From spreadsheets and sensors to AI-ready datasets

Data quality · 12 September 2026 · AI Collective Mind

“Ready” depends on the requirement

A dataset can be useful for one purpose and unsuitable for another. A daily aggregate might answer a market-research question but omit the detail needed for equipment monitoring. Before changing formats or commissioning labels, define the intended use and the criteria by which the buyer will judge a source.

Understand the observation unit

For a table, explain what each row represents. For a sensor stream, document what was measured and when. For an archive, distinguish documents, pages and versions. Quantity without a unit is hard to interpret; a count of files may say little about coverage or diversity.

Keep context with the values

Document units, languages, time zones, collection methods and known gaps. Explain whether numbers were measured, estimated or derived. Instrument changes, shifts in product identifiers or revised research methods can affect comparisons across time. Good preparation makes these changes visible rather than silently smoothing them away.

Review quality against agreed criteria

Potential work includes normalising formats, checking duplicates, assessing missingness, reviewing transcripts or defining annotation rules. Record who performed a review, the method used and its limitations. Do not present hypothetical specialist capabilities as already completed work or claim a quality score that has not been measured.

Keep rights attached to every source

Combining collections does not make their restrictions disappear. Preserve provenance, permission evidence and allowed uses alongside technical documentation. When personal information may exist, assess identification risk and the relevant requirements before processing. A replacement identifier is not proof of anonymity.

Pilot before scaling

A controlled sample or pilot can test whether preparation, quality and delivery match the requirement. Agree permissions, handling, compensation and acceptance criteria first. Recurring feeds and live-system access need their own authorised technical review. A description-first enquiry is the place to begin, not an invitation to connect a production database.