The recent data governance blog series lays all of this out clearly. It answers the questions that have defined the field for a decade: Who owns this data? Who can access it? Is it accurate? Where did it come from? These are the right questions. But they are no longer the only ones.
There is a layer underneath all of them that classic data governance never had to govern explicitly, because for a long time, we could get away with leaving it implicit. That layer is “meaning”. And in the age of AI, leaving it implicit has become the most expensive shortcut in the enterprise.
The question data governance never asked
Data governance governs the data plane: the rows, the tables, the pipelines, the access events. It is very good at this.
What it doesn’t govern is the semantic plane: what a term actually means, whose definition wins when two domains disagree, and how that meaning is supposed to travel across system boundaries.
For years, this was tolerable. A human analyst would quietly reconcile the fact that “active customer” meant one thing in finance and another in sales. The interpretation happened in someone’s head, silently, and the organisation absorbed the cost as reconciliation work, slightly-off dashboards, and the occasional heated meeting.
AI removes that tolerance. When an AI system or an agent crosses the same boundary, it does exactly what we do under uncertainty: it meets a familiar term, collapses the ambiguity without noticing, and acts as though meaning were settled. Except it does so at machine speed, across thousands of decisions, with full confidence. Automation doesn’t reduce the ambiguity. It scales it.
So the governance question evolves. “Who owns this data?” becomes: “Who owns what this data means?” And that is a different question, with a different answer, requiring a different kind of governance.
Three planes, not one
The shift becomes clearer once we stop treating the data platform as the whole architecture. Modern AI-ready architecture is distributed across three interdependent elements:
- Data: Facts from operational systems, governed today by data governance.
- Content: Documented intent and constraints living in systems like SharePoint, M-Files, or Confluence: agreements, policies, design decisions.
- Knowledge structures: Explicit models of the organisation and its domain: ontologies, business metadata, process and enterprise-architecture models.
Data governance covers the first plane thoroughly. The other two are mostly ungoverned, or governed by accident. Yet these are precisely the planes AI depends on to interpret anything correctly. An agent grounded only in the data plane will over-generalise or hallucinate, because the meaning it needs lives in content and knowledge structures it was never given.
Knowledge governance is the extension of governance across all three planes and especially across the semantic and knowledge layers that data governance leaves untouched.
What knowledge governance actually adds
This isn’t a new department or a parallel bureaucracy. It is the same governance instincts, applied one layer down.
- Ownership of definitions, not just data. A data owner is accountable for a customer-data domain. A meaning owner is accountable for what “customer” *means* — the definition, the synonyms, the relationships, the exceptions. Often these are the same person wearing a second hat. The point is that the definition becomes a governed asset with an owner, not folklore that everyone assumes they share.
- Stewardship of ontologies and the semantic layer. Data stewards already translate business rules into technical implementation. Semantic stewardship extends that to maintaining the vocabulary ontologies and business glossaries that say what concepts mean, and keeping them aligned as the business changes.
- Governance of context-assembly rules. In a knowledge ecosystem, context isn’t stored centrally. It’s assembled at query time from data, content, and knowledge structures. Someone has to own the rules that decide which signals get assembled, which constraints apply, and which definitions win. Today, in most organisations, no one owns this. It is the most consequential ungoverned logic in the AI stack.
- A standard for the semantic layer, but only one slice of it. The Open Semantic Interchange specification went live in early 2026, giving the industry a vendor-neutral way to define metrics, dimensions, and relationships so business logic moves consistently across BI and AI tools. This is a real step forward, and knowledge governance should adopt it. But it deliberately governs analytics semantics, the meaning of structured data inside the platform. Contractual terms, process rules, design rationale, and organisational applicability live outside it. The semantic layer is necessary. It isn’t the whole meaning.
Semantic sovereignty is a governance principle, not a loophole
The instinct of traditional governance is to standardise: one global model, one universal vocabulary, adopted by everyone. We have watched enterprise-wide conceptual modeling exercises fail this way for years. Forcing a universal vocabulary in an open organisation produces the illusion of alignment while hiding disagreement underneath — alignment on paper, disagreement in reality.
Knowledge governance has to be honest about this. The principle is semantic sovereignty: each domain retains the right to define its own terms locally, and that local meaning stays explicit and governable until alignment is established deliberately, rather than assumed by default. This is the same logic that took data governance from rigid centralisation to federated, data-mesh-style models — the people closest to the work are best placed to govern it, within shared guardrails. Meaning is no different.
Sovereignty is not permission for every domain to invent its own reality. It is a refusal to pretend agreement exists where it does not. False shared meaning is worse than acknowledged local meaning, because false shared meaning fails silently.
Where to enforce, and where to relax
Knowledge governance does not require formalising every definition in the organisation. That cost rarely pays off, and an information architect who declines to build an enterprise-wide conceptual model is usually making a rational trade-off, not rejecting semantics.
The discipline is knowing where ambiguity becomes expensive — revenue decisions, compliance boundaries, automated actions — and enforcing determinism there first. Where meaning drives consequence, it must be explicit enough to act on. Everywhere else, governed-but-loose is fine.
The new governance question
- Data governance asked: Can we trust this data?
- Knowledge governance asks: Can we trust what the organisation, and now its AI, understands?
The organisations that win the next wave won’t be the ones with the cleanest tables. They will be the ones that treat meaning as a first-class, owned, federated asset, and govern it before their AI systems start acting on assumptions no one ever wrote down.