Korea AI Basic Act requirements for foreign AI suppliers

The supplier should prepare a short product record identifying the underlying model, its own development, each integration, hosting locations and the customer-facing service. It should distinguish the support assistant from the optional decision component and record who can change either component's functions.

Korean law recognises an AI developer that develops and provides AI, and an AI-using business operator that uses developer-supplied AI to provide AI products or services. A supplier developing its own system may be a developer. A supplier building a service around another developer's model may be an AI-using business operator. Its activities can engage both descriptions (AI Basic Act, Article 2(7)).

The Korean fintech's position needs equal care. Using a purchased system, adjusting ordinary settings or changing its interface does not necessarily make the customer an AI-using business operator. Official guidance depicts a bank using supplied loan-screening AI as a user. A customer that develops or controls the AI functionality and provides the resulting AI service requires a different assessment (MSIT/NIA, High-impact AI Determination Guidelines, corrected 29 April 2026, pp. 13–17).

A distributor arranging sales does not acquire all supplier duties through its commercial label. If it integrates AI, supplies its own AI service or accepts a domestic-agent appointment, record those additional functions separately. The contract should reflect the statutory analysis; a clause calling the fintech the “deployer” cannot determine its Korean legal status.

Before providing an AI product or service, the business operator must review whether its AI is high-impact. The statutory test combines a listed use area with a significant effect, or risk of effect, on life, physical safety or fundamental rights. Loan screening is expressly listed. Selling to a fintech does not make every feature high-impact (AI Basic Act, Articles 2(4) and 33(1)).

A support assistant confined to product information, navigation and routing enquiries can remain outside that category. Its generative character still creates transparency obligations. The classification may change if the assistant determines eligibility, recommends credit outcomes or handles a process that materially affects a customer's rights.

The optional decision-support component therefore needs a workflow assessment. Record what it predicts, who receives the output, its weight in the decision and what the human reviewer can actually examine and override. Calling the output “advisory” is insufficient where staff routinely follow it or lack the information needed to challenge it.

Official guidance recognises significant intermediate influence on a final decision. It also distinguishes genuinely ancillary information from AI that materially shapes the outcome. A human signature alone does not establish the former (MSIT/NIA, High-impact AI Determination Guidelines, pp. 79–83).

Where classification remains materially uncertain, seek confirmation from the Ministry of Science and ICT, or MSIT. The request requires product, training-data and process/output information. The ordinary response period is 30 days, with one extension of up to 30 days for stated reasons. A request for reconfirmation can be made within ten days of receiving the response (AI Basic Act, Article 33; Enforcement Decree, Article 25).

This procedure addresses high-impact status. The supplier must still resolve the separate data, financial and business-permission conditions affecting the proposed service.

The Act distinguishes advance notice that a product or service uses high-impact or generative AI, marking of generated outputs, and recognisable disclosure for realistic synthetic audio, images or video. These requirements should be tested against the support interface, API responses, exported transcripts and any media features (AI Basic Act, Article 31(1)–(3)).

The decree allows several methods of advance notice, including the interface, documentation or contractual delivery. Generated-output marking can be perceptible to people or machine-readable. The machine-readable route requires an accompanying text or audio notice at least once. It does not create a universal requirement to apply both marking methods to every output (Enforcement Decree, Article 23(1)–(3)).

For this service, the supplier should provide a Korean-language notice that users can understand, specify the marking mechanism and test whether the fintech's interface preserves it. The parties should agree who supplies the wording, who displays it and who approves changes. API metadata that disappears during integration cannot perform the intended disclosure function.

The decree contains qualified exceptions, including obvious AI use and exclusively internal use by the AI business operator. A business-to-business sales contract does not establish either exception. Record the actual user, delivery channel and purpose before relying on one (Enforcement Decree, Article 23(4)).

The distributor should receive approved sales descriptions and a process for referring technical or legal questions. Marketing should not describe the assistant as a human adviser or promise an approved lending function before that function has passed its release conditions.

If the decision component is high-impact, the responsible operator must implement risk management, technically feasible explanations, user protection, human oversight and supporting documentation. Explanations concern the final result, principal criteria and an outline of training data; they do not automatically require disclosure of the entire model or dataset (AI Basic Act, Article 34(1)).

The decree requires publication of specified summaries and the supervising person's name and contact details, subject to its trade-secret qualification. It also requires the relevant implementation evidence to be retained for five years. The supplier should nominate the document owner and give the customer a usable evidence package before activation (Enforcement Decree, Article 27(1)–(2)).

An upstream developer's work can count towards certain duties where the downstream AI-using business operator has not substantially changed the system's functions. This treatment is limited to the measures actually performed and the first three duties in Article 34(1). It does not transfer every responsibility to the model developer. Secure contractual access to the necessary evidence, update notices and cooperation (Enforcement Decree, Article 27(3)–(4)).

Article 35 requires an effort to assess effects on fundamental rights. It does not impose a universal completed impact-assessment certificate as a launch prerequisite. For a lending-related component, a documented assessment remains a sound release condition because it supplies evidence for risk controls and customer approval. Include effects on vulnerable users under the amended provision (AI Basic Act, Article 35(1)).

A separate safety regime applies only where the decree's cumulative conditions are met: training computation of at least 10²⁶ floating-point operations, use of state-of-the-art AI technology and the specified broad, serious risks. Generative AI or a high training-compute figure alone does not establish that trigger. Obtain model information before excluding it (AI Basic Act, Article 32; Enforcement Decree, Article 24).

Article 36 applies to an AI business operator with neither a Korean address nor a Korean business office that meets a prescribed threshold. The stated absence of an office does not resolve the address or threshold facts.

The decree provides four alternative triggers:

• Previous-year total sales of at least KRW 1 trillion.
• Previous-year AI-service sales of at least KRW 10 billion.
• At least one million domestic users as a daily average over the preceding three months, measured at the previous year-end.
• A previous administrative fine under Article 43(1)(3) for failing to comply with an order under Article 40(3).

Corporate revenue uses the previous business year. Foreign-currency revenue uses the average exchange rate for the previous year, or previous business year for a corporation. The revenue limbs are not expressed as Korean sales alone. The supplier should preserve its calculation, counting method and source records (Enforcement Decree, Article 29).

Where appointment is required, the operator must designate the agent in writing and report the appointment to MSIT. The agent must have a Korean address or business office. Its functions include relevant safety-result submissions, high-impact confirmation requests and support for Article 34 measures, including document accuracy and currency (AI Basic Act, Article 36(1)–(2)).

The distributor can undertake that role only through an appropriate appointment and the ability to perform it. The mandate should cover access to records, response authority, escalation and continuity if the distribution relationship ends. Relevant breaches by the agent are attributed to the appointing operator. Article 36 itself does not require the supplier to incorporate a Korean subsidiary (Article 36(3)).

The main AI Basic Act obligations commenced on 22 January 2026. As at 5 November 2026, the Korean consolidation includes amendments effective on 21 July 2026, and the decree in force is effective from 20 August 2026. An October amendment scheduled for 2027 does not replace that operative text (Act Nos. 20676, 21311 and 22076, supplementary provisions; Presidential Decree No. 36580).

MSIT announced an enforcement transition of at least one year, generally deferring fact-finding investigations and administrative fines while supporting compliance. It retained scope for exceptional investigation in serious cases. This policy does not suspend the statutory duties or the separate privacy and financial regimes (MSIT, AI Basic Act implementation announcement, January 2026; AI Basic Act Helpdesk Casebook, 1 April 2026, p. 37).

The Act provides fines of up to KRW 30 million for missing the advance notice, failing to appoint a required domestic agent and disobeying a suspension or corrective order. Other breaches can lead to investigation and orders under Article 40; an output-marking error is not an independently listed Article 43 fine trigger (AI Basic Act, Articles 40 and 43(1)).

The supplier and fintech must map prompts, customer records, outputs, logs, backups and support access. Korean hosting does not eliminate an overseas transfer if personnel in Vietnam can access personal data. Korea's Personal Information Protection Act (PIPA) expressly includes overseas access, entrusted processing and storage within its transfer rules (Personal Information Protection Act, Article 28-8(1)).

The customer must establish the applicable transfer basis and required notices or disclosures. Separate consent is one route. The contract-necessity route for outsourcing or storage concerns the data subject's contract and its statutory conditions; the supplier's commercial agreement alone is insufficient. Entrusted-processing terms must also address purpose, security, supervision and subprocessing, for which Article 26 requires the entruster's consent (PIPA, Articles 26 and 28-8).

Personal credit-information outsourcing needs a separate Credit Information Act check. Its safeguards and permitted-use rules apply, and subprocessing requires an applicable exception recognised by the Financial Services Commission (FSC); PIPA entruster consent alone does not establish one (Credit Information Use and Protection Act, Articles 17, 32–33).

The foreign recipient must also assess onward transfers to another country under PIPA's further-transfer rule. If the supplier is itself a PIPA personal information controller, check the separate domestic-representative requirements. An AI Basic Act appointment does not automatically satisfy that mandate (PIPA, Articles 28-11 and 31-2; PIPA Enforcement Decree, Article 32-3).

For covered Korean financial companies and electronic financial businesses using cloud services, assess materiality, provider controls, continuity, security and the required internal decision. Event-triggered reporting obligations also apply. Systems processing unique identifiers or personal credit information through cloud services generally must be installed in Korea, subject to specified exceptions. An overseas API needs an architecture and regulatory basis that satisfies that rule (Electronic Financial Supervision Regulation, FSC Notice No. 2026-29, Article 14-2(1)–(4), (7)–(8)).

The label “fintech” does not establish the buyer's regulated category. Obtain its licence or registration position and identify applicable network-separation controls or documented relief. A distributor's assurance is insufficient evidence of a customer-specific exemption (Electronic Financial Supervision Regulation, Article 15).

Decision functions also require separate treatment of automated-decision rights and applicable credit-information rights. A service that itself performs regulated credit assessment needs a permissions analysis beyond its software classification (PIPA, Article 37-2; Credit Information Use and Protection Act, Articles 4 and 36-2).

The FSC's revised financial AI guidelines took effect on 22 June 2026 as self-regulatory guidance. Use the customer's applicable approval process to agree testing, oversight and change controls while preserving the distinction between guidance and binding rules (FSC, revised financial AI guidelines announcement, 18 June 2026).

Vietnam's AI Law applies to development and service activities carried out in Vietnam. Korean sales do not establish an exemption for those activities. Classify both components under the Vietnamese rules as well as the Korean test (Vietnam, Law on Artificial Intelligence No. 134/2025/QH15, Articles 1–2 and 9–10; Decree No. 142/2026/NĐ-CP, Articles 5–14).

The Vietnamese high-risk list contains specified automated banking functions. Its credit-decision entry includes conditions concerning final decisions, direct use without independent credit-officer approval and the loan-value criterion prescribed by the State Bank of Vietnam. A Korean high-impact classification cannot replace this different assessment (Decision No. 33/2026/QĐ-TTg, Article 2 and Annex IV(2)).

Medium- and high-risk providers must prepare the classification dossier and notify the Ministry of Science and Technology before use, subject to applicable transition. High-risk systems also require the applicable conformity assessment. Record any transition relied on: Decision No. 33 requires compliance before 1 March 2027 for systems introduced during the six months after its 15 August 2026 commencement. Its wording is broader than the AI Law's pre-existing-system provision, so establish its application before deferring a statutory pre-use duty (AI Law, Articles 10, 13 and 35; Decree No. 142, Articles 13–14; Decision No. 33, Article 4).

Personal-data processing has a separate permission question. Operating automated processing for a customer, performing creditworthiness scoring or providing other covered AI processing can engage the Ministry of Public Security's processing-service certificate regime. Establish the actual service and the supplier's certificate position before enabling that operation; supplying software and operating customer processing require separate analysis (Decree No. 356/2025/NĐ-CP, Articles 21–25).

The data map must distinguish Korean-origin records, data collected in Vietnam and onward access by model or infrastructure providers. Allocate processing roles and establish the relevant processing and transfer assessments. Vietnamese law does not provide a general exemption merely because the customer or data originated abroad (Personal Data Protection Law No. 91/2025/QH15, Articles 1, 20–22, 30 and 37; Decree No. 356, Articles 17–20).

A procedural amendment takes effect on 15 November 2026. It changes assessment and certificate procedures while retaining the service categories and substantive certificate conditions. Use the procedure applicable to the dossier's receipt date (Decree No. 392/2026/NĐ-CP, Articles 42–49).

The supplier, distributor and fintech need one consistent operating description across their agreements. Define permitted support functions, prohibited decision uses and the conditions for activating the optional component. Include the model version, integration responsibilities and the evidence each party must supply.

The supplier should contract for access to upstream risk, training-data and testing information. The customer agreement should allocate notices, explanations, human review, complaints, incident handling and evidence retention. The data-processing schedule should control reuse for training, support access, subprocessors, deletion and return of records. A separate domestic-agent mandate is needed where the threshold assessment requires one.

Changes to the model, scoring logic, data sources, hosting or subprocessors should trigger an assessment of classification, customer approvals and required notices. Agree advance notice, evaluation, rollback and suspension rights. Licentium's discussion of AI model change control for Malaysian suppliers to Singapore banks develops those contractual controls in a different regulatory setting; the Korean permissions and duties must be assessed separately.

The launch owner should require four recorded decisions:

1. Product and legal owners approve the actor map, component classifications, permitted uses and domestic-agent calculation.
2. Engineering and customer-service owners test notices, output marking, escalation and access restrictions in the delivered interface.
3. Privacy, security and the customer's financial compliance owners approve the data route, procurement evidence and any required permissions or reporting arrangements.
4. The responsible business owner authorises each component against its completed conditions, with the decision function disabled until its additional controls are ready.

The facts most likely to change this sequence are the decision component's actual influence, the customer's regulated status, overseas access, the supplier's revenue and user counts, and who operates the processing service. Preserve those facts with the release decision and reassess them when the service changes.

For decisions spanning several destinations, AI Projects Across Jurisdictions addresses the broader choice of operating and contractual structure.