Preparing your data for AI without boiling the ocean first
"Get your data in order first" is advice that has stalled more AI projects than it has saved. Some data problems will sink your feature and most will not, and telling them apart is the useful skill.
By Quality AboveAll · · 8 min read
- Scope data work to the specific corpus your first feature touches, not to the whole organisation.
- Contradictory and outdated documents are the most damaging problem, because retrieval surfaces them with new authority.
- Permissions metadata is non-negotiable if the feature will serve more than one class of user.
Resist the platform-first instinct
A full data warehouse programme before any AI feature ships delays value by quarters and frequently builds for requirements nobody has validated. The alternative is narrower: identify the documents or tables your first feature actually needs and get those into shape.
This has a second benefit. Building the feature teaches you what the data problems really are, which is far more reliable than predicting them, and it produces a concrete requirement for any broader data work that follows.
The problems that genuinely break things
Contradiction is the worst. If three documents state different refund windows because two were never retired, retrieval will surface whichever scores highest and assert it confidently. The AI did not create that inconsistency, it made it visible and authoritative.
Missing permissions metadata is the second. If you cannot tell which documents a given user may see, you cannot build a multi-tenant or role-aware feature safely, and retrofitting access control after launch is far harder than including it in ingestion.
AI does not make your documentation problems worse. It makes them fluent, confident, and quotable.
The problems you can live with
Inconsistent formatting, mixed file types, imperfect metadata and varied writing styles are largely tolerable. Modern extraction and retrieval handle heterogeneity far better than template-era tooling, and the effort to normalise everything rarely pays for itself.
Duplication is usually survivable too, as long as duplicates agree. Near-identical documents waste some retrieval slots and can be handled at the ranking layer rather than by an upstream deduplication project.
Structure content so it can be retrieved
Documents written with clear headings and self-contained sections chunk well, because a section boundary is a meaning boundary. Long undifferentiated prose forces arbitrary splits that cut through the middle of explanations.
Where you control the content, small editorial changes pay off disproportionately: expand acronyms on first use in each section, avoid references like "as described above" that break when a chunk travels alone, and keep the subject explicit. The chunking mechanics are covered in embeddings explained.
Build ingestion as a pipeline, not a migration
Content changes, so a one-off import guarantees the index drifts out of date. Treat ingestion as a running pipeline with incremental updates, deletion handling, and visibility into what was indexed when.
Deletion handling is the detail teams forget. A document removed from the source that remains in the index means your AI feature is quoting a policy you retracted, which is exactly the failure mode you adopted retrieval to avoid. Governance framing is in AI model governance.
Frequently asked questions
Do we need a data warehouse before doing AI?
No. Scope data work to the corpus your first feature uses. Broader data platform work should be justified by requirements the first features reveal.
How clean does the data need to be?
Clean enough that it does not contradict itself and carries the permission metadata your feature needs. Formatting inconsistency is largely tolerable.
What about data we are not allowed to send to a model provider?
Identify it explicitly before choosing an architecture, then either exclude it, redact it, or use a deployment where data does not leave your control.
Told your data is not ready for AI? A free 30-minute consultation will separate the problems that actually block a first feature from the ones that can wait.
