxury, the most practical response is to prepare the ground: accurate, localised product information, a presence in the right channels, carefully assessed partnerships and use cases that improve the service. Agents may become an important interface, but they do not replace trust in a brand or guarantee product discovery. The advantage lies in being understood clearly without giving up control of the experience and data.