Industry
Part of The honest guide to data governance as of 2027
Which data governance metrics are actually worth a target?
Data governance measured by state, not activity: governed coverage, detection lag, disputed concepts, and which of them can safely carry a target.
Governance functions are measured by what they produce, because output is easy to count. Policies written, entries cataloged, issues raised. None of it says whether anything got better.
This page proposes a different set: measures of whether the estate is in a better state, plus an honest account of which ones can carry a target and which cannot.
What to take away
- Measure state, not activity. A catalog entry count rises whether or not anyone trusts the catalog.
- Three measures do most of the workcoverage of the governed list, time to detect a failure, and the count of numbers still calculated two ways.
- Some of the best measures here are subjective. Record them anyway, consistently, by the same person.
The measures worth keeping
| Measure | How to compute it | What a bad reading means |
|---|---|---|
| Governed coverage | Share of the governed list with an owner, a written definition, and a date | The list is aspirational |
| Definition age | Median time since each governed definition was reviewed | Nobody is looking |
| Detection lag | Median time from a data failure occurring to someone noticing | Readers are your monitoring |
| Reader-reported failures | Share of failures first reported by a reader rather than a check | Same problem, sharper |
| Access review completion | Share of roles reviewed on schedule | Permissions only ever accumulate |
| Disputed concepts | Count of concepts calculated two ways in two places | The core problem, unresolved |
| Change notices issued | Share of definition changes announced before they landed | Silent changes are the expensive kind |
| Exception age | Median age of open exceptions to governance rules | Exceptions have become the rule |
Eight is more than most functions need. Pick the first three and the sixth if you are starting.
Eight Governance Measures
- Governed coverageowner, definition, date
- Definition agemedian time since review
- Detection lagfailure to noticing
- Reader-reported failuresshare found by readers
- Access review completionroles reviewed on schedule
- Disputed conceptscalculated two ways
- Change notices issuedannounced before landing
- Exception agemedian age of open exceptions
Why detection lag matters more than error rate
Error rate is nearly impossible to measure honestly, because you only count the errors you found. Detection lag is measurable exactly, and it is the number that determines the damage: a wrong figure caught in an hour costs a conversation, and the same figure caught in a quarter costs every decision made in between.
The related measure is who found it. If most failures are first reported by a reader, your quality checks are decorative, whatever their pass rate says. Formal frameworks for data quality list many dimensions worth checking, and it is still true that the practical question is whether anything is watching at all.
Where a data feed has an agreed delivery commitment, treat it like any other service-level agreement: state the promise, measure against it, and report the breaches rather than the average.
The measure most functions are missing
The count of concepts still calculated two ways in two places. It is subjective, it needs a person to maintain a list, and it is the closest thing to a real outcome measure that governance has.
Build it by walking the reporting estate once and recording every place a concept is computed. Where two computations disagree, that is one dispute. Track the number quarterly. It moves slowly and it cannot be gamed by producing documents.
What not to measure
Policies published, catalog entries created, meetings held, issues logged, training completions. Every one of these can be increased in a week without anything improving.
The temptation here is structural rather than dishonest. These are the numbers that are already available, and reaching for them is the streetlight effect applied to a function's own reporting. Resist it by choosing the measures before anyone asks for a report.
Which of these can carry a target
Targets belong on measures inside somebody's control that cannot be satisfied without doing the work: access review completion, change notice coverage, exception age. Set them, review them, and expect them to be met.
Which Measures Can Carry a Target
Can carry a target
- Access review completion
- Yes
- Change notice coverage
- Yes
- Exception age
- Yes
- Disputed concepts
- Detection lag
Track only, no target
- Access review completion
- Change notice coverage
- Exception age
- Disputed concepts
- Yes
- Detection lag
- Yes
Targets do not belong on disputed concepts or on detection lag. Both depend heavily on what you happen to discover, and a target on either produces underreporting. Track them as trends, discuss them, and never attach them to anyone's assessment.
Reporting this to a sponsor
One page, quarterly, with four things: the governed list and its coverage, detection lag with the reader-reported share, the disputed count, and one paragraph naming what was settled this quarter and what is stuck.
The last paragraph is the one that gets read. Naming what is stuck, with the specific concept and the specific person who has not decided, is the only reporting mechanism that reliably unsticks anything.
Related reading on this site
The measured substance is in data governance, and the scope and tiers behind the governed list are in bounding governance work.
The records these measures count are in twelve governance records, and the failure modes they detect are in nine governance mistakes. The general question of designing a metric that resists gaming is in metric shapes and their traps.
Common questions
Our sponsor wants a single governance score.
Offer coverage of the governed list instead, which is one number, honest, and directly actionable. A composite score built from eight measures hides exactly the movements a sponsor should see.
How do we measure detection lag if we do not know when something broke?
Use the timestamp of the earliest affected record rather than the moment of discovery. It is approximate and it is consistently approximate, which is enough for a trend.
The disputed count went up. Have we got worse?
Probably not. It usually rises for the first year because you are finding disputes that always existed. Say that when you publish it, before someone else notices the direction.
Should quality test pass rates be reported?
Only alongside the count of tests and what they cover. A ninety-nine percent pass rate across tests that check nothing meaningful is a number that actively misleads whoever reads it.
