Data quality is not a single score. It breaks down into six measurable dimensions: accuracy, completeness, freshness, consistency, uniqueness, and validity. Each needs its own KPI, and each must be re-enforced at every data refresh, not audited once and forgotten.
Most teams talk about data quality as if it were one thing. It is not. When a dashboard shows the wrong revenue figure or an AI model makes an odd prediction, the root cause is almost always a failure along one specific axis. Naming that axis is the first step to fixing it.
The discipline breaks data quality into six dimensions. Understanding each one - how to measure it, why it matters, and how to keep it healthy - turns a vague complaint ("our data is bad") into a concrete, trackable engineering problem.
What are the six data quality dimensions?
The six dimensions describe distinct ways a dataset can be right or wrong. A record can be perfectly accurate yet arrive too late to be useful. It can be complete yet contain duplicates. Each dimension is independent, which is exactly why you need all six. This six-part breakdown is not a K.IOP invention: it is a recognised industry baseline, formalised by the DAMA UK Working Group in its paper The Six Primary Dimensions for Data Quality Assessment (Source: DAMA UK, 2013).
- Accuracy - does the value match reality?
- Completeness - is anything missing?
- Freshness (timeliness) - is the value current enough?
- Consistency - does it agree with itself across systems?
- Uniqueness - is each entity represented once?
- Validity - does it conform to the rules and formats it should?
Let us take them one at a time.
How do you measure accuracy?
Accuracy is the degree to which a value reflects the real-world thing it describes. A customer's registered address is accurate if a letter sent there arrives.
Example: A supplier's VAT number is stored as valid in format but belongs to a different company. The field passes a format check but fails accuracy - and the tax filing built on it is wrong.
Accuracy is the hardest dimension to measure because it requires a trusted reference. Practical KPIs include the match rate against an authoritative source (a company register, a verified address service, a golden record) and the exception rate flagged during reconciliation. You measure accuracy by comparing, not by inspecting a field in isolation.
How do you measure completeness?
Completeness asks whether all the data that should be present actually is - at field level, record level, and dataset level.
Example: A CRM holds ten thousand leads, but forty percent have no country field. Any regional revenue analysis built on that data silently excludes those leads, skewing the result.
Useful KPIs:
- Percentage of populated values per critical field
- Ratio of records meeting a minimum required-field threshold
- Volume completeness - did today's load contain the expected number of rows?
Completeness is not about filling every field. It is about ensuring the fields that matter for a given business use are present. A missing middle name rarely matters; a missing IBAN on a payment record always does.
How do you measure freshness?
Freshness, or timeliness, measures whether data is current enough for the decision it supports. The acceptable window depends entirely on the use case.
Example: An inventory dashboard refreshed nightly is fine for procurement planning but useless for a live order-promising engine that needs stock levels updated within minutes.
Measure freshness with:
- Data latency - the gap between an event happening and it appearing in your systems
- Refresh lag - time since the last successful load versus the agreed schedule
- The share of records older than the tolerance defined for their purpose
Freshness failures are insidious because the data looks correct. It is simply describing a world that no longer exists.
See how a refinery approach enforces all six dimensions at every refresh.
How do you measure consistency?
Consistency means the same fact holds the same value everywhere it appears. When two systems disagree, at least one is wrong - and often nobody knows which.
Example: A customer's status is "active" in the billing system and "churned" in the CRM. Support, finance, and marketing each act on a different truth.
KPIs for consistency:
- Cross-system match rate for shared entities
- Number of reconciliation breaks between source and target
- Referential integrity violations (orphaned records, broken keys)
Consistency is where data reconciliation earns its keep. Comparing two systems record by record and surfacing every difference is the only reliable way to know where they diverge.
How do you measure uniqueness?
Uniqueness means each real-world entity appears exactly once. Duplicates inflate counts, split histories, and corrupt any metric that involves counting customers, products, or transactions.
Example: "ACME Ltd", "Acme Limited", and "ACME LTD." are three records for one company. Revenue is spread across all three, so the account never looks large enough to prioritise.
Measure uniqueness with:
- Duplicate rate - records identified as referring to the same entity
- Number of fuzzy-match clusters requiring merge
- Post-deduplication reduction in record count for a known entity type
Uniqueness cannot be solved by exact matching alone. Real duplicates differ in spelling, spacing, and formatting, which is why deterministic fuzzy matching and clear merge rules are essential.
How do you measure validity?
Validity is conformance to the rules a field should obey - format, type, range, and allowed values. It is the most mechanical dimension and the easiest to automate.
Example: A date of birth in the year 2189, a phone number with letters, or an order quantity of minus five all fail validity even before anyone checks whether they are accurate.
KPIs:
- Percentage of values passing format and pattern rules
- Range and domain violation counts
- Business-rule pass rate (for example, end date must follow start date)
Validity is your first line of defence. It cannot confirm a value is true, but it reliably catches values that cannot possibly be true.
Why do these dimensions matter for BI and AI?
Business intelligence inherits every flaw in its inputs. A dashboard is only as trustworthy as the least reliable dimension feeding it. One duplicated account or one stale figure is enough to make executives doubt the entire report - and once trust is gone, adoption follows. The cost is not abstract: Gartner estimates that poor data quality costs organisations an average of $12.9 million per year (Source: Gartner, 2021).
AI raises the stakes further. Models trained or prompted on inconsistent, incomplete, or duplicated data learn the noise along with the signal. Poor uniqueness biases training sets. Weak freshness means a model reasons about an outdated reality. Invalid values produce unpredictable behaviour. Data quality is not a nice-to-have upstream of AI; it is the precondition for it.
How does a refinery approach enforce quality at every refresh?
The common mistake is treating data quality as a one-off cleanup project. You profile the data, fix what you find, produce a report, and move on. Weeks later the same problems return, because every new load reintroduces them. This is also why cleanup stays so expensive: data professionals report spending, on average, 45% of their time simply getting data ready - loading and cleansing it - before they can put it to work (Source: Anaconda, 2020).
Quality has to be enforced continuously, at the point where data flows. This is the logic of a data refinery. The Data Refinery engine behind K.IOP - K-Refine - applies the same deterministic pipeline on every refresh:
- Profiling measures each of the six dimensions and surfaces where they break.
- Business rules enforce validity and completeness against your definitions, not generic defaults.
- Deduplication resolves uniqueness with deterministic fuzzy matching and explicit merge logic.
- Reconciliation compares systems to enforce consistency and flag accuracy exceptions.
Every action is deterministic and recorded in a full audit trail - no guesswork, no hallucinations. That matters for governance and for regulated environments in the UAE, where the PDPL (Federal Decree-Law No. 45 of 2021) and frameworks in the DIFC and ADGM expect you to demonstrate how personal and business data is handled and corrected.
K.IOP sits within the K-Technology & Services group and follows a simple arc - Connect, Refine, Operate - with its sister brand K inventory covering asset and inventory data. The point is not a single clean-up. It is quality that holds at every refresh.
If you want to know where your organisation stands today, a structured Data governance maturity assessment is a practical starting point.
FAQ
Are there only six data quality dimensions?
Six is the widely used core set, but some frameworks add integrity or accessibility as separate dimensions. The six here cover the failures that most affect reporting and AI, which is why they are the standard baseline.
Which dimension should we fix first?
Start with validity and completeness, since they are the cheapest to automate and block the most obvious errors. Then move to uniqueness and consistency, which usually cause the biggest analytical distortions.
Can one KPI summarise all six?
You can build a composite score, but it hides where the problem is. Track each dimension separately so a failure points you directly to the cause rather than an average that means little.
How often should data quality be checked?
On every refresh. Quality degrades as new data arrives, so a pipeline that re-profiles and re-enforces rules at each load is far more reliable than a periodic audit.
Does data reconciliation replace the other dimensions?
No. Reconciliation is the strongest tool for consistency and accuracy, but validity, completeness, freshness, and uniqueness each need their own checks. The dimensions are complementary, not interchangeable.
