Data Quality: Dimensions, Checks, and Responsibilities
Decisions are only as good as the data behind them. Poor quality data wastes time, damages trust, and leads to wrong conclusions. Data quality work gives teams a vocabulary and a set of routines for preventing that.
Common dimensions of quality
- Accuracy: does the value reflect reality?
- Completeness: are required values present?
- Consistency: do different systems agree with each other?
- Validity: does the value follow the allowed format or range?
- Uniqueness: is each real-world thing recorded once?
- Timeliness: is the data fresh enough for its use?
Different uses prioritize different dimensions. A monthly report tolerates some delay, whereas a stock availability display does not.
Where problems come from
Human entry errors, system integration mismatches, changes in definitions, missing validation at the point of capture, and unrecorded manual fixes all contribute. Often the issue is not a bug but a misunderstanding, such as two departments counting “active customer” differently.
Practical checks
- Schema checks: required fields exist with the right types.
- Range checks: ages, prices, and dates fall within plausible limits.
- Referential checks: every order refers to a real customer.
- Duplicate checks: similar records are flagged for review.
- Freshness checks: data was updated within the expected window.
- Volume checks: counts are not suddenly far above or below normal.
- Distribution checks: the mix of categories has not shifted strangely.
Run checks automatically and report results where people will see them.
Profiling before fixing
Profile a dataset first: count blanks, list distinct values, view minimums and maximums. Profiling reveals what is actually in the data, which is often different from what documentation says.
Fix at the source
Cleaning downstream is a stopgap. Repeatedly fixing the same problem in reports means the cause remains. Add validation to forms, standardize picklists, and agree on definitions. Feedback loops from data users to data producers matter greatly.
Roles and accountability
- Data owners decide meanings and acceptable quality levels.
- Stewards monitor and coordinate fixes.
- Engineers implement checks and pipelines.
- Users report anomalies.
Without named responsibility, quality becomes everyone’s problem and therefore no one’s.
Keep a data dictionary
Document each field: meaning, format, source, and owner. A shared dictionary prevents arguments and helps newcomers.
What AI can and cannot do
Assistants can suggest rules, find unusual patterns, or standardize messy text. However, they may also “correct” values incorrectly or fill gaps with guesses. Any automated repair should be logged and reviewable.
A realistic goal
Perfect data does not exist. Aim for data that is fit for purpose, with known limitations written down, and a process for improving it where it matters most.
A small worked scenario
Suppose a retailer notices that revenue per region differs between the finance report and the sales dashboard. Investigation shows that one system records orders on the day they are placed and the other on the day they ship. Neither is wrong; the definitions simply differ. The fix is a written agreement on which date drives each report, plus a label on both dashboards. Many quality incidents turn out to be definition gaps like this, which is why conversations between data producers and data users often accomplish more than another cleaning script.
Start where it hurts
Rank datasets by how many decisions depend on them, and apply the checks above to the top few first. Expanding coverage gradually is far more sustainable than attempting to certify every table at once.