About this site 한국어

Unofficial explanatory translation of the Korean AI-Ready Data (AIRD) draft standard. The Korean text prevails.

About the standard

At a glance

Type · OverviewReading time · about 3 minAll readers

In shortAgents cannot confirm facts that are not recorded with the data. The AI-Ready Data standard specifies the methods for four activities on the facts needed for judgment: record, measure, judge and use.

AI-Ready Data Standard · At a glance. Don't guess. Judge. A standard that records facts in the data so agents can judge on evidence. Illustration: data.csv (original preserved) and a manifest listing reference date 2026-08-31, units · codes recorded, terms of use recorded, quality tier measured, marked Quality-Ready; an agent judges on evidence. TTA standard “AI-Ready Data” Parts 1–3 · Expected to be enacted in December 2026.
Don't guess. Judge.The AI-Ready Data standard records facts in the data so agents can judge on evidence instead of guessing.Download this slide · PNG (3200×1800)
01 Background. Data users are shifting from people to agents. Chatbot: data → AI answers → a person judges · acts; if it is wrong, a person corrects it at the end. Agent: data → AI judges → AI acts; gaps in the record are acted on as they are. What an agent can judge = the context recorded in the data.
Data users are shifting from people to agentsWhen agents act on data directly, no person is left to catch gaps, so an agent can only judge within the context recorded in the data.Download this slide · PNG (3200×1800)
02 Problem. You see the catalog; the facts needed to use it are below the surface. Above the waterline, catalog information: title · agency · format · update cycle → findable. Below: five kinds of context and more, that usually exist only in the data steward's head — Meaning → invented rules; Source · timing → wrong basis; Quality → inaccurate values; Linkage → wrong joins; Terms of use → acting out of bounds; + α: task-specific context. → Risks an agent creates when nothing is recorded.
You see the catalog; the facts needed to use it are below the surfaceIf meaning, source and timing, quality, linkage, terms of use and task-specific context are not recorded, agents create risks.Download this slide · PNG (3200×1800)
03 Need. People ask; agents guess. A person asks “What's the reference date?” and the steward fills it in. An agent does not ask; it assumes “Probably this year”. If each organization fills the gaps its own way, agents have to guess again for every dataset. That is why a standard is needed: 1 Same format — learn to read it once, read any dataset. 2 Same criteria — “Ready” means the same in every organization. 3 Verifiable evidence — re-checkable by anyone, including machines.
People ask; agents guessA common format, common criteria and verifiable evidence let agents read any dataset without guessing anew each time.Download this slide · PNG (3200×1800)
04 Definition. AI-Ready Data (AIRD), Standard Part 1, 3.1: Data in the Quality-Ready or Purpose-Ready state that meets the common readiness requirements for AI use. Quality-Ready: quality confirmed · Purpose-Ready: prepared for a purpose — a name covering both states, not a separate file or state. Common readiness requirements: meet all three · unmeasured data never qualifies, however good its quality, and the state cannot be self-declared. Measure and judge: quality measured, judged as a tier. Machine-readable record: down to the layer the state requires. Verifiable evidence: judgment basis and provenance history.
Data in the Quality-Ready or Purpose-Ready state that meets the common readiness requirements for AI useAI-Ready Data names two judged states, and data qualifies only when it is measured, recorded in machine-readable form and backed by verifiable evidence.Download this slide · PNG (3200×1800)
05 Role and scope. Four activities, current and extension. Originals are never overwritten; cleaned results become new versions or derived datasets. Current · Parts 1–3, expected to be enacted in December 2026: Record (Part 3) — manifest: description, judgment basis and lineage, machine-readable. Measure (Part 2) — diagnostic report · tier: 7 quality dimensions; quality tier and diagnostic maturity. Judge (Part 2) — readiness state · record: decision gates judge the move to the next state. Use (Part 3) — judgment with evidence: agents read the record and judge; no evidence, no action. 2027 extension · follow-on parts: Part 4 Transformation Pipeline and Governance; Part 5 Discovery, Exchange and Use — wider scope, same activities. Outside the scope: AI models · security and domain ontologies (combined with external standards); trade · contracts · pricing (commercial side of distribution); AI ethics · bias (excluded from quality dimensions). Thresholds → operating guidelines; platforms → implementation guidelines.
Four activities, current and extensionParts 1–3 set the current scope of record, measure, judge and use; the 2027 follow-on parts (Part 4 Transformation Pipeline and Governance, Part 5 Discovery, Exchange and Use) extend the same activities.Download this slide · PNG (3200×1800)
06 How to apply. A state is not declared; it is judged at a gate. Raw: raw data, as received from the source. Gate G1 → Discoverable: findable; discovery layer complete; values unchanged, only the format converted. Gate G2 → Quality-Ready: quality confirmed; diagnostic maturity DM-2 · official tier Tier 1 or higher; cleaned results become new versions · derived datasets. Gate G3 → Purpose-Ready: prepared for a purpose; quantitative requirements of a registered purpose profile; purpose-specific derivatives · purpose tier; ML, RAG, KG — six types incl. DL, statistics and fine-tuning, judged per purpose. AI-Ready Data: a name covering both Quality-Ready and Purpose-Ready. Suspended: a tier is suspended on integrity failure or a missed promised update, and restored after re-verification.
A state is not declared; it is judged at a gateData moves from Raw to Discoverable, Quality-Ready and Purpose-Ready only by passing gates G1 to G3, and a tier can be suspended.Download this slide · PNG (3200×1800)
07 Where it applies. Public sector and enterprises, one procedure. ① Discoverable — findable, available now: public data portal and agency portal catalogs; internal data catalogs and data exchange products. ② Quality-Ready — quality evidence, once tools ship: publishing the quality of open data; inspecting purchased data. ③ Purpose-Ready — fit for purpose, once criteria are set: RAG over legal texts and electricity demand statistics; internal document search and machine learning data. Start not by building a system, but with one task to hand to an agent · step-by-step methods in “Preparing data”: 1 one task · questions, 2 official copy · steward, 3 write the manifest, 4 meaning · quality, 5 purpose · boundaries, 6 evaluate · maintain.
Public sector and enterprises, one procedurePublic bodies and enterprises follow the same steps, starting with one task to hand to an agent rather than a new system.Download this slide · PNG (3200×1800)
08 Expected effects. From agents that guess to agents that judge. Users · agents: from filling gaps by guessing to judging four things, then acting — fitness · reliability · joinability · usability; hold without evidence. Data providers: from answering the same questions again to the manifest answering — limits of the data are stated up front in the record. Organization · governance: from rules people must uphold to records machines check — from review after use to judgment before use. Start now: Stage 1, write the manifest; no diagnostic tools or guidelines needed. Expected to be enacted in December 2026. Expected effects are design goals, not yet measured.
From agents that guess to agents that judgeYou can start now with stage 1, writing the manifest. The expected effects are design goals and have not yet been measured.Download this slide · PNG (3200×1800)
1 / 9← → keys · swipe

Next steps by role

The schedule is in Current status and release plan.