Data quality metrics change behaviour when they are owned, tied to a pipeline gate, and painful for the person who can fix the emit. Volume, freshness, uniqueness, and a few field checks on the datasets that feed executive numbers beat a catalogue of unread scores.
Most quality scores do not change anything
A dashboard of red, amber, and green that nobody is paged for is interior decoration. People learn to ignore it, then argue about the business metric the broken data produced.
Quality work starts from a different question: whose week gets worse when this check fails? If the answer is "the data team, eventually," the metric will not change producer behaviour.
Pick metrics that map to an action
Use a short set. Each one must imply a next step.
Freshness
Did the dataset arrive by the time the job needs it? Freshness is a clock, an owner, and a page. It is not an average lag that hides a missed Monday.
Completeness of keys
Are the fields you join on present? A null business key is not a "soft" issue. It is a broken row. Quarantine it.
Uniqueness of the grain
Does the grain you published still hold? Two current rows for one customer in a type-2 dimension, or duplicate invoice ids in a fact, will multiply every downstream metric.
Validity of the fields that define money or identity
Status codes should be in the allowed set, and currencies should exist. Emails should at least be shaped like emails if you claim they are emails. Do not validate every string. Validate the fields that, if wrong, publish a wrong executive number.
Stability of volume
A drop or spike relative to the recent pattern is a prompt to look, not an automatic fail. Combine it with freshness so a silent source does not look "healthy" at zero rows.
That is enough. Accuracy against the real world is the goal, but you cannot score it without a second source. Use reconciliations on the few datasets where a second source exists, and do not pretend a regex is accuracy.
Field notes on unifying data, governance and getting analytics teams to one source of truth.
Every Monday, five minutes.
Put the metric on a gate, not only on a wall
A quality metric that cannot fail a pipeline is a suggestion. Suggestions lose to ship dates.
Fail ingest on contract breaks: missing keys, type mismatches, duplicate grains. Fail the silver build when uniqueness of current rows breaks. Do not fail gold to hide silver. Microsoft Fabric's medallion best-practices series on 3 September 2026 said gold must not compensate for upstream quality, and a heroic coalesce in a mart is the opposite of a metric that changes behaviour.
When the gate fails, the alert goes to the owner on the contract, not to a shared inbox. Shared inboxes are how quality becomes nobody's problem.
Make the score visible where work already happens
Do not add a sixth dashboard. Put the last run's freshness and fail count on the same page as the executive metric, or in the pull request that changes the model.
Citi Global Wealth's move to one cloud platform, with silver as the unified data model and reporting compressed from weeks to days, was reported by Computer Weekly on 8 September 2026. Faster reporting makes silent quality issues louder. If numbers land in days, a stale or duplicated key is an operational incident, not a quarterly cleanup.
A thirty-day way to install this
Week one: pick the five datasets behind the metrics leadership already argues about. Week two: add freshness, key completeness, and grain uniqueness tests with owners. Week three: fail the pipeline on those three. Week four: review the failures with producers and delete any test that did not earn an action.
If a test never fails and never could fail, it is not a metric. If a test fails constantly and everyone ignores it, it is also not a metric. Tune or delete.
What not to measure yet
Do not score "trust" on a one-to-five scale. Trust is the absence of incidents, not a survey.
Do not roll one company-wide quality index. It will average a broken billing feed with a tidy reference list and look fine.
Do not let ML anomaly detectors be the first line. They are extras after the gates exist. A model that flags volume without an owner is another ignored wall.
LensAIght is a unified data intelligence platform, so quality on the modelled dataset can show up in chat, notebooks, and dashboards together. That only helps if the gate already stopped bad rows. Otherwise you have three surfaces confidently repeating the same error.
Frequently asked questions
Which data quality metrics actually change behaviour?
Freshness, key completeness, grain uniqueness, and validity on money or identity fields — each with an owner and a pipeline gate. Scores without a gate rarely change anything.
Should we track accuracy as a percentage?
Only where a second source exists to reconcile against. Otherwise you will invent a percentage from rules that are not accuracy.
Who should be paged when a check fails?
The producer owner for emit and schema. The model owner if the grain broke in transformation. Never only a generic data inbox.
How many checks per table?
Enough to protect the grain and the fields that feed published metrics. A long suite that everyone skips is worse than three checks that fail the build.
Where do quality metrics live in medallion architecture?
Measure bronze against the contract. Measure silver against the unified grain. Do not use gold to hide either.
How do we stop people ignoring the red tiles?
Make red stop the pipeline and the publish. A tile that cannot block is optional, and optional quality loses.
Book a demo to see a managed data platform running on your own sources
LensAIght
