Do we need a data lake or lakehouse before starting with AI?

No. Consolidated storage helps at scale, but it is neither necessary nor sufficient for a first use case. What matters is whether the specific data a use case requires is accessible, authorized, sufficient and understood. Many organizations with mature platforms still fail on authorization, and many without one deliver successfully.

How good does data quality need to be?

Good enough for the decision it supports, with the gaps quantified. A recommendation that a human reviews tolerates more noise than an automated action with financial consequence. The failure is not imperfect data — it is imperfect data whose limitations are undocumented and therefore invisible to the people relying on the output.

Should unstructured data be included in the readiness assessment?

Yes, and with the same four conditions. Documents, tickets and transcripts frequently carry more of the value than structured tables, but they are also where authorization and retention questions are least likely to have been answered. Assess them explicitly rather than assuming they are unconstrained.

Who owns data readiness — IT or the business?

Accountability sits with the business function that produces and uses the data; IT owns the access route and the platform. When readiness is delegated wholly to IT, definitions and permitted-purpose questions go unanswered because the answers were never IT's to give.

What would change the assessment?

A change in the decision the model supports. Narrowing a use case from an automated action to a human-reviewed recommendation can move it from blocked to deliverable without any change to the data itself, because the readiness bar is set by consequence.

의사결정 인텔리전스

AI 데이터 준비도: 실제 운영이 요구하는 조건

필요한 것은 완벽한 데이터가 아니라 대상 업무에 대해 위치·권한·최신성·품질을 설명할 수 있는 상태입니다. 데이터 정비는 전제 조건이며 도입과 병행 설계하는 것이 현실적입니다.

Decision intelligenceWritten by NirjiX AI AdvisoryPublished January 2026Last reviewed February 20269 min read

Direct answer

핵심 답변

아래 상세 분석은 영문 원문 그대로 제공됩니다. 영문 전체 가이드 보기

Estate maturity is the wrong test

Ask whether the enterprise data estate is ready for AI and the answer is always no, in every organization, permanently. Estates contain decades of accumulated systems, and remediating all of them before starting is an argument for never starting.

The useful question is narrower: for this specific use case, is the data we need reachable, permitted, complete enough and understood? That question is answerable within weeks and frequently produces a surprising result — the blocker is authorization or lineage documentation, not quality.

This reframing matters commercially. Estate-level programs consume years of budget before producing a business outcome, and they are usually the first thing cut when priorities shift. Use-case-scoped readiness work delivers value while leaving durable improvements behind.

The four conditions, and how to test each one

Test them in this order. Accessibility and authorization fail more often than quality, and they fail earlier.

Data readiness conditions, the practical test for each, and the typical remedy when it fails.
Condition	The practical test	Remedy when it fails
Accessibility	Can the delivery team obtain a working sample within two weeks through a supported route?	An access pattern and environment, not a new platform.
Authorization	Is there a named owner who can approve this use, and does the purpose fall within existing consent and contracts?	Data ownership assignment and a documented purpose review.
Sufficiency	Does coverage span the population and period the decision needs, with known gaps quantified?	Backfill, supplementary source, or narrowing the use case scope honestly.
Interpretability	Can someone explain what each field means, how it is produced and where it breaks?	Lineage and definition documentation from the owning function, not from IT alone.
Freshness	Does the data arrive fast enough for the decision cadence the use case requires?	Pipeline change, or a decision to operate on a slower cadence deliberately.

Signals that data will block delivery later

No single named owner for the primary dataset — the most reliable predictor of a stalled use case.
Definitions that differ between the reporting layer and the operational system that generates the record.
Historical data that changed meaning after a system migration, with no documented cut-over.
Quality that is acceptable for aggregate reporting but not for record-level decisions.
Personal or contractual data whose permitted purposes were never assessed for analytical or model use.
Manual correction steps applied downstream that the source system never receives back.

These surface late and expensively. Look for them during assessment rather than during deployment.

How to run a use-case data readiness review

Two to three weeks per use case is normally sufficient.

01
Write the decision first
State the decision the model will inform, its cadence, and the consequence of being wrong. Readiness standards for a pricing recommendation and for a triage suggestion are not the same, and treating them identically over-engineers one and under-protects the other.
02
Trace the data to its source
Follow each required field back to the system that creates it, not the warehouse that stores it. Most interpretation errors originate in the gap between those two.
03
Assess a real sample
Measure coverage, completeness and drift on actual records for the period the use case needs. An inventory of tables is not evidence about quality.
04
Resolve permission explicitly
Confirm the purpose is permitted under the contracts, notices and consent that apply, and record the conclusion. Discovering this after build is the most expensive sequencing mistake available.
05
Record the residual gaps
Publish what remains imperfect and how the design compensates. A use case shipped with documented limitations is defensible; one shipped with undocumented ones is not.

NirjiX view

The NirjiX view

We treat unresolved data ownership as a gate rather than a score. A use case whose primary dataset has no accountable owner is not a low-readiness candidate — it is not yet a candidate, and the work in front of it is governance, not engineering.

We also resist the sequencing that says all data must be fixed before AI begins. In practice the fastest route to a governed estate is a small number of high-value use cases that force ownership, definitions and access patterns to be resolved for the data that actually matters.

Frequently asked executive questions

Do we need a data lake or lakehouse before starting with AI?: No. Consolidated storage helps at scale, but it is neither necessary nor sufficient for a first use case. What matters is whether the specific data a use case requires is accessible, authorized, sufficient and understood. Many organizations with mature platforms still fail on authorization, and many without one deliver successfully.
How good does data quality need to be?: Good enough for the decision it supports, with the gaps quantified. A recommendation that a human reviews tolerates more noise than an automated action with financial consequence. The failure is not imperfect data — it is imperfect data whose limitations are undocumented and therefore invisible to the people relying on the output.
Should unstructured data be included in the readiness assessment?: Yes, and with the same four conditions. Documents, tickets and transcripts frequently carry more of the value than structured tables, but they are also where authorization and retention questions are least likely to have been answered. Assess them explicitly rather than assuming they are unconstrained.
Who owns data readiness — IT or the business?: Accountability sits with the business function that produces and uses the data; IT owns the access route and the platform. When readiness is delegated wholly to IT, definitions and permitted-purpose questions go unanswered because the answers were never IT's to give.
What would change the assessment?: A change in the decision the model supports. Narrowing a use case from an automated action to a human-reviewed recommendation can move it from blocked to deliverable without any change to the data itself, because the readiness bar is set by consequence.

Continue

Related intelligence

이 주제의 핵심 가이드

우리 회사가 AI 도입 준비가 되었는지 어떻게 판단하는가?

이 페이지는 의사결정의 한 부분만 다룹니다. AI 준비도 진단에 대한 NirjiX 종합 가이드에서 전체 그림을 확인하십시오.

AI 준비도 진단 →

함께 보기

연관 가이드

같은 영역의 관련 가이드

파일럿에서 프로덕션으로: AI 프로그램이 멈추는 지점
정체는 예측 가능한 다섯 개 관문에서 발생합니다. 출시되지 않아도 손해를 보는 사업 책임자가 없고, 파일럿 데이터가 수작업으로 조립되어 거버넌스 하에서 운영 주기로 확보되지 않으며, '충분히 좋음'의 기준이 없어 승인이 종결되지 않고, 주변 업무 프로세스가 재설계되지 않았으며, 운영 주체가 정해지지 않은 경우입…

반대 영역의 동일한 의사결정

GCC 확장: 출범 이후의 발전 경로
확장은 인원 증가가 아니라 담당 업무의 복잡도를 높이는 방향으로 진행됩니다. 정형 처리에서 분석, 설계, 의사결정 지원으로 책임 범위를 단계적으로 옮기는 것이 가치의 원천입니다.

Transparency

Sources and methodology

This page reflects NirjiX advisory practice rather than a survey or a vendor benchmark. The structure of the assessment — the dimensions, the maturity language and the sequencing logic — is the same framework used inside the NirjiX AI readiness assessment and the AI plan builder.

Where we describe patterns ("most organizations discover…"), we are describing what we observe across client engagements, not a measured statistic. We deliberately avoid quoting market numbers we cannot verify, because an AI investment case built on borrowed statistics collapses the first time a CFO tests it.

Any figure that ends up in your own plan should come from your own data: your cost base, your cycle times, your error rates, your volumes. The assessment and plan builder are designed to force that discipline.

Assess readiness against a real use case

The AI assessment scores data foundations together with ownership, governance and execution capacity, so the blocker is identified before the build is funded.

Run the AI assessment Build your AI plan

Outputs are preliminary and intended for advisor validation before funding decisions.