← Straight Fact

Straight Fact · desk research and BespokeData methodology · 18 Nov 2026

Complete data can still be wrong

2 min read · Free article

A populated field is not necessarily a reliable field. If every order has a promised date, a completeness check can pass while the dates are copied from the wrong source, revised without history or attached to duplicate orders. The practical takeaway is to choose checks that match the decision. Counting non-empty values is a useful start, but it cannot tell you whether those values represent what happened or what was agreed.

The UK Government Data Quality Framework treats completeness and accuracy as distinct dimensions [Q1]. That distinction matters in ordinary business reporting. A manager choosing which customers to contact needs trustworthy commitments, not merely dates that satisfy a format rule. Validity can establish that a date parses; accuracy requires some comparison with an appropriate authorised record. Do not describe a successful format check as verification of the customer's original promise.

Consider a hypothetical extract of 1,000 distinct orders, with 80 missing promised dates. Completeness is 920 divided by 1,000, or 92%. Now imagine checking a documented sample of 40 populated dates against the original confirmations and finding four disagreements. That sample has 36 agreements out of 40, or 90%. It is a sample result under its stated selection rule, not proof that exactly 10% of all populated dates are wrong. The two checks answer different questions.

Look at where the missing and disputed records occur. A problem concentrated in one location or a recently changed system deserves a different response from a random scattering. If late orders are more likely to have missing promises, dropping those records can make performance look better. Preserve the unknown category and report its size. Correcting a denominator silently is not a data-quality improvement; it changes the question being answered.

A practical action is to add three rows to your evidence register: field population, permitted format and agreement with an authorised reference. Give each a denominator, extraction date, owner and limitation. Ask the owner to investigate disagreement mechanisms before replacing values. Preserve the original file and document corrections separately so a later reviewer can reproduce the result.

This is a hypothetical example and desk-research application, not a finding about your business. A sample may miss important failure patterns and does not establish every field's accuracy. Use the full AI readiness guide to connect these checks to a specific use, permission boundary and pilot gate rather than treating a completeness percentage as a universal readiness score.

Dated method & source register

Read and reviewed 18 Nov 2026. Desk research and original BespokeData recommendations and hypothetical examples; no primary study, independent expert review or Atora generation.

Read the full guide & worksheet →

Turn the insight into a next step.

Explore focused help →