Home › Cybersecurity & GRC Career Guides › Data Quality Skills

Data Quality Skills

Get new GRC career guides by email. We publish most weeks. No spam, and we never sell or share your address with anyone.

Data quality work answers one question: is this data fit for the decision somebody is about to make with it? Not whether it is perfect, which is unachievable and would not be worth paying for. Fit for a stated purpose.

Where the data came from and what depends on it is a related but separate discipline, covered under lineage.

The dimensions, and the one that causes the arguments

Completeness, accuracy, consistency, timeliness, validity and uniqueness are the standard set. Five of them are testable with a rule. Accuracy is the exception, and it is where most of the difficulty lives, because accuracy means agreement with the real world and your systems do not contain the real world.

You can check that a date of birth is a valid date. Confirming it is the customer's actual date of birth requires a source outside the system. Teams that say they measure accuracy are usually measuring validity, and knowing the difference is a quick way to sound like you have done this.

Fitness beats perfection

The same dataset can be excellent for one use and dangerous for another. Marketing addresses that are ninety-four percent correct are fine for a campaign and unacceptable for regulatory reporting.

So the useful question is never "is this data good". It is "good enough for what, and who decided". Programs that chase a global quality score produce dashboards. Programs that set thresholds per use case produce decisions.

Profile before you write rules

The instinct is to start writing validation rules. Profile first: distributions, null rates, cardinality, format variation, outliers, and how all of that has moved over time.

Profiling routinely reveals that the actual problem is not the one anyone was discussing. A field is populated but with a default value nobody noticed. Two systems use the same field name for different things. A file that arrives daily has been arriving with the same timestamp for three weeks.

Fix upstream or keep paying forever

Every quality problem can be corrected in the pipeline. Doing so guarantees you will correct it again every day, forever, and that the next consumer of that data will hit the same problem and fix it differently.

Upstream work is slower and political, because it means asking the team that generates the data to change something for someone else's benefit. It is also the only permanent fix, and being able to make that case is a large part of what separates a senior data quality role from a junior one.

What AI changed

Training data problems do not surface as errors. They surface as a model that behaves oddly for a subgroup, months later, with no obvious cause.

That has made two things newly valuable. Representativeness, which is a quality question the traditional dimensions do not cover: is this data actually about the population the model will be used on. And label quality, since a supervised model inherits the judgment of whoever labelled the training set, including their inconsistencies.

Neither shows up on a standard data quality dashboard, which is exactly why people who understand both are being hired.

Where to go next

Frequently Asked Questions

What are data quality skills?

The ability to determine whether data is fit for a specific purpose, profile it to find what is actually wrong, define measurable rules and thresholds, and drive fixes upstream rather than patching them repeatedly in a pipeline.

What are the data quality dimensions?

Completeness, accuracy, consistency, timeliness, validity and uniqueness. Five are testable with a rule. Accuracy is the difficult one, because it means agreement with the real world and your systems do not contain the real world.

What is the difference between accuracy and validity?

Validity means the value conforms to expected format and rules, such as a date of birth being a real date. Accuracy means it is the customer's actual date of birth, which requires a source outside the system. Teams that claim to measure accuracy are usually measuring validity.

How good does data need to be?

Good enough for the specific use, decided by a named person. Addresses that are ninety-four percent correct are fine for a marketing campaign and unacceptable for regulatory reporting. Global quality scores produce dashboards; per use case thresholds produce decisions.

Why profile data before writing rules?

Because profiling frequently shows the real problem is not the one being discussed. A field populated entirely with an unnoticed default value, two systems using one field name for different things, or a daily file arriving with the same timestamp for weeks.

Should data quality be fixed upstream or in the pipeline?

Upstream wherever possible. Pipeline fixes work and guarantee you will apply them forever, while the next consumer hits the same problem and solves it differently. Upstream work is slower and political because it asks the producing team to change something for someone else's benefit.

How does AI change data quality work?

Training data problems do not appear as errors. They appear months later as a model behaving oddly for a subgroup. Two concerns become important that traditional dimensions miss: representativeness of the population the model will serve, and label quality, since a supervised model inherits the labeller's inconsistencies.

What tools are used for data quality?

Great Expectations, dbt tests, Soda, Monte Carlo and Collibra are common, alongside a great deal of SQL. Employers care more about whether you can define the right rule than which platform enforces it.

What jobs require data quality skills?

Data quality analyst, data steward, data governance analyst, data engineer, and AI governance roles assessing training data representativeness and labelling consistency.