Standards
7 min

Taxonomy Governance: Decision Rights, Records and Cadence

Key takeaways

  • Governance is the reason a taxonomy still says something accurate a year later. It has no deliverable.
  • Slow approval is the root cause of most drift. People route around an expensive process.
  • Split structure from meaning: the steward owns process, the domain approver owns what a term means.
  • Four fields per change: who proposed, who approved by name, when, and what it replaced.
In this article
Share

What is taxonomy governance?

Taxonomy governance is the set of decisions and roles that keep a controlled vocabulary true. Who may add a term. Who approves a change in meaning. What happens to a term when it is retired, and how often anyone checks.

It is the part of the work that has no deliverable. A taxonomy is a visible artifact somebody can point at. Governance is the reason it still says something accurate a year later. Its absence shows up as drift, and never as a missing document.

Why do vocabularies drift?

Four causes, and none of them involve anybody behaving badly.

CauseWhat it looks likeWhat it produces
Slow approvalAdding a term takes a fortnightPeople use free text, and the vocabulary stops describing reality
Unassigned branchesThe owning team reorganised and nobody reassignedA branch nobody maintains, visibly current and quietly wrong
No duplicate checkTwo near-identical terms added months apartQueries split across both and return partial sets
Retirement without a pointerA term is deleted and nothing replaces itEverything tagged with it is stranded

Slow approval is the root of most of the others. When the official route is expensive, people route around it, and every workaround adds a term the vocabulary does not know about. A governance model that optimises for correctness and ignores latency produces a vocabulary that is correct and ignored.

How should the decision rights split?

Three roles, and conflating any two of them is the common failure.

The three roles

  • Steward. Runs the process, checks for duplicates, maintains the structure. Usually central, usually one or two people.
  • Domain approver. Decides what a term means in their area, because they are the only person who holds that. Distributed, one per branch.
  • Consumer. The systems and teams reading the vocabulary. They raise problems and do not get a vote on meaning.

The failure mode when steward and approver collapse into one is a central team authoring meaning it does not have. The failure mode when they separate too far is a federated model with no structural consistency, which produces several taxonomies wearing one name.

A workable split gives the steward authority over structure and process, and the domain approver authority over meaning. When those two disagree, the steward escalates instead of overruling.

What has to be recorded?

Four fields per change, and anything less is not a review trail.

What a change record holds

  • Who proposed it, so the context is recoverable.
  • Who approved it, by name, because "the team agreed" is not an approver.
  • When, because the question that matters later is what this meant in March.
  • What it replaced, so the previous meaning is still reachable.

The last one is the one that gets skipped and the one auditors ask for. A definition that changed with no record of what it changed from leaves every historical report unexplainable, because nobody can reconstruct which version produced which number.

Expressing this in an open standard helps, because the record then travels with the vocabulary. SKOS carries notes, editorial history and replacement pointers, and SHACL can require that a concept has a definition and an owner before it is valid.

How often should anyone look at it?

Quarterly for a review pass, and immediately when a consuming system reports a problem.

The quarterly pass asks three questions per branch. Is the owner still here and still in that role. Has anything been added without going through approval. Is anything in here now wrong because the business changed. Twenty minutes per branch, and it catches the slow failures before they compound.

The incident path matters more. When a system or a person reports that a term is wrong or missing, that has to reach the approver within days. A vocabulary where reporting a problem leads nowhere teaches people to stop reporting, and after that the only signal you get is the drift itself.

What should you measure?

Coverage is the usual metric and it measures the wrong thing. A vocabulary can be exhaustive and still be routed around by everyone who has to use it.

Four numbers worth tracking

  • Time to decide. How long a proposed term waits for an approver. This predicts whether people keep proposing.
  • Unapproved additions. Terms that appeared without going through the loop. A rising count means the loop is too slow.
  • Branches with a live owner. Anything below total is a branch quietly going stale.
  • Change records with all four fields. The gap here is what an auditor will find.

The first and the last are objective and automatable. The middle two need somebody to look, which is itself the point: a governance model nobody checks is a governance model nobody has.

Watch time to decide most closely. It is the earliest signal and the one that moves before anything visible breaks, because people stop proposing long before the vocabulary looks wrong.

How does this sit inside data governance?

A taxonomy is one of the things a data governance programme governs, and the review trail it produces is the same evidence an audit asks for. The overlap is real and the scope is different.

Data governance covers access, quality, lineage and policy across the estate. Taxonomy governance covers one narrow question inside it: what the terms mean and who said so. Folding the second entirely into the first tends to lose it, because meaning is slower and less measurable than the other concerns and gets deprioritised against them.

The practical arrangement is that taxonomy governance reports into the wider programme and keeps its own cadence and its own approvers. Data governance for AI covers the wider frame, and taxonomy and ontology management covers running a vocabulary and a model together.

So what should I take away?

Split structure from meaning, record four fields on every change, and make the approval loop fast enough that nobody routes around it. Latency is a governance property, and a slow process produces an inaccurate vocabulary just as reliably as no process at all.

TopQuadrant's position is that governance is what separates a vocabulary from a list, and that it has to be in the data instead of in a side document. When the approver, the date and the replaced version travel with the concept, the question of what a term meant in March is answerable from the vocabulary itself. When they live in a spreadsheet beside it, they survive exactly as long as the person who maintained it. Taxonomy management covers the surrounding discipline.

Give your AI the context it's been missing

See how the TQ Data Foundation turns your enterprise knowledge into trusted, Al-ready context.