How to define data quality thresholds: a guide with examples by dimension

Why "high quality" isn't an operational threshold

We already covered the 6 dimensions of data quality — completeness, accuracy, consistency, timeliness, validity and uniqueness. Knowing those dimensions exist is the first step. The second — the one almost nobody takes — is putting a number on them: what percentage of completeness is acceptable for this specific dataset? How old can a piece of data get before it stops being "timely"? Without that figure, "high quality" is an opinion, not a control you can audit.

The right threshold depends on the use, not the data itself

The most common mistake is defining a single threshold for the entire organization. An email field used for marketing can tolerate a 5% error rate without serious consequences. That same field used for electronic invoicing can tolerate practically none. The threshold doesn't describe the data — it describes how much it can fail before the use you're putting it to stops being reliable.

Starting thresholds by dimension, with real examples

Completeness

Percentage of required fields that are filled in. Common starting threshold: 95-98% for critical business fields (contact email, national ID, date of birth in an HR system), and 80-90% for secondary fields (secondary phone, internal notes).

Real example: a CRM with 50,000 contacts where the "industry sector" field is filled in for 60% of records. If that field feeds a lead scoring system, a 90% threshold forces you to prioritize enrichment before trusting the score.

Accuracy

Percentage of records that match the source of truth (cross-checked against an external system, or manually validated on a sample). Starting threshold: 97-99% for financial or regulatory data, 90-95% for operational data.

Real example: a shipping address dataset verified against a postal validation service. If 8% of addresses aren't deliverable, your accuracy threshold (say, 95%) tells you to act before launching a mass mailing campaign.

Consistency

Percentage of records without contradictions across systems (the same customer with the same status in the CRM and in the billing system). Starting threshold: 98-100% for key identifiers (customer ID, national ID), more flexible (90-95%) for descriptive attributes.

Timeliness (freshness)

Maximum acceptable age of data before it's considered stale. This isn't a percentage — it's a time window: 24-48 hours for operational data (stock, prices), 30-90 days for customer profile data, annual for structural data (a company's industry sector).

Validity

Percentage of records that meet the expected format or range (an email in a valid email format, a birth date that isn't in the future). Starting threshold: 99%+ — it's the cheapest dimension to control with automated validation at the point of entry, so the threshold should be the strictest of the six.

Uniqueness

Percentage of records without exact or near-exact duplicates. Starting threshold: under 1-2% duplicates on key identifiers. A customer dataset with 8% duplicates isn't just a quality problem — it artificially inflates any business metric calculated on top of it (number of customers, average revenue per customer).

How to document the threshold, not just the number

A threshold without justification is as arbitrary as having no threshold at all. Document, for each one: which dimension it measures, the exact value, which domain it applies to, who approved it (technical and business), and when it was last reviewed. This is exactly the structure followed by the Data Quality Rules Template — with starting thresholds already defined by dimension and worked examples, so you don't start from a blank sheet.

Reviewing the threshold isn't optional

A threshold defined once and never reviewed tends to go stale in either direction: too lax if the business has grown more demanding, or unattainable if it was set without real baseline data. A quarterly review is reasonable for most domains; monthly if the data feeds high-impact automated decisions (like an AI system).