"Is our data any good?" is a hard question to answer, because data can be good in one way and poor in another. A customer list can have every field filled in and still hold addresses that are two years old.
Data practitioners deal with this by splitting quality into six dimensions. DAMA UK, the UK chapter of the Data Management Association, set out six core dimensions in a 2013 paper, and the UK government adopted them in its Data Quality Framework in 2020. The list is not fixed: GOV.UK calls it "not a prescriptive list", and DAMA NL's research found a longlist of 60 dimensions.
This article explains each dimension with an example, then gives a five-step way to use them without measuring everything.
What data quality means
The Government Data Quality Framework defines quality as fitness for purpose: "is this data set good enough for what I want to use it for?"
That definition matters because it makes quality depend on the use. Figures that are fine for a quarterly trend may be useless for a decision that has to be made today.
The framework is also clear about the cost of ignoring the question. "Poor or unknown quality data weakens evidence, undermines trust, and ultimately leads to poor outcomes."
The six dimensions at a glance
| Dimension | The question it answers | Example from GOV.UK guidance |
|---|---|---|
| Completeness | Is all the data needed for this use present? | Missing information on a patient's allergies |
| Uniqueness | Is each thing recorded only once? | Two patient records that refer to the same person |
| Consistency | Do values agree within a record and across data sets? | A date of birth that differs between two data sets |
| Timeliness | Is the data available when it is expected and needed? | A hospital bed allocation system that needs the most up-to-date data |
| Validity | Does the data fit the expected format, type and range? | An eye colour of "blue" is valid |
| Accuracy | Does the data reflect reality? | A person's height recorded as 15 cm |
Each dimension in practice
The examples below are illustrations from everyday business data.
Completeness
Data is complete "when all the data required for a particular use is present and available to be used". The words "for a particular use" matter. An empty fax number field is not a quality problem. A customer record with no sector is a problem if you report sales by sector.
How to check: for each use, list the fields it requires and measure the share of records that have them.
Uniqueness
Uniqueness measures duplicates. The same customer entered twice, once as "J. Smith Ltd" and once as "John Smith Limited", inflates customer counts and splits their history across two records.
How to check: look for duplicate identifiers, then for near-matches on name, address and email.
Consistency
Consistency means values do not conflict within a record or across data sets. If finance reports one revenue figure for a client and the CRM reports another, one of them is wrong, or the two systems define revenue differently.
How to check: reconcile the same measure across the systems that hold it, and write down the agreed definition. Time periods often trip people up here: a monthly total from a system that closes on the last working day will not match one that closes on the last calendar day.
Timeliness
Timeliness asks whether data is available when it is expected and needed. The GOV.UK guidance points out that "timeliness means different things for different uses". Month-end figures that arrive on day ten are fine for an annual review and too late for a pricing decision.
How to check: agree when each use needs its data, then measure the gap between that and when it arrives.
Validity
Validity is "the extent to which the data conforms to the expected format, type, and range". A UK postcode in the wrong pattern is invalid. So is a date of death in the future.
How to check: apply format and range rules where the data is entered, so invalid values are rejected at the start.
Accuracy
Accuracy means the data reflects reality. It is the hardest dimension to measure, because you need something trustworthy to compare against, such as the original document or a call to the customer.
How to check: sample records and compare them with the source.
Two distinctions people miss
Valid is different from accurate. The GOV.UK guidance gives the example that a value of "blue" for eye colour is "valid but incorrect if it is actually brown". Format checks catch invalid data. They cannot tell you a well-formed value is wrong.
Complete is different from accurate. In the guidance's words, "a full data set may still have incorrect values". A dashboard showing 100% of fields populated says nothing about whether the contents are right.
The guidance also states that "a single dimension will not be sufficient to assess the quality of your data". One overall quality score hides which problem you have.
How to use the dimensions: five steps
- Start from one decision or report. Choose the one your leadership team relies on most.
- List the data it depends on. The Government framework's second principle is to "prioritise efforts on the data which is most critical". For one report, that is usually a short list of fields.
- Choose the dimensions that matter for that use, and set a threshold for each. Ask the people who use the report how they can tell when its figures look wrong, and set thresholds for what the use requires rather than for perfection. A monthly board pack might need 98% completeness on five fields and figures no more than three working days old.
- Measure and share the result. Publish your data quality measures alongside the report, so readers know how far to trust the figures.
- Fix problems at the source and name an owner. Cleaning data downstream has to be repeated every month. A check at the point of entry is done once. Where a report depends heavily on data from a supplier or another team, agree service levels with them against your data quality measures.
Repeat for the next report once the first is under control.
Why this matters more with AI
AI tools make data quality problems visible quickly, because they produce confident output from whatever they are given.
UK government research on AI adoption found that 21% of businesses cited data complexity, meaning missing or unstructured data, as a barrier. Among the businesses that cited it, 70% rated it a significant barrier, and among large businesses 84% did.
An organisation that knows the quality of the data behind its top priorities can say which AI uses are safe to start now and which need groundwork first.
Automating data quality checks as part of your data pipeline reduces manual reviews and the back-and-forth about whether the data can be trusted. For Spirits Spyder, this cut the time taken to complete tasks by 80% and made the data more consistent (client review on Clutch).
Common mistakes
- Measuring everything. Quality work spread across every data set never finishes. Tie it to a use.
- Expecting perfection. Waiting for every measure to be perfect means the data never gets used. Define "good enough" thresholds instead.
- Reporting one score. It cannot tell you whether to fix duplicates or chase late data.
- Cleaning without fixing the cause. If the entry form allows the error, it will come back.
Our data governance service covers ownership, quality checks and the rules around them. For more on the cost of poor data, read Why messy data is costing your business and Building confidence in your data foundations.
Sources
- Government Data Quality Hub, The Government Data Quality Framework, 3 December 2020.
- Government Data Quality Hub, Meet the data quality dimensions, 24 June 2021.
- DAMA UK Working Group, The Six Primary Dimensions for Data Quality Assessment, October 2013. The download is for DAMA UK members only.
- Department for Science, Innovation and Technology, AI Adoption Research, IFF Research and Technopolis, updated 13 February 2026.
- DAMA NL Data Quality Working Group, Dimensions of Data Quality.
- Spirits Spyder, client review of Data Understood, Clutch.
Contains public sector information licensed under the Open Government Licence v3.0.
