Data quality
Also: DQ, Data Quality, Information quality, Qualite des donnees
The degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.
What it is
Data quality measures how well a dataset serves the purpose it is used for. It is not an abstract ideal but a judgment relative to a specific use: data that is good enough for a marketing segmentation may be unfit for financial reporting. Quality is usually broken down into six widely used dimensions:
- Accuracy: values correctly describe the real world (a customer's address is their actual address).
- Completeness: required fields and records are present (no missing revenue lines).
- Consistency: values agree across systems (the same customer has the same status in CRM and billing).
- Timeliness: data is current enough for the decision (yesterday's inventory, not last month's).
- Validity: values conform to defined rules and formats (a date is a real date, a country code exists).
- Uniqueness: no unintended duplicates (one record per customer, not three).
Why it matters
Every downstream use inherits the quality of its inputs. Dashboards, regulatory filings, machine learning models and customer communications all degrade when the underlying data is wrong, stale or duplicated. The cost is rarely visible as a single line item; it shows up as rework, wrong decisions, compliance risk and eroded trust in data.
How it is used in practice
Organizations operationalize data quality through:
- Profiling: scanning data to discover patterns, ranges and anomalies.
- Rules and validation: encoding expectations (for example, `age between 0 and 120`).
- Scorecards and thresholds: tracking dimension scores over time and alerting when they drop.
- Remediation: fixing at the source, deduplicating, and enriching from trusted references.
- Ownership: assigning data stewards accountable for specific domains.
Quality is measured continuously, not once. A common practice is to attach a data quality score (percentage of records passing rules) to each critical dataset.
Worked example
A company has 100,000 customer records. Checks reveal:
- 4,000 duplicate records (uniqueness fails).
- 7,000 missing email addresses (completeness fails).
- 2,000 invalid postal codes (validity fails).
If a marketing campaign targets by email, roughly 7 percent of customers are unreachable and duplicates inflate reported reach. Deduplicating and validating first raises campaign accuracy and lowers wasted spend. The same cleaned dataset then feeds finance and AI teams, so one fix compounds across functions.
Frequently asked questions
What does data quality actually mean?
Data quality measures how well a dataset serves the purpose it is used for. It is a judgment relative to a specific use, not an absolute standard: a customer file that works for marketing segmentation can be unfit for financial reporting. It is usually assessed along six dimensions, accuracy, completeness, consistency, timeliness, validity and uniqueness.
What are the six dimensions of data quality?
Accuracy (values describe the real world), completeness (required fields and records are present), consistency (values agree across systems), timeliness (data is current enough for the decision), validity (values respect defined rules and formats) and uniqueness (no unintended duplicates). Each dimension fails differently, so each needs its own checks: a duplicate record is a uniqueness problem, an empty email field is a completeness problem.
Who in a leadership team is concerned by data quality?
Marketing, finance, data and AI leaders all inherit the quality of the same inputs. A CMO sees it as unreachable contacts and inflated reach, a CFO as filings that need rework, a CDO as eroded trust in dashboards, an AI lead as models trained on wrong values. That shared dependency is why one cleanup effort compounds across functions.
How do you measure data quality in practice?
By encoding expectations as rules (for example, age between 0 and 120), running them continuously, and tracking the share of records that pass. A common practice is to attach a data quality score, the percentage of records passing rules, to each critical dataset, with thresholds that trigger an alert when scores drop. Profiling comes first to discover patterns, ranges and anomalies before writing the rules.
What does poor data quality cost on a 100,000-record customer file?
Take a file with 4,000 duplicates, 7,000 missing email addresses and 2,000 invalid postal codes. An email campaign cannot reach roughly 7 percent of customers, while the duplicates inflate reported reach, so both the spend and the performance reading are wrong. Deduplicating and validating before the send raises targeting accuracy, and the cleaned file then serves finance and AI teams too.