What is data quality?
Data quality describes whether data is fit for the purpose it serves. The same dataset can be good enough for one use and poor for another. Customer addresses that are correct only to the city level work for a regional sales map but fail for delivery routing. For that reason, quality is measured against rules written for a specific use, usually along six dimensions.
- Accuracy means a value matches the real world, such as a phone number that reaches the right patient.
- Completeness means required fields are filled, such as an insurance member ID on every scheduled visit.
- Consistency means the same fact agrees across systems, such as a birth date that matches in scheduling and billing.
- Timeliness means data is current enough for its use, such as insurance eligibility checked shortly before the appointment.
- Validity means values follow the required format and rules, such as a diagnosis code that exists in the current code set.
- Uniqueness means each real-world entity appears once, such as one record per patient.
Why data quality matters
Poor data quality shows up as rework, returned shipments, denied claims, and decisions made on wrong numbers. AI systems spread these errors further and faster. A machine learning model trained on mislabeled or incomplete records learns those errors and repeats them at scale. A language model answering from a knowledge base quotes an outdated policy as confidently as a current one. An agent that writes to systems of record can copy a bad value into several systems in minutes.
Data quality in practice
A multi-specialty physician group loses time and revenue to claims denied for registration errors. Common causes are missing member IDs, misspelled names, and coverage that had already ended. The group adds quality rules at the front desk and in a nightly check. Member IDs must match the payer’s format, and names and birth dates must match the eligibility response. Coverage must be verified within a set number of days before the visit. A weekly report shows error rates by clinic, and the clinics with the most errors get targeted training. The cleaner registration data later feeds a no-show prediction model as well.
What to watch for
- Write quality rules with the people who use the data, since they know which errors matter.
- Measure quality continuously in pipelines, ideally with data observability, because a one-time assessment goes stale.
- Fix recurring problems in the source system, and use data cleansing for the backlog that already exists.
- Check training and retrieval data before blaming an AI model for bad answers.