# Regulation and compliance
> **What you will learn**
> Which layer of the law your initiative lands on, what is due when, and what
> gets missed by misreading the "grace period."
> **As of August 2026.** Dates, amounts, and thresholds may be amended. Check the
> current statute before relying on a figure.
> This chapter is an operational overview, not legal advice. If there is any
> chance your use qualifies as high-impact AI, bring legal in.
## It is already in force
In [Phase 2](/guide/ax-phase2-assessment) you checked "governance and security
requirements" as one line item. This chapter unfolds that line item, and the
first fact to establish is this:
> **Korea's AI Framework Act took effect on 22 January 2026.**
This is not something to prepare for — **the obligations already exist.** Korea
is the second jurisdiction after the EU to bring comprehensive AI legislation
into force.
## Does it apply to us?
The most common misconception is **"we don't build AI, so it doesn't apply."**
The law splits businesses into two types.
| Type | Who | Example |
|---|---|---|
| **AI developer** | develops and supplies AI directly | companies that build models |
| **AI user-business** | **supplies a product or service using someone else's AI** | most companies that wire in an API |
**The second category is overwhelmingly larger.** A shop with a chatbot, a SaaS
with AI summaries, a call centre with automated triage — all here. Overseas
businesses are covered too when they affect users in Korea.
## Three layers
Obligations stack. Each layer down is heavier.
```mermaid
graph TD
A["Layer 1 — every AI business
baseline duties"] --> B["Layer 2 — using generative AI
notice and labelling"]
B --> C["Layer 3 — high-impact AI
risk management, explanation, oversight, 5-year retention"]
```
### Layer 3 — is it high-impact AI?
**This is the judgement that matters.** AI used in areas that "may materially
affect, or pose a risk to, human life, physical safety, or fundamental rights"
is high-impact AI. The fields named in the law:
| Field | How it shows up in practice |
|---|---|
| Medical devices and digital medical devices | diagnostic support |
| Healthcare delivery systems | patient triage |
| Energy supply, drinking-water production | plant control |
| Nuclear energy | safety-related decisions |
| Criminal investigation using biometric data | face matching |
| **Hiring, loan screening and similar rights/obligations decisions** | **CV screening, credit assessment** |
| Operation of transport systems | autonomous driving, traffic control |
| Public-service decisions | benefit eligibility |
| **Student assessment** | **grading, placement** |
**The two bold rows catch ordinary companies most often.** If AI does the first
pass on job applications, that is high-impact AI.
A high-impact classification brings five requirements.
```
□ Establish and operate a risk management plan
□ Establish and run a way to explain how the AI reached its decision
□ Establish and operate user protection measures
□ Human management and oversight
□ Produce documentation confirming the above, and retain it for 5 years
```
**The last line is the one most often missed.** It is not enough to have taken
the measures — you need **documentation confirming you took them**, kept for
**five years**. That leads straight into [the audit
trail](/guide/ax-audit-trail).
### Layer 2 — three duties for generative AI
If you use generative AI, these three apply even without a high-impact
classification.
| Duty | What | Where |
|---|---|---|
| **Prior notice** | tell users **before** providing the service that it runs on AI | terms of service, on-screen notice |
| **Output labelling** | mark output as AI-generated | visible (logo, wording) or invisible (watermark, metadata) |
| **Deepfake labelling** | audio, images, or video hard to distinguish from reality must be **clearly recognisable** as such | explicit on screen |
**There are exemptions from labelling.**
```
✓ Where it is obvious from the product/service name or the screen
✓ Where it is used only for the business's own internal work
```
The second matters. **Internal workflow automation is not subject to
labelling.** But the moment it reaches an outside user — a [chat
widget](/guide/cn-chat-widget), an [email
agent](/guide/cn-email-agents) — layers 1 and 2 both apply.
## "We're in the grace period" is the riskiest reading
The government is running a grace period of at least one year. Here is where the
misunderstanding creeps in.
> **The grace period suspends administrative fines, not the legal obligations
> themselves.**
The duties have existed since 22 January 2026. Even during the grace period, a
serious incident — loss of life, infringement of rights — triggers
investigation. And failing to give the required notice carries a fine of **up to
30 million won**.
**In practice the fine is not the big risk.** Deferring on the strength of the
grace period and then having to reconstruct documentation retroactively costs far
more. Five-year retention documents cannot be manufactured after the fact.
## Personal data is a separate axis
Complying with the AI Framework Act does not discharge the Personal Information
Protection Act. The operational standard is the Personal Information Protection
Commission's **Guide to Personal Data Processing for the Development and Use of
Generative AI**, issued in August 2025.
The guide splits the generative-AI lifecycle into four stages with criteria for
each.
| Stage | What to settle here |
|---|---|
| 1. Purpose setting | what the personal data is used for |
| 2. Strategy | which data, on what legal basis |
| 3. Training and development | training-data handling and safeguards |
| 4. Deployment and operation | runtime controls and user rights |
It also **separates the roles of model developer and model user** — the
responsibilities differ, and most companies should read against the **model
user** column.
## If you have users abroad
| Framework | Character | Where it fits |
|---|---|---|
| **EU AI Act** (Reg 2024/1689) | legally binding | mandatory with EU users. Fines up to €35M |
| **NIST AI RMF** | voluntary | your internal risk-management operating model |
| **ISO/IEC 42001** | certifiable | external proof for customers and regulators |
The common 2026 combination is **NIST for internal operations, ISO 42001 for
external proof, and the applicable law in each jurisdiction for legal duties.**
Standing all three up typically takes 8–12 months.
## When, by phase
| Point | What to do |
|---|---|
| Phase 1 — problem definition | first-pass judgement on whether this touches a high-impact field |
| **Phase 2 — assessment** | **confirm high-impact status, settle the personal-data basis, start legal review** |
| Phase 3 — POC | design notice/labelling, place the human oversight points |
| Phase 4 — rollout | produce documentation and start the 5-year retention process |
**It clusters in Phase 2** for the same reason [Phase
2](/guide/ax-phase2-assessment) gave: approvals and reviews take longer than the
technical work.
## Common misconceptions
### "But we don't build models"
**You are covered as an AI user-business.** Supplying a product or service using
someone else's AI brings you in scope, and in practice this is the far larger
group. "We didn't build it" is not a defence.
### "If it's internal only, is there nothing to do?"
**Generative-AI labelling is exempt; the rest remains.** High-impact fields such
as CV screening apply regardless of whether the use is internal. And personal
data law does not distinguish internal from external at all.
---
## Check yourself
**1. Why does the Act apply to a company that builds no models?**
Answer
**Because it is classified as an AI user-business.** The law distinguishes AI
developers from AI user-businesses, and supplying a product or service using
someone else's AI puts you in the latter — by far the larger group.
**2. Why is "we can prepare later, it's the grace period" wrong?**
Answer
**The grace period suspends fines, not obligations.** The duties have existed
since 22 January 2026, and serious incidents still trigger investigation. In
particular, documents subject to five-year retention cannot be created
retroactively.
**3. What does AI first-pass CV screening add?**
Answer
**It qualifies as high-impact AI, bringing five requirements** — a risk
management plan, an explanation method, user protection measures, human
management and oversight, and documentation of all of it retained for five
years.
---
With the legal requirements settled, time to build and measure →
[Phase 3 — design and run a POC](/guide/ax-phase3-poc)