Data quality: why two teams bring different figures to the same meeting

When two teams bring different figures for the same metric, there is almost never a calculation error: there are two definitions coexisting. Data quality is measured across six dimensions — accuracy, completeness, consistency, timeliness, uniqueness and validity — and that committee argument usually starts in the third.

Below: the six dimensions in plain language, which one fails in each typical symptom, what to fix first, what it costs to leave it, and how to sustain it without a permanent project.

The question "which of the two figures is right?" has an uncomfortable answer: frequently both, because they measure different things under the same name. And a useful one: the problem is not in the report, it is upstream.

To fix it you have to be able to name exactly what is failing, and that is what decomposing quality into dimensions is for.

The dimensions are not an academic taxonomy. They exist because each one has a different cause and a different owner: completeness is usually a capture problem at the point of entry, consistency is a definition problem between two teams, and timeliness is an architecture problem in how the data moves. Treating all three as "bad data" sends the same generic fix to three problems that share nothing but a symptom.

The six dimensions, in plain language

Accuracy: the value reflects reality. An address that exists but is not the customer's is complete and false at once.

Completeness: the fields you need are populated. Usually measured badly, because a field containing "N/A" counts as filled.

Consistency: the same fact is represented identically across systems. This is the one that produces two figures in a committee.

Timeliness: the value is current enough for the decision being made. Yesterday's inventory is fine for planning and not for promising a delivery.

Uniqueness: each entity appears once. Duplicate customers inflate counts and ruin any per-customer analysis.

Validity: the value satisfies the rule that applies to it. An identifier in the wrong format is valid to the system that accepted it and not to the one receiving it.

Which one fails in each typical symptom

If totals do not reconcile between teams, the problem is consistency: there are two definitions. If counts look high, check uniqueness before anything else. If a segment analysis concludes nothing, it is usually completeness in the field doing the segmenting.

If the decision arrived late, the problem is timeliness rather than quality in the strict sense: the value was correct and arrived after it was needed.

Naming the dimension matters more than it sounds. "The data is bad" produces a project; "the customer identifier is inconsistent between billing and CRM" produces a fix.

What to fix first

The temptation is to clean everything. The alternative that pays is to ask which decision is being made badly because of this data, and correct only what that decision needs.

And to correct at source, not in the report. A value fixed in the dashboard will be wrong again next time, and it creates a third version of the truth. Upstream correction costs more the first time and stops costing afterwards.

Where the source cannot be touched — a third party's system, for instance — the cleaning rule should live in exactly one place along the path and be documented, not scattered across five reports nobody knows exist.

What it costs to leave it

The visible consequence is committee time spent reconciling figures, and it is the least serious one. The serious consequence is that the organisation starts deciding by instinct, because the data stopped being an authority: when every figure is arguable, the better arguer wins.

The second is duplicated work. Each team builds its own version of the truth rather than depend on a number it does not trust, and the organisation pays three times for the same analysis.

Neither appears in a budget line, which is why a data quality initiative competes at a permanent disadvantage against work whose benefit can be invoiced. Naming both costs explicitly is usually what gets it funded.

How to sustain it without a permanent project

By measuring little and always. Three or four quality indicators over the data that supports important decisions, reviewed on a fixed cadence, achieve more than an exhaustive annual audit.

And with an owner per domain. Without one, quality degrades between audits at the same speed it was corrected.

The cadence matters more than the depth. A check that runs weekly and looks at four fields catches a break within days; one that runs annually and looks at four hundred catches it after the decisions have already been made on bad data.

The definition is the artefact, not the dashboard

The durable output of a data quality effort is not a cleaner table: it is a written definition of each important metric, agreed by the people who use it, stored where anyone can find it without asking.

It sounds bureaucratic and it is the cheapest thing on this list. "Active customer" written down once — with the inclusion rule, the exclusion rule and the date the definition took effect — ends an argument that otherwise recurs every quarter with different participants.

It also gives the next change somewhere to land. When the business decides that active now means ninety days rather than sixty, the change is made in one place, dated, and the history stays interpretable. Without that record, an old report and a new one disagree and nobody can say which definition each one used.

The failure mode to avoid is a catalogue nobody reads. Definitions belong close to where the figures are consumed — in the report, next to the number — not in a governance repository that requires knowing it exists.

The check that costs nothing and finds the most

Before building any quality framework, there is one exercise worth running: take the three figures that reach the executive committee and trace each one back to the records that compose it, by hand, once.

It takes an afternoon and it usually surfaces more than a tool would in a month — a filter nobody remembers applying, a date range that shifts, a category that was renamed and half the history not migrated.

It also does something a tool cannot: it tells you whether anyone in the organisation can currently answer where a number came from. If nobody can, that is the finding, and no amount of monitoring fixes it.

Where to place the effort

Trust in a figure is lost at the weakest stretch of its path, and that stretch is rarely the one suspected. Measuring it is cheaper than guessing: a data and analytics maturity assessment scores source, quality, governance, model and consumption separately.

Governance of the data is also read against the framework that applies — Habeas Data in Colombia, the federal data protection law in Mexico — and not only against a generic maturity model. From there comes an analytics plan with defensible priorities.

Frequently asked questions

Why do two teams present different figures for the same metric?

Almost always because two definitions coexist, not because of a calculation error. It is a consistency problem, and it is solved by agreeing and writing the definition rather than recalculating the report.

How is data quality measured?

In six dimensions: accuracy, completeness, consistency, timeliness, uniqueness and validity. Naming which one is failing is what allows fixing the source instead of patching the report.

Should all the data be cleaned?

No. It pays more to ask which decision is being made badly because of a data item and correct only what that decision needs. A general clean-up consumes the budget before demonstrating value.

Fix it in the report or at the source?

At the source wherever possible: a value fixed in the dashboard will be wrong again next time and creates a third version of the truth. If the source belongs to a third party, the rule must live in one documented place along the path.

What is the cheapest useful first step?

Take the three figures that reach the executive committee and trace each back to its records by hand, once. It takes an afternoon and usually finds more than a tool would in a month.

How is it sustained without a permanent project?

Three or four indicators over the data that supports important decisions, reviewed on a fixed cadence, plus an owner per domain. Cadence matters more than depth.