An enterprise data team prepares a customer dataset for a machine learning project. The records contain all expected fields, but customer identifiers and outcome labels do not line up consistently, so the team cannot yet tell whether the dataset is suitable for training.
The UK Government Data Quality Framework describes data quality in terms of fitness for purpose. The framework explains that required quality levels vary according to the intended purpose and identifies dimensions including completeness, uniqueness and timeliness. Completeness describes the degree to which records are present.
For AI applications, quality standards need to cover how the fields relate to the target task. CUBIG’s data quality for AI article explains why checking fields one by one may not reveal whether the data correctly represents the outcome the model must predict.
Common dimensions addressed by data quality standards include:
- Completeness: Check whether required records and fields are present, as described in CUBIG’s entry on data quality dimensions.
- Uniqueness: Identify duplicate records that could distort analysis or model training.
- Consistency: Check whether related fields and records follow compatible rules and definitions.
- Timeliness: Assess whether information is current enough for its intended use.
- Accuracy: Evaluate whether recorded values correctly represent the underlying entities or events.
How are data quality standards different from data quality dimensions?
- Data quality dimensions describe the aspects used to evaluate data, such as completeness, uniqueness and timeliness.
- Data quality standards define the agreed expectations, thresholds and rules applied to those dimensions.
- A dimension provides a way to describe quality, while a standard establishes what level is acceptable for a particular purpose.
Where do data quality standards break in AI work?
Generic standards can overlook relationships that matter to a specific prediction task. In a CUBIG analysis of the synthetic UCI AI4I 2020 predictive maintenance dataset, published in November 2026, the dataset contained 10,000 records and zero missing values, yet the failure label disagreed with the five failure-mode flags in 27 records. All 27 records would have passed a completeness check.
The example illustrates why completeness alone cannot establish whether data is suitable for model training. Teams need additional checks for label consistency, feature relationships and task-specific assumptions. Data quality monitoring can help identify changes over time. AI-ready data also requires checks tied to the intended AI task.