Watching the Data Quality

A system can run for years producing records that look complete and are not. These four checks are cheap and catch most of it.

Promptness

Proportion of entries created within one working day of the work.

The strongest single indicator, because accuracy falls sharply with delay.

Segment by team. A team with a long tail is telling you the tool does not fit their work.

Watch the trend rather than the level. A rise means friction increased somewhere — a new mandatory field, a category list that grew, an app update.

Completeness

People with no entries in a period who should have some.

Recorded hours against contracted hours, in aggregate, which reveals systematic under-recording.

Unapproved entries at close.

Missing clock-outs, which indicate a usability problem rather than carelessness.

Correction rate

Entries edited after creation, as a proportion.

By cause where a reason is captured.

A rising rate points upstream: a rule applied wrongly, a category people cannot interpret, an integration writing bad data.

A rate near zero is also a signal. It usually means corrections are impossible rather than unnecessary, and people are living with wrong records.

Reference data hygiene

Active projects nobody has recorded against in six months.

Entries against a catch-all or general code, as a proportion. Rising means the structure no longer fits the work.

People in the system who left.

Codes reused after retirement, which silently corrupts historical reporting and which should be impossible by configuration.

Reporting the four

On one page, quarterly, with the trend.

Alongside any analysis built on the data, so a reader knows what weight to apply.

With a sentence of interpretation each, written by the owner.

To the people recording the time as well as to management, which is the audience most often omitted and the one whose behaviour determines the numbers.

Acting on them

Promptness falling: reduce friction, not reminders.

Catch-all rising: prune the category list.

Corrections rising: look at the rule that changed.

Completeness gaps: check the integration first, before concluding people stopped recording.

In every case the first hypothesis should be that the system changed, not that people did. It is right more often, and being wrong about it in the other direction costs the cooperation the records depend on.

The zero-correction signal

A number that looks good and usually is not.

A correction rate near zero rarely means the records are right.

It usually means correction is impossible — the period locks too early, the workflow is painful, or people were told not to.

And so people live with wrong records, which surfaces in a dispute rather than in a report.

Check by asking a few people whether they have ever had an entry they could not fix.

Report it with the analysis

Numbers without their quality are numbers with unknown weight.

Put the four checks on the same page as any analysis built on the records.

State the recording method: prompt entry and weekly reconstruction are not the same data.

Name the known gaps.

A report that states its own limits is trusted; one that does not is tested once and then discounted entirely.

Investigate unusual signals

An anomaly should start a data-quality check rather than trigger an automatic judgement. See the additional details, then document false-positive review and escalation before using the signal operationally.