The word ontology has two distinct lives. In philosophy, it refers to the study of what exists. In information science, it refers to a formal specification of what exists within a particular domain - the entities that matter, the relationships among them, and the rules that govern how those relationships can be interpreted by machines.

The information science definition is the one with practical stakes today, as organizations deploy AI systems that depend on structured semantic frameworks to produce consistent, auditable outputs.

That practical definition has been in development since the 1980s, refined through academic research at institutions including Stanford University and formalized through standards published by the World Wide Web Consortium. What emerged from that work is a discipline that sits at the intersection of logic, linguistics, and software engineering.

This field is now drawing renewed attention as the limits of unstructured AI systems become visible in regulated and high-stakes environments.

Understanding ontology engineering as a field requires separating it from several adjacent concepts it is frequently conflated with. A taxonomy organizes things into categories. A schema defines the structure of a database. A knowledge graph stores entities and their relationships.

An ontology does something more specific: it encodes the logic that allows a machine to derive new facts from existing ones, without being explicitly told each conclusion in advance.

Ontology Engineering at a Glance


  • Ontology engineering produces formal models that allow machines to reason about meaning, not just retrieve data - a distinction that separates it from schemas, taxonomies, and conventional databases.
  • The W3C's OWL and RDF standards form the underlying technical substrate; Protégé, TopBraid EDG, and AllegroGraph represent the main tooling layer used in production environments.
  • Life sciences is the most mature deployment vertical; financial services has the most ambitious shared standard (FIBO) and the most uneven adoption.
  • Ontologies support inference and interoperability but cannot self-populate, adapt automatically to domain changes, or substitute for data governance decisions made by people.
  • The rise of retrieval-augmented generation has made ontology-grounded knowledge graphs an active research and engineering frontier, with structured semantic scaffolding increasingly used to constrain and ground large language model outputs.

The Standards Underneath


The technical foundation of modern ontology engineering is a pair of W3C standards: the Resource Description Framework, known as RDF, and the Web Ontology Language, known as OWL. RDF provides a graph-based data model that represents information as subject-predicate-object triples.

A statement such as "Goldman Sachs is a financial institution" becomes a triple in which Goldman Sachs is the subject, a relationship type is the predicate, and financial institution is the object. Any fact expressible in that three-part structure can be stored and queried within an RDF system.

OWL, which the W3C standardized and formalized in OWL 2, extends RDF with formal logic. It allows ontology builders to define classes, specify properties of those classes, declare constraints on how they can relate, and express equivalences across systems.

The key capability OWL adds is inference: given a set of defined rules, an OWL-capable system can derive conclusions that are not explicitly stated. If a bond is defined as a financial instrument, and financial instruments are subject to certain regulations, the system can infer without being told separately that the bond is subject to those regulations.

Between RDF's flexibility and OWL's logical expressiveness sits SKOS - the Simple Knowledge Organization System, also a W3C standard - which provides a lighter-weight vocabulary for representing controlled vocabularies, thesauri, and taxonomies. SKOS is less expressive than OWL but far easier to implement.

Many organizations use it as a first step before committing to full ontology modeling. These three standards - RDF, OWL, and SKOS - form the backbone of nearly all production ontology work.

The standard query language for RDF data is SPARQL, also maintained by the W3C. It allows users and systems to query a triple store using pattern-matching logic, retrieving entities and relationships that satisfy specified conditions across what can be very large graphs.

The combination of OWL for modeling and SPARQL for querying defines the operational core of what practitioners call the semantic web stack.

More Technology Articles

IPFS and the Enterprise Compliance Gap in Decentralized Storage
artistic representation of massively distributed data storage
VRML and the Platform Problem of the First 3D Web
VRML-based render of a Japanese city
Physical AI’s Human Data Supply Chain Is Taking Shape
artistic representation of a chinese humanoid robot trained on human motion data
The Governance Dependency Problem in Digital Asset Custody
artistic representation of a stack of contracts held in place on a metallic pin

The Tooling Layer


The most widely used tool for ontology development is Protégé, a free, open-source editor originally developed at Stanford University and first released in 1999. Protégé has a large global user community and has been applied in projects ranging from academic research to government standards development.

It supports full OWL 2 editing, integrates with automated reasoners including HermiT and Pellet, and is available both as a desktop application and as a browser-based collaborative environment called WebProtégé. The tool has been used in the development of the National Cancer Institute Thesaurus and in the World Health Organization's construction of the International Classification of Diseases 11th revision.

At the enterprise end of the tooling spectrum, TopQuadrant's TopBraid Enterprise Data Governance platform, known as TopBraid EDG, is among the most established commercial products for ontology management at scale. It is built on RDF knowledge graphs and uses the W3C's SHACL standard - the Shapes Constraint Language - to define schemas and validate data against ontological rules.

TopBraid EDG ships with pre-built ontology models for data-governance asset types, including business terms, data sources, reference datasets, taxonomies, and enterprise governance entities. It also supports role-based access control, automated cataloguing, and integration with external data pipelines.

For organizations requiring a graph database optimized for RDF and ontology-driven reasoning, Franz Inc.'s AllegroGraph is a widely deployed commercial triple store. It stores RDF triples, supports SPARQL querying, and handles ontology-based inference at scale.

Ontotext, a Bulgarian knowledge graph company, offers GraphDB, another commercial triple store with strong OWL reasoning support, used extensively in media, publishing, and life sciences. Cambridge Semantics, founded by former IBM researchers in 2007, markets its platform under the Altair Graph Studio brand following a corporate acquisition, positioning it as a semantic data fabric layer that integrates structured and unstructured enterprise data sources.

These tools share a common characteristic: they are built for practitioners with training in knowledge representation and formal logic. Protégé's learning curve is navigable for a graduate student or experienced knowledge engineer, while TopBraid EDG is designed for data governance teams in large organizations.

None of these products are self-configuring. The ontology has to be designed, reviewed by domain experts, validated against real data, and maintained as the domain evolves. That labor cost is a central feature of the field.

Palantir Technologies represents a distinct approach to ontology in enterprise deployments - one examined in detail in a prior Beige Media analysis. Rather than providing a standards-based ontology editing environment, Palantir builds a proprietary ontology layer into its Foundry and Gotham platforms as an operational integration mechanism, with embedded engineers constructing and maintaining the model alongside the customer.

Where the Field Has Taken Root


Life sciences and biomedical research represent the most mature deployment environment for formal ontologies. The National Cancer Institute Thesaurus, maintained by the NCI and developed with Protégé, is a controlled vocabulary and ontology covering cancer-related diseases, drugs, genes, and anatomical structures.

The Gene Ontology, a collaborative project launched in 1998, provides a shared vocabulary for describing gene and protein functions across species. The Open Biological and Biomedical Ontology Foundry curates a suite of interoperable domain ontologies that now covers anatomy, phenotype, environment, and chemical entities.

These systems are used for data integration across research databases, for annotation of experimental results, and for enabling cross-institutional queries against heterogeneous datasets. FDA-facing clinical data submissions and adverse-event reporting workflows rely heavily on controlled terminologies, where consistent language is a precondition for machine-assisted review.

The breadth of ontology adoption in life sciences reflects both the technical complexity of the domain and the existence of institutional infrastructure - funding, consortia, standards bodies - that supported years of foundational work before commercial applications became viable.

In financial services, the equivalent effort is the Financial Industry Business Ontology, known as FIBO, developed under the auspices of the EDM Council and standardized by the Object Management Group. FIBO models financial instruments, legal entities, markets, and regulatory relationships using OWL, with the goal of providing a machine-readable common vocabulary across institutions.

As of the 2026 Q1 production release, FIBO spans thousands of classes covering foundations, business entities, securities, derivatives, financial instruments, loans, and market data. Contributors to its development include Goldman Sachs, State Street, Citigroup, Deutsche Bank, Wells Fargo, and the Commodity Futures Trading Commission, among others.

FIBO's adoption trajectory illustrates a tension that runs through enterprise ontology broadly. The ontology is technically sophisticated and represents genuine consensus among domain experts. Its implementation, however, requires specialized OWL tooling, trained ontologists who understand both the formalism and the financial domain, and organizational commitment to aligning internal data to a shared external standard.

Third-party FIBO implementation commentary has argued that many financial institutions remain on the sidelines, and that even early adopters have struggled to scale FIBO use beyond dedicated centers of excellence within their organizations.

Intelligence and defense applications of ontological systems exist but are substantially less documented in open sources. What is documented reflects a consistent underlying requirement: large institutions managing data across many separate systems, with different classification levels, different formats, and different provenance chains, need a way to reason across those systems without manually reconciling them.

Ontological frameworks offer a mechanism for that, though the specific architectures deployed in classified environments are not subject to public review.

What Ontologies Can and Cannot Do


An ontology that is well-designed and properly populated can do several things conventional data management systems cannot. It can enable semantic interoperability: two organizations using the same ontology to describe their data can exchange it and have it land correctly in the other's system, because the shared definitions eliminate ambiguity about what the terms mean.

It can support auditable reasoning chains: when a system reaches a conclusion based on ontological inference, the chain of rules applied to reach that conclusion can in principle be reconstructed and examined. It can also enable cross-domain queries that would otherwise require manual data wrangling, by providing a unified conceptual layer above disparate databases.

Those capabilities come with structural constraints that have produced consistent failure modes in enterprise deployments. Ontologies do not self-populate. The data has to be mapped to the ontology's object definitions, a process that requires domain expertise and engineering work, and that produces technical debt when the underlying systems change.

Large ontologies are also brittle in practice: adding new classes or relationships can break existing inference rules, and ontologies covering complex domains tend to accumulate inconsistencies over time as domain knowledge evolves and the original design assumptions no longer hold.

A more fundamental constraint is that ontologies formalize consensus, but cannot produce it. When two teams within the same organization define a financial instrument differently, or when two regulatory agencies use the same term to mean different things, an ontology cannot resolve that disagreement - it can only make the disagreement explicit in machine-readable form.

The history of large ontology projects, including FIBO, reflects significant effort spent on the human problem of achieving terminological agreement before any technical modeling can begin.

There is also a well-documented tension between the two dominant assumptions that govern reasoning over knowledge systems. The closed world assumption, which underlies most traditional relational databases, treats any fact not recorded as false. The open world assumption, which underlies OWL-based ontologies, treats any fact not recorded as unknown.

That difference has practical consequences: ontology-based systems designed under the open world assumption can produce unexpected inference behavior when users expect database-style completeness. Practitioners who move from SQL environments to ontology systems often encounter this mismatch, and it contributes to the pattern of over-engineered specifications and deployment failures that have characterized parts of the semantic web's enterprise history.

Ontology in the AI Era


The emergence of large language models has created a new context for ontology engineering and, with it, a new wave of interest from organizations that had not previously engaged with the field. The core problem is familiar: large language models generate fluent outputs from unstructured text, but without a structured semantic layer, those outputs are difficult to constrain to a specific domain, to verify against authoritative data, or to audit after the fact.

Retrieval-augmented generation - the practice of supplementing model outputs by retrieving relevant information from external sources - addresses part of this problem. However, it only works if the retrieval system can locate semantically relevant content rather than lexically similar text.

Knowledge graphs built on ontological foundations are being used as retrieval substrates in this architecture. A 2024 preprint later published at EMNLP 2025 demonstrated ontology-grounded retrieval-augmented generation approaches, in which a model's outputs are constrained by hypergraph representations of domain knowledge derived from formal ontologies.

The practical advantage is that an ontology-grounded retrieval system can pull results based on semantic relationships. It can infer, for example, that a query about a particular drug class should retrieve information about all members of that class, not just documents that happen to mention the specific string used in the query.

The limitation that carries over from classical ontology engineering is the same: the quality of the outputs depends on the quality of the ontology. An ontology built hastily, without domain expert review, or without a clear scope will produce retrieval results that reflect its gaps and inconsistencies.

The current wave of interest in knowledge graphs as LLM grounding infrastructure is generating both genuinely rigorous implementations and a larger number of informal graph structures that use ontological terminology without the formal substrate those terms imply. The distinction matters for any application where the auditability of the reasoning chain is a requirement.

For organizations in regulated industries - financial services, healthcare, legal, government contracting - the question is not whether AI systems should have structured semantic scaffolding, but how much of that scaffolding needs to meet the formal standard ontology engineering has historically required.

The answer may be domain-dependent. A customer service chatbot grounded in a product catalog knowledge graph has different auditability requirements than an AI system used in credit risk assessment or clinical decision support. Ontology engineering has developed the technical vocabulary and tooling to address the harder case.

How broadly those tools get applied will depend on how clearly the field articulates what it can deliver and for whom.

What decades of ontology engineering have produced is not a solved problem but a set of tested methods for making the problem tractable. The standards exist. The tooling exists. The documented failure modes exist in the literature.

The open question is whether the current moment of AI infrastructure investment will produce deployments that draw on that body of work, or whether the same patterns of over-promising, under-governing, and deferred maintenance that have characterized prior cycles of enterprise knowledge management will repeat under a new set of product names.

Sources


Article Credits

Portrait of Michael LeSane
Michael LeSane
Editor