Korea AI Basic Act AI Compliance Checklist: Ten Actions Before the Grace Period Closes

The grace period on administrative fines under South Korea's AI Basic Act closes on 22 January 2027, one year after the Act and its Enforcement Decree took effect. Between now and then, the Ministry of Science and ICT issues corrective orders rather than penalties, which makes this the cheapest period in which to discover a gap.

This checklist assumes an enterprise operating an AI-enabled product with Korean users, calling hosted models over HTTP. It runs in the order an inspection tends to move: scope and standing first, then notification and labelling, then supervision and evidence. Each item pairs the action with the artifact it produces, because MSIT reads artifacts and a corrective order is written against the missing one.

1. Run the high-impact determination and date it

Assess whether each AI system materially affects human life, safety, or fundamental rights, working through the named sectors: healthcare, energy, transportation, hiring, and biometric analysis. Evidence is a written determination naming the system, the sector, the reasoning, and the date it was made. This document is the one an inspection reads first, and it sets the scope of every obligation below.

2. Test the domestic representative thresholds against actual numbers

Check annual revenue above KRW 1 trillion, AI service revenue above KRW 10 billion, and average daily Korean users above one million across the preceding three months. Crossing any one of them triggers the requirement for a designated local agent. Evidence is a threshold assessment refreshed on a schedule, since the user-count threshold moves with growth and a company can cross it without a board decision.

3. Reconcile your AI feature inventory against production traffic

List every product surface that calls a model, then check it against the endpoints actually receiving requests. Evidence is a destination log naming every model endpoint called during the period. This item exists because notification duties attach to features, and features ship behind flags faster than notices get updated. It is also how shadow AI discovery surfaces the calls nobody registered.

4. Serve prior notification and keep proof it was served

Notify users in advance that the product or service uses AI, for every feature on the reconciled inventory. Evidence is the notice artifact paired to the inventory from item 3. Failing to notify is one of the three violations that carries a KRW 30 million administrative fine directly, and it usually fails through drift rather than intent.

5. Evaluate the generative AI labelling rule in the response path

Apply labels indicating content was produced by generative AI, with heightened notice where sound, image, or video is difficult to distinguish from human-created work. Evidence is a per-response record showing the rule was evaluated and what it produced. A label applied in a rendering template can be described to an inspector; a rule evaluated in the response path can be shown firing against a specific output on a specific date.

6. Bind a named identity to every model request

Route AI traffic so a named user or agent, rather than a shared API key, is attached to each prompt. Evidence is a per-request record showing which principal sent which prompt to which model. The meaningful-explanation duty and the supervision duty both resolve to questions about a specific person's specific interaction, and a single service credential collapses every caller into one. The architecture is set out in identity-aware AI gateway design.

7. Classify prompt content before it leaves the network

Evaluate the classification of the context window at the boundary, since a prompt assembles fragments from several stores into a payload no catalogue ever classified as a unit. Evidence is a per-request classification decision. For high-impact systems in healthcare or hiring, this is the record that shows which category of personal data reached which endpoint, which is the question a Korean inspection and a PIPA review both arrive at.

8. Make human supervision produce a trail

Implement intervention and supervision so ambiguous or high-risk requests are held, escalated, or blocked, with a fail-closed default. Evidence is a decision record showing which requests were held, under which rule, and who resolved them. Documenting that a mechanism exists satisfies a reading of the statute; showing it operating on 14 requests last quarter satisfies an inspector.

9. Keep the evidence write path outside the application

Ensure the compliance record is produced by a system the application making the AI call has no custody over, so it survives selective logging, later modification, and crash-loss between the model response and the log commit. The reasoning sits in You Own the AI Liability, Not the Vendor, and it matters more here than usual because the domestic representative answering MSIT is a different person from the engineer who owns the logging code.

10. Rehearse the inspection response before you need it

Confirm you can answer, for an arbitrary past window, which principals sent which classifications to which endpoints under which policy revision. Evidence is a reconstructable timeline query with a known runtime. My honest take: this is the item that decides the outcome, because obstruction of an inspection carries a fine while a genuine control gap during the grace period carries a corrective order. Being slow to answer is treated worse than being imperfect, and that ranking surprises most teams. AI audit log retention requirements covers how far back the window has to reach.

What the ten items resolve to

Items 3 through 9 all fire at one place: the HTTPS request carrying a context window to a model endpoint. The controls mapping lays out which obligation lands on which enforcement point, and the audit evidence view covers what each artifact looks like when MSIT asks for it.

DeepInspect

Seven of these ten items produce evidence at the boundary between your users or agents and the LLM APIs they call, and DeepInspect sits at that boundary. It binds identity to each request, classifies the prompt before transmission, enforces per-role and per-route policy with a fail-closed default, records the destination endpoint and version, inspects the response, and writes a signed per-decision record before the response returns.

For an operator with Korean users, items 6 and 9 are the two that set how every other answer reads, because identity binding makes a record meaningful and write-path independence makes it evidence. A domestic representative facing an MSIT inquiry then works from a query rather than a reconstruction. Book a technical deep dive at deepinspect.ai.