What Is Data Quality? Dimensions, Checks and AI Impact

BusinessOperations and adoptionPublished Updated By Simon Budziak

Data quality is how well data fits the job you need it for. It is judged on dimensions such as accuracy, completeness, consistency, timeliness and validity, and for AI it decides whether a model answers from facts or from errors, because a model repeats bad data with the same confidence as good data.

Which data quality dimensions matter most?

Dimension Question Example failure
Accuracy Does the value match reality? Wrong VAT number on a supplier
Completeness Is everything required present? Contracts missing end dates
Consistency Do systems agree? Customer tier differs between CRM and billing
Timeliness Is it current enough? Last quarter’s price list in a sales assistant
Validity Does it follow the format and rules? Dates stored as free text

ISO/IEC 25012 defines a fuller model with 15 characteristics (ISO/IEC 25012:2008, read 12 November 2026). Quality is always relative to a use, so the bar for a board report and for an agent that pays invoices are different.

Why does AI raise the cost of poor data quality?

A person reading a report notices that a number looks odd. A RAG assistant quotes the outdated policy it retrieved, and an agent acts on the wrong bank detail. AI removes the human who used to catch bad data by accident. The EU AI Act asks that training, validation and testing data for high-risk systems be relevant, sufficiently representative and, to the best extent possible, free of errors and complete, in Article 10(3) (Regulation (EU) 2024/1689, read 12 November 2026). Fixing it upstream is data cleaning plus data governance that says who owns each fix.

How do we check data quality before we trust it?

In the automated pipelines we run, every data source gets a shape check separate from the judgment about its content. An error object, an HTTP 400, an empty response and a silently filtered row all look like “nothing interesting today” unless something asserts that the source actually answered. So an empty source is treated as a fault, not as a result. We also distinguish missing from low: when we join measured search data to a list of terms, a term with no matching row is marked unmeasured, never zero, and the rules treat it as neutral. When an agent hands work to another agent, the controller checks the content of the input files, such as non-empty counts per column, not only that the schema is right. A value you did not measure is not a value of zero.

Written with AI assistance and reviewed by Simon Budziak. The production notes come from systems Soba Labs builds and runs.

Frequently asked questions

What are the main dimensions of data quality?

The common ones are accuracy, completeness, consistency, timeliness, uniqueness and validity. Pick the few that matter for one use case and measure those, rather than scoring every table on every dimension.

Who is responsible for data quality?

The business owner of the data sets the quality bar, because only they know what good enough means for their process. Data stewards and engineers then measure it and fix the sources, not just the copies.

How does poor data quality affect AI?

A model trained or grounded on wrong, missing or stale data gives wrong answers fluently. In retrieval systems the effect is immediate, because the assistant quotes whatever document it finds, outdated or not.

Summarize this page with
ChatGPTClaudePerplexity

See this working in a system we built