The Enterprise AI Data Readiness Checklist

4. Governance & permissions

Ready when: you know who's allowed to see what, and your AI system can enforce it. An AI that retrieves across permission boundaries is a data-leak incident. Quick test: can you attach access-control metadata to each document and filter retrieval by the requesting user's permissions?

5. Privacy & PII

Ready when: you know where personal/sensitive data lives and have a plan to mask, redact, or exclude it before it reaches the model or logs. Quick test: do you know which fields contain PII, and is there a redaction step in the pipeline?

6. Freshness

Ready when: the data updates on a cadence that matches the use case, and your pipeline re-ingests changes (you're not answering from a six-month-old snapshot). Quick test: how does a document edited today reach the AI system — and how long does it take?

7. Coverage & volume

Ready when: you have enough of the right data to answer the questions users will actually ask — not too little (gaps) and not a firehose of irrelevant noise. Quick test: take 10 real user questions — is the answer present in your corpus?

8. Labeling (only if you're training/fine-tuning)

Ready when: you have labeled examples of sufficient quality and quantity. (For most enterprise use cases, RAG or prompting beats fine-tuning — see fine-tuning vs RAG vs prompting before you invest here.) Quick test: do you have hundreds+ of clean, representative labeled examples?

How to use this

1. Score each dimension red / amber / green for your specific use case.
2. Fix the reds before building — especially access, quality, and governance. They're the ones that quietly sink projects.
3. Don't wait for perfect. Green on the dimensions your first use case needs is enough to start; you don't need the whole enterprise clean.

FAQ

Do I need all 8 green before starting? No — only the ones your first use case touches. A narrow, well-scoped use case on a clean data source beats a broad one on messy data.

What's the most commonly underestimated dimension? Governance/permissions and freshness. Teams nail access and quality, then discover the AI leaks restricted content or answers from stale data.

We don't have labeled data — can we still do AI? Yes. Most enterprise use cases use RAG or prompting, which need good source data, not labels. Labeling only matters if you're fine-tuning.

How does this connect to RAG quality? Directly. Retrieval can only surface what's accessible, clean, and permitted. Data readiness is RAG readiness.

AI readiness is mostly data readiness. Audit access, quality, structure, governance, privacy, freshness, coverage, and (if training) labeling — fix the reds your use case depends on, and you've removed the reason most AI projects stall.

Get your score: Enterprise AI Readiness Assessment. More: RAG systems explained and RAG chunking strategies.

Assessing whether your data is ready for AI and want an expert review? Let's talk.