Data quality: the dimensions that define whether you can trust your data

Quick summary: Data quality is the degree to which data serves its intended use. It is not measured with a single yardstick, but by dimensions: completeness, uniqueness, accuracy, consistency, timeliness and validity. A piece of data can be "right" and still be useless if it is outdated, duplicated or off-standard.

"But the data is there"

A phrase that shows up a lot in projects: "the data is there, the field is filled in". Being filled in doesn't mean being right. A CNPJ (Brazil's company registration number) can be in the right field and still belong to another company. An address can exist and be three years out of date. A material can have a description, and that description can say nothing useful. Data quality is about trust, not about full fields.

The point is that quality is not one thing. That is why we talk about dimensions.

The main dimensions of data quality

Completeness is about not missing what matters. A supplier master record without bank details, or a material without a unit of measure, is incomplete for its use. Completeness is not filling in every existing field, it is filling in what the process needs.

Uniqueness is about having no duplicates. Each real entity should have a single record. Two records for the same customer violate uniqueness and break any count, any consolidated view.

Accuracy is about the data reflecting reality. Does the name match the document? Is the price the real price? Accuracy is the hardest dimension to measure on its own, because it often requires checking the data against a reliable external source.

Consistency is about the data not contradicting itself across systems. If the customer is "active" in the CRM and "inactive" in the ERP, there is an inconsistency. Consistency is what lets different areas arrive at the same number.

Timeliness is about the data being current. An address that was correct in 2022 may be wrong today. Outdated data is wrong data that looks right.

Validity is about the data respecting format and domain rules. A birth date in the future, a CEP (Brazilian postal code) with the wrong number of digits, a code outside the allowed list: all of that is invalidity.

Whoever treats these dimensions separately can see where the base really fails, instead of saying generically that "the data is bad".

Why data quality is so much work

Because bad data almost never screams. It doesn't crash the system. It goes through. The order is placed, the invoice is issued, the report is generated. The problem shows up later, far from the source, disguised as something else: the purchase that came out more expensive, the tax paid in excess, the customer who received at the wrong address, the demand forecast that missed.

Since the symptom appears far from the cause, it is rare for anyone to connect the dots and say "this was bad data". That is why data quality needs to be measured proactively, with indicators, and not only when the problem blows up.

How to measure data quality

Measuring starts by choosing the rules that matter for each domain. In a material master, for example: percentage of items with a description in the standard, percentage with a valid NCM (Mercosur Common Nomenclature, Brazil's tax classification code), percentage without duplicates, percentage with a coherent unit of measure. Each rule becomes an indicator, and the set becomes a quality dashboard.

The value of measuring is not just knowing where things are bad. It is tracking the evolution. Good governance makes quality rise over time and sustains the level. Without measurement, the company thinks it improved because it did a cleanup push, and doesn't notice that the base degraded again months later.

Data cleansing is not the end, it is the beginning

Many companies treat quality as an event: they hire a data cleansing, clean the base, celebrate and move on. Months later, the mess is back. That happens because data cleansing without governance is bailing water with a bucket. As long as the entry process allows bad data, the base will get dirty again.

Sustainable quality comes from validating at entry, with a clear rule, so the error isn't born. Fixing later is always more expensive than preventing it at the source. On cleaning up the base, see saneamento de dados.

Data quality and AI

AI models are only as good as the data that feeds them. An AI trained on a base with duplicates, empty fields and wrong classifications will learn those flaws. Worse: it will present them with confidence, which makes the error harder to notice. Before dreaming of sophisticated AI, the most profitable step is usually to raise the quality of master data. It is less glamorous and more effective.

A practical example of how the dimensions combine

Take a supplier master record to see the dimensions at work together. A supplier may have the CNPJ filled in (looks complete), but that CNPJ belongs to another branch (accuracy failure). The name may be correct, but also be recorded under a small variation in another record (uniqueness failure). The bank details may be right, but be two years old, from before the supplier changed banks (timeliness failure). And the tax status field may say "regular" in the record while the official system says something else (consistency failure).

None of these problems crashes the system. The payment is scheduled, the purchase is released, everything seems to work. Until the money lands in the wrong account, or the purchase comes from a non-compliant supplier. The point is that data quality is not a single grade; it is a set of checks that must pass together. Looking only at "the field is filled in" is seeing one dimension and ignoring the other five.

Who is responsible for data quality, anyway

A question that always comes up: whose job is data quality? The answer that tends to go wrong is "IT's". IT looks after the system, the integration, the infrastructure. But it doesn't know whether the supplier changed banks, whether the material is classified in the right family, whether the customer is the same as that other record. The business is who knows that. That is why data quality is a shared responsibility, with the business defining what is right and IT providing the tools to sustain it.

When quality is dumped entirely on IT, it dies, because IT has no way to judge the content of the data. And when it is dumped entirely on the business without tools, it also dies, because there is no scale. Sustainable quality lives in the partnership: rules coming from the business, measurement and automation coming from technology, and someone clearly responsible for tracking the indicators. Without that clear division, quality ends up belonging to everyone and no one, which is the recipe for it never really improving.

MDM Academy: data quality as a skill

Knowing how to measure and improve data quality is one of the most sought-after skills in governance. MDM Academy, 4MDG's school, trains professionals who master these dimensions in practice, based on 4MDG's experience in data cleansing and governance projects for large databases. If you want to learn to diagnose a base, define indicators and build quality that lasts, the MDM Academy training was designed for that.

Data quality is not perfection

A misunderstanding that hurts quality projects is thinking the goal is perfect data. It isn't. Chasing perfection in every field of the whole base is expensive, slow and, in most cases, unnecessary. The quality that matters is quality for the use. A field nobody uses doesn't need to be flawless. A field that decides tax, payment or purchasing needs to be right. Good quality work prioritizes: it identifies which data, in which dimensions, has real impact, and concentrates effort there.

That logic avoids two bad extremes. On one side, neglect, which treats all data as irrelevant and lets the base rot. On the other, perfectionism, which spends energy polishing data that doesn't change any decision. The balance lies in linking quality to consequence: where an error is costly, rigor is demanded; where it isn't, good enough is accepted. Defining that criterion is a business decision, not a technical one, and it is what makes a quality program sustainable instead of a bottomless pit. Smart quality is about focusing effort where it pays off, not about chasing an ideal that never arrives.

Frequently asked questions about data quality

What is data quality? It is the degree to which data serves its intended use, measured by dimensions such as completeness, uniqueness, accuracy, consistency and timeliness.

What are the dimensions of data quality? The most commonly used are completeness, uniqueness, accuracy, consistency, timeliness and validity.

How do you measure data quality? By defining rules per domain, turning each one into an indicator and tracking the evolution on a quality dashboard.

Why does the base get bad again after data cleansing? Because data cleansing without governance at entry doesn't stop the error from coming back. You need to validate at the source.

Does data quality influence AI projects? Directly. AI learns from data; a bad base produces a bad model, with errors presented convincingly.