The six dimensions work together. Accurate but incomplete data can still mislead. Complete but outdated data may not reflect current conditions. Consistent data can still repeat a flawed definition. Practitioners should therefore ask for a clear data-quality profile before training, evidence of checks during evaluation and monitoring after deployment. Each issue should have an owner, a likely effect and an agreed treatment, such as correcting, completing, standardising, refreshing, excluding, relabelling or collecting better evidence. The aim is not perfect data; it is known, controlled and suitable data for the AI’s purpose, risk and users.
Relevance asks whether the data genuinely helps the AI learn the intended task and whether unnecessary information has been removed. More data is not automatically better. Irrelevant fields can add noise, increase cost, create privacy risk or encourage the model to use a proxy that should not affect the outcome. Practitioners should ask why each source is needed and whether it will be available at the point the AI is used. The ICO recommends reviewing relevance throughout AI development and removing personal information that is not justified (ICO, nd a).
Consistency asks whether the same formats, categories, units, labels and meanings are applied across the dataset. Problems arise when ‘United Kingdom’, ‘UK’ and ‘GB’ are treated as different places; when costs are mixed between pounds and pence; or when one team defines an ‘active customer’ differently from another. Practitioners can reduce confusion by agreeing definitions and transformation rules before data is combined. Timeliness then asks whether the data is recent enough and available when needed, because real-world patterns change over time.
Integrity concerns whether data remains trustworthy as it is collected, transferred, combined, stored and changed. Practitioners need confidence that records have not been corrupted, changed without approval or separated from important context. Useful controls include source records, version control, access permissions, audit trails, checksums, lineage and clear ownership. Integrity helps a team explain which data and labels trained a model and why they were approved.
Completeness asks whether the necessary records, fields, groups and situations are present. Missing values may be easy to spot, but missing coverage can be harder. A dataset may include every required column yet exclude night shifts, rare faults, remote workers or customers who use accessibility support. The AI may then perform well on familiar cases and poorly where practitioners most need reliable support. Completeness means both filled records and realistic coverage of the situations the system will face.
Accuracy asks whether the values, labels and descriptions reflect reality. For example, if customer emails are labelled with the wrong topic, an AI tool may learn the wrong pattern and route future enquiries incorrectly. Accuracy can be checked against a trusted source or ‘ground truth’ through validation rules, sampling and expert review. Practitioners should remember that clean inputs reduce risk, but they do not guarantee every AI output will be correct.
Data Quality means the data is fit for the purpose it will be used for. A dataset can be good enough for one task but unsuitable for another. For example, customer comments may help an AI tool identify common themes, but they may not be complete enough to decide whether a complaint was resolved. ISO’s 5259 series addresses data quality for analytics and machine learning, while NIST identifies accuracy, completeness, integrity, consistency, timeliness and relevance as important quality components (ISO, 2024a; NIST, 2024a).
Unstructured data does not fit neatly into fixed fields. Examples include emails, policy documents, free-text notes, photographs, audio, video and call transcripts. This data often contains useful detail about real work, but it can be harder to search, label and compare. Practitioners should check whether files are duplicated, out of date, confidential, inconsistently labelled or separated from the context needed to interpret them safely.
Structured data follows an agreed format, such as rows and columns in a spreadsheet or database. Examples include customer ID, order date, product code and transaction value. This makes it easier to filter, compare and validate. However, practitioners should not assume that tidy data is trustworthy. A spreadsheet can still contain outdated records, wrong labels, duplicated entries or definitions that differ between teams.
Data is recorded information that can be stored, processed and used as evidence. A dataset is an organised collection of records. In a workplace, this could include sales transactions, maintenance logs, customer feedback, HR case notes, photographs of defects or chat transcripts from a support desk. For AI training, the data provides the examples the system learns from, so practitioners need to understand both what is included and what is left out.
Action Point
Choose one dataset that could support an AI use case in your organisation. Define the task in plain language, identify whether the data is structured, unstructured or mixed, then assess accuracy, completeness, integrity, consistency, timeliness and relevance. For each dimension, record one piece of evidence, the possible impact on practitioners’ decisions and the action needed before the data is approved for use.