How well data fits the purpose of an AI use: accurate, complete, consistent, timely, and covering the cases that matter, with correct answers to test against.
Definition
Data quality for AI is the degree to which the data an AI use depends on is fit for that use. Common dimensions are availability, accuracy, completeness, consistency, timeliness, representativeness, and the existence of ground truth, meaning agreed correct answers to test against. Permission to use the data, the handling of sensitive content, and clear ownership belong beside them. Poor data in gives poor answers out, and a polished tool can hide the problem.
History
Data quality has been studied for decades, and standards such as the ISO 8000 series set out concepts and measures. In operations, the Lean idea of building quality in at the source applies directly: fix how data is entered, not only how it is cleaned.
When to Use
Assess data quality before choosing a use, before a pilot, and again at each review, because it drifts.
Step-by-Step
- List the data the use needs and where each item comes from.
- Score the dimensions with the people who create and use the data.
- Check accuracy on a random sample against the source and report the interval.
- Build the golden test set with experts.
- Fix the source: mistake-proof entry and define the codes.
- Decide: go, prepare first, or stop.
Examples
- Work-order data with 31% blank failure codes and 78% agreement in a sample check: prepare first.
- A complaint form changed to a drop-down list so the category cannot be left blank.
Common Pitfalls
- Assuming volume means quality.
- No accuracy check.
- No ground truth.
- One-time cleaning while the source stays bad.