A report is only as trustworthy as the data behind it, and the failures that damage trust are rarely dramatic. They are quiet: a figure that changed without explanation, a total that does not sum, a period that does not match its label. A short, repeatable set of data-quality checks catches these before the reader does.

This guide sets out the checks to run before every report cycle, how to make them fast enough to survive deadline pressure, and how to build them into the workflow. It is the companion to Data Quality for Reporting.

The checks are not there to be thorough. They are there to be run, every cycle, without fail.

The dimensions the checks cover

Data quality is not one property. Six dimensions matter for reports, and each can fail independently.

Dimension Question it answers Report failure if weak
Accuracy Is the figure correct? Wrong numbers
Completeness Is anything missing? Gaps and understated totals
Timeliness Is it current enough? Stale figures presented as current
Consistency Does it match other reports? Conflicting numbers
Validity Is it in the right format and range? Nonsensical values
Uniqueness Are records counted once? Double-counting

A good check set covers all six. A report can be accurate and still fail on consistency, if the same metric is defined differently in two places.

The checklist

Run this before every report cycle.

Check Question
Source Does every headline figure trace to a named source?
Definition Does each metric match the dictionary definition?
Prior period Have changes since last cycle been explained?
Completeness Are all expected entities and lines present?
Totals Do totals equal their parts?
Period alignment Does the reporting window match the label?
Units and currency Are units and currency stated and consistent?
Outliers Have surprising figures been investigated?
Estimates Are estimates and forecasts labeled?
Version Is the data snapshot recorded?

The list is short on purpose. A checklist that takes an hour will be skipped; one that takes ten minutes will be run.

Making the checks fast

The discipline of data-quality checks fails when the checks are slow. Keep them fast.

  • Automate the mechanical ones – totals summing, completeness against an expected list.
  • Reconcile against the prior period, where the comparison does the checking for you.
  • Flag outliers automatically, and investigate only the flagged.
  • Keep the checklist to ten items, runnable in minutes.

The aim is a habit, not a project. A fast checklist that runs every cycle beats a thorough one that runs occasionally.

Who runs the checks

The checks need an owner, and it should be the person accountable for the report rather than whoever has time.

  • The report owner runs the checks, because they own the output.
  • The verifier independently confirms the headline figures against source.
  • The approver relies on the completed checklist before signing off.

On a small team the owner may do all three, but the checklist makes the checks explicit rather than assumed.

When a check fails

A failed check is not a crisis; it is the system working.

  • Investigate the cause, not just the figure – a wrong number usually has a process cause.
  • Decide the fix – correct the figure, or qualify it if the source is confirmed correct.
  • Record the finding so the next cycle does not repeat it.
  • Feed it to the post-cycle review, where the process cause is addressed. See The Post-Cycle Review.

A check that never fails is either a very clean process or a check that is not being run.

Automating the checks

The mechanical checks are the ones to automate, because they are the ones that get skipped.

  • Totals and subtotals can be checked automatically against their parts.
  • Completeness can be checked against an expected list of entities or lines.
  • Prior-period comparison can flag changes beyond a threshold.
  • Outlier detection can flag figures outside a normal range.
  • Units and formats can be validated against the metric dictionary.

The human checks remain the ones requiring judgment: whether a cause is true, whether an outlier is real, whether the report is consistent with last cycle.

A worked check cycle

A ten-minute cycle shows how the checks fit together.

  • Minute 1-2: confirm every headline figure traces to a named source.
  • Minute 3-4: reconcile figures against the prior period and explain changes.
  • Minute 5-6: check totals and completeness against the expected list.
  • Minute 7-8: investigate flagged outliers.
  • Minute 9: confirm period labels, units and estimates.
  • Minute 10: record the snapshot version and sign the checklist.

The cycle is deliberately short. A check that takes minutes runs every month; one that takes an afternoon is deferred until it is too late.

When a check fails

A failing check is information, not a crisis.

  • Stop and assess before the report proceeds.
  • Trace the failure to its source, using the source log.
  • Decide: correct the data, correct the report, or disclose the limitation.
  • Record the failure and the fix, so the check improves.
  • Do not override silently – an overridden check is a future correction.

The point of the checks is to find problems before a reader does. A failing check that is investigated is the system working.

A minimal starting checklist

If you do nothing else, do this.

  • Confirm every headline figure has a named source.
  • Reconcile each headline figure to the prior period and explain the change.
  • Check totals against their parts.
  • Check completeness against an expected list.
  • Confirm units, periods and estimate labels.
  • Record the snapshot version and sign the checklist.

Six steps, minutes to run, and they catch most of what reaches a reader.

Common mistakes

  • No checklist. Checks are done from memory, and memory skips them.
  • A checklist that is too long. It is abandoned under deadline pressure.
  • Checks with no owner. They happen when someone remembers.
  • Fixing the figure, not the cause. The same error recurs next cycle.
  • Recording nothing. No way to see whether checks are improving quality.

Frequently asked questions

What data-quality checks should a report include?

Source traceability, definition alignment, prior-period reconciliation, completeness, totals, period alignment, units and currency, outlier review, estimate labeling, and version recording.

How often should the checks be run?

Before every report cycle, on the frozen data snapshot, so the checks are against a fixed target rather than a moving one.

Who should run data-quality checks?

The report owner, with the verifier independently confirming the headline figures. The checklist makes the checks explicit, even when one person holds several roles.

What do you do when a check fails?

Investigate the cause, correct or qualify the figure, record the finding, and feed the process cause to the post-cycle review so it does not recur.

Are data-quality checks the same as reconciliation?

Reconciliation between reports is one of the checks; the full set covers accuracy, completeness, timeliness, consistency, validity and uniqueness as well.

Which checks can be automated?

The mechanical ones – totals, completeness, prior-period comparison, outlier detection, and unit or format validation. The judgment checks remain human.

How long should the checks take?

About ten minutes on a frozen snapshot. The aim is a habit run every cycle, not a thorough audit run occasionally.

What should you do when a check fails?

Stop, trace the failure to its source using the source log, decide whether to correct the data, correct the report or disclose the limitation, and record the failure and fix. Never override a failing check silently.

Can you run checks on a report you inherited?

Yes, and you should. The checks do not depend on knowing the history; they depend on the report having named sources and a stated period. Where those are missing, establish them first.

Do checks slow the cycle down?

They add minutes and prevent hours. The aim is a short, repeatable cycle on a frozen snapshot, not an open-ended audit.

How do you check a figure you cannot verify against a source?

Rely on reconciliation instead: compare it to the prior period, to a related figure, and to any independent source. Where none exists, label the figure as an estimate and say so.

What is the difference between a data-quality check and a review?

A check is mechanical and repeatable – totals, completeness, thresholds. A review is judgmental – does the story make sense, does the figure match reality. You need both, in that order.

Should the checks be documented?

Yes. A short checklist, signed each cycle, is what turns the checks from a good intention into a repeatable control – and the record is what shows the report was checked. A one-page checklist is enough.

Do checks replace a second pair of eyes?

No, they reduce how much a reviewer has to catch. A mechanical check finds the broken total; a reviewer finds the sentence that contradicts the number.

Next step

Build the checklist into your reporting workflow, automate the mechanical checks, and run the rest every cycle. Download the Data-Quality Checklist to start, and see Reconciling Numbers Across Reports for the reconciliation step in detail.

Sources

  • Gartner, Magic Quadrant for Data Quality Solutions (July 2020), survey of 154 enterprise customers: the cost of poor data quality.
  • IBM Institute for Business Value (2025): data quality as a top priority for chief operating officers.
  • Harvard Business Review (Nagle, Redman and Sammon, 2017): 47% of newly created data records contain at least one critical error.

Numbers are cited from their sources and dated. Where a source is a vendor benchmark, the sample size is stated.