When metadata Carries Meaning
Three systems in the same bank each contain a field called customer_status. In the onboarding platform, it records whether identity verification is complete. In core banking, it shows whether an account is active. In the credit system, it holds a credit rating class. A report that combines the three columns looks consistent, and its numbers are wrong.
A data catalogue would list all three fields correctly, with name, data type, system and owner. It would still not tell you what customer_status means in each place, or that the three values cannot be compared.

From labels to connected context
Most organisations already collect three kinds of metadata:
Technical metadata describes structure and location: tables, columns, data types, formats and storage.
Operational metadata records what happens to data, such as load times, pipeline runs and usage statistics.
Business metadata covers definitions, classifications and responsibilities, often kept in a business glossary.
Semantic metadata connects the three. A glossary entry can define “verification status”. Semantic metadata adds which fields in which systems represent that concept, which similar-looking fields do not, and how the concept relates to others. It is a connecting layer over the metadata you already have, written in a form that machines can process as well as people.
The connections are typed relationships. For the onboarding field customer_status, five questions cover most of what a reader needs to know:
What does it mean? It represents the concept verification status, with the values pending, verified and rejected. Account status is a different concept.
Which source is authoritative? The onboarding platform. Copies in the data warehouse are derived from it.
How was it produced? The identity verification process sets it.
Who owns the meaning? A business owner in customer onboarding approves the definition.
Where is it used? In the onboarding report and in access decisions. A policy excludes it from marketing analytics.
Two more relationships deal with the classic errors. The credit system’s customer_status points to a different concept, so nobody merges it by accident. The CRM field client_state points to the same concept, so anyone searching for verification status will find it.

Each of these relationships can be written as a simple statement: customer_status represents verification status. The Resource Description Framework (RDF) stores data in exactly this subject, predicate, object form, and several open W3C standards build on it. SKOS describes vocabularies and taxonomies with preferred terms, synonyms, and broader and narrower concepts. DCAT and PROV-O cover catalogue entries and provenance, so that descriptions and lineage can move between tools. OWL adds formal definitions for cases where systems have to apply rules to the model.
Standards make these relationships exchangeable between tools. Which definition is right remains an organisational decision: someone has to agree which concepts matter, how they are defined and who may change them.
Model only what a use case needs
A frequent pitfall is to start by modelling the entire organisation. That often ends in a large ontology that is technically sound and rarely used, after months of debate about definitions that no concrete decision depended on.
Starting with one use case works better. For example:
helping analysts find the authoritative source for a regulatory report
separating customer, account holder, authorised user and counterparty
tracing how a management indicator is calculated
flagging which datasets contain personal or sensitive data
giving a search or AI assistant reliable context about internal data
The use case defines the minimum semantic information. A regulatory report needs the reported concept, its authoritative source, the calculation rule, lineage, an accountable owner and the applicable policy. Data discovery gets by with synonyms, related concepts, classifications and access conditions, and everything else can wait.
It also defines how formal the model has to be. Formal models need specialist skills, every relationship needs maintenance, and automated rules can draw surprising conclusions when a definition is too broad. So it pays to stop at the first level that solves the problem. A controlled vocabulary is enough when the main problem is that people use different words for the same thing. A taxonomy helps when users mainly need to move from general to specific concepts, for example from customer to retail customer. An ontology is worth the effort when systems must interpret relationships or derive facts automatically, for example that every account holder is also a customer.

A use case also makes success measurable. Before starting, record a baseline: how long analysts need to find the authoritative source for a figure, how often reports show conflicting numbers for the same concept, or what share of critical datasets has an owner and an agreed definition. Measure again after a few months. If nothing has moved, either the model has not reached the tools people work in, or it answers questions nobody asked. Organisations that report against the FAIR principles can use the same measures as evidence for findability, interoperability and reuse.
Keep meaning current
Meaning drifts. Terms get redefined and systems get replaced, and new regulation brings new classifications with it. A field can keep its meaning while its source changes, or keep its source while its calculation changes. Without upkeep, semantic metadata becomes one more outdated reference. Four practices keep it current.
Split ownership by responsibility
A business owner approves definitions, and a data owner is accountable for a dataset. An architect maintains the links to applications and interfaces, while a governance function sets classification and quality requirements. “Owned by the data team” is too broad to act on.
Separate editorial from semantic changes
Renaming a label is usually harmless. Changing what “active customer” means can affect dashboards, policies, models and interfaces. Versioning and impact analysis show which kind of change is on the table and who has to review it.
Validate the structure automatically
Machines can check rules such as “every critical dataset has an owner and a definition” or “every field is linked to exactly one concept”. For RDF-based models, the W3C standard SHACL expresses such rules. Validation confirms that the agreed structure is complete; whether a definition is correct stays with the owner.
Show context where people already work
Semantic metadata is most useful inside the tools people already use: catalogue search, BI dashboards, API documentation, data contracts and approval workflows. That is where someone selects, interprets, changes or approves data, and where the context makes a difference to the decision.
Metadata carries meaning when it is connected
Metadata carries meaning when it links a data item to the concept it represents, its authoritative source, the process that produced it, the owner of its definition and the places where it is used, and when someone keeps those links current.
A practical first step is to pick one data item that is regularly misunderstood in your organisation and try to answer the five questions. What does it mean? Which source is authoritative? How was it produced? Who owns the meaning? Where is it used? The questions you cannot answer show where to start. Which data item causes the most confusion in your organisation? I’d like to hear what the five questions turn up, and how far a model needs to go for your use case.
Further reading
RDF 1.1 Primer: https://www.w3.org/TR/rdf11-primer/
SKOS Primer: https://www.w3.org/TR/skos-primer/
PROV-O: https://www.w3.org/TR/prov-o/
OWL 2 Overview: https://www.w3.org/TR/owl2-overview/
SHACL: https://www.w3.org/TR/shacl/
FAIR Principles: https://www.go-fair.org/fair-principles/
Next posts

Building Enterprise-Grade Web Filtering with AdGuard Home and Wazuh SIEM
How DNS-based threat detection integrates into a security monitoring stack

AI Agents as Microservices - Why AgentOps Doesn't Need a New Platform World
AI agents are workloads, not magic. Why we don't need separate AgentOps silos, but should integrate agents into proven cloud-native platforms, Kubernetes, and domain ownership.

A Practical Framework for Agent-Driven Application Modernization
Learn how agent-driven development supports application modernization by turning complex migrations into a structured process of analysis, planning, execution, and documentation.