Data quality: measuring it before anyone argues about it

Data quality is usually discussed as a complaint — the report is wrong, the addresses are stale, half the records have no phone number. It becomes tractable when it is expressed as measurable dimensions with rules attached, because then it can be reported on, targeted and argued about with numbers. The alternative is a permanent state of everyone believing the data is bad and nobody being able to say how bad or where.

The data quality dimensions worth measuring

• Completeness: is the field populated where it should be? Measured as a percentage against the records where it is required, not against all records.
• Validity: does the value conform to the rule — a date that exists, a country code from the list, an email with a plausible shape.
• Accuracy: does it match reality? The hardest and most expensive to measure, because it needs an external reference.
• Consistency: does the same fact agree across systems? This is where most reporting disputes actually originate.
• Uniqueness: how many duplicate records represent the same thing.
• Timeliness: how old is the value, and is that within the tolerance the process needs.

Turning complaints into rules

1. Take the last three data problems that caused real work and express each as a rule that would have caught it.
2. Attach the rule to a field defined in the data dictionary, so what is being measured is not itself ambiguous.
3. Set a threshold that reflects the business need rather than perfection — ninety-five per cent completeness on a field used for postal delivery may be fine, ninety-five per cent on a regulatory identifier is not.
4. Measure on a schedule and report the trend, because a single bad number provokes argument and a trend provokes action.
5. Assign each failing rule to the owner of the field, not to the data team, who can measure the problem but cannot usually fix its cause.

Fix the cause, not the records. A cleansing project that corrects two hundred thousand rows without changing the form that created them buys about six months. The cheapest data quality work is almost always a validation at the point of entry — a required field, a dropdown instead of free text, a check that runs before the record is saved.

What the numbers should drive

• Entry controls where validity fails: most bad data is typed, and most of it could have been rejected at the keyboard.
• A single source of truth where consistency fails, which is a decision about systems rather than a cleaning exercise.
• Deduplication with an agreed matching rule, recorded, because merging records is irreversible and disputes follow.
• Retirement of fields nobody maintains — an unused field with a poor score is a candidate for deletion under your data retention policy, not for a cleaning project.
• A visible owner per dataset, since unowned data reliably degrades regardless of tooling.

Manufacturing solved a version of this first. Total quality control puts responsibility for quality into every step rather than at a final inspection, with the people doing the work checking it as they go.

Loading a list is the cheapest moment to fix it. A contact import should deduplicate in the source file, carry consent and source across, and drop the columns nobody has filled in for years.

Where the rules and results live

Ettex Records holds the rules, their thresholds, the owner and the measured result over time, so the quality picture is a record with history rather than a spreadsheet regenerated whenever someone complains. Pair it with the data dictionary, which supplies the definitions the rules test against. Ettex does not profile databases, does not run validation over external systems and does not cleanse data — the measurement has to come from wherever the data lives. What this holds is the agreement about what good looks like and the evidence of whether it is being met.