Webinars/On-Demand

Data Is Not Knowledge: Ontologies, Context, and the Infrastructure of Meaning

September 30, 2026

  • Jessica Talisman

    Semantic Engineer and Author, The Ontology Pipeline

  • Matthew Mullins

    CTO, Coginiti

What you'll learn

  • The database holds what finance, HR, and accounts care about. Most of what a business actually does lives outside it, so a knowledge program has to capture more than the database.
  • Knowledge does not emerge from data on its own. The DIKW pyramid suggests it does, and there is no research behind the pyramid.
  • You cannot shop for an ontology. Buying the tool before building the practice costs you sovereignty, interoperability, and a system designed for your own requirements.
  • The ontology pipeline runs controlled vocabulary, metadata schema, taxonomy, thesaurus, ontology. Each stage produces an artifact you can deploy and test with AI before starting the next.
  • A controlled vocabulary with definitions is useful on day one: loaded into an agent as a skills file, it improves accuracy and cuts prompting and token spend.
  • Teams do not have to agree on one definition. Model the disagreement so every definition has a place and a context.
  • A sustainable ontology program is about seven roles. Most companies fund one person and expect all seven.

Organizations have spent two decades getting better at managing data. Data management has budgets, job titles, and an established toolchain. Knowledge management, as a discipline, mostly has none of those inside the enterprise, and data teams are now being asked to deliver it because AI systems need business context that the warehouse does not hold.

Coginiti CTO Matthew Mullins spoke with Jessica Talisman, a semantic engineer, information architect, and librarian, and the author of The Ontology Pipeline. The conversation started from her essay arguing that data management is not knowledge management, and worked through what the second discipline involves and how a data team can start on it.

What the database leaves out

Talisman came to the distinction through client work. She started where much of the industry did, with semantic layers and the question of what knowledge could be found in the warehouse. What she found is that the database mostly records what concerns finance, HR, and accounts. What the business does day to day, and why, is somewhere else: documents in SharePoint and cloud storage, Confluence pages, Jira tickets, Slack threads, people's heads.

Her conclusion was that she could not build knowledge infrastructure from the database alone. Capturing that context requires a program whose scope is wider than the data.

Mullins asked whether this is the old split between data management and document management under a new name. Talisman's answer was that knowledge management is much broader than documents. It covers provenance, lineage, attribution, and ownership, and it tracks ideas over time. In library science, one of the core commitments is that predecessors are not deleted: things stay linked and accounted for so that an object can be understood in context.

The two fields share vocabulary, which adds to the confusion. Both talk about lineage and both talk about semantics, and they mean related but different things. Talisman's reading of the history is that the fields forked in the mid-1990s. The same explosion of information that produced big data, with its focus on collection, storage, and query, also produced the semantic web. One stream followed data and the other followed knowledge, and what the industry is living through now is the two converging.

What data people get wrong first

Asked what people from a data background misunderstand most, Talisman said she had a long list, and gave two items from it.

The first is labels. Semantics is a discipline with roots in linguistics that long predates business intelligence, and a name like table1_c_customer4 is not a label a person or a machine can read. Mullins recognized the pattern from customer data models: sophisticated structures with cryptic column names, as if brevity were the goal, in an era when autocomplete removes the cost of typing the full name. He made a related point about metrics. No business has "revenue"; it has annual recurring revenue, monthly recurring revenue, and a dozen other specific things, and the model should say which.

The second is audience. Teams build for each other before they build for users. Designing with taxonomies, ontologies, and metadata means designing for readers outside your own context, human and machine, which requires setting aside what you already know about your company. Talisman sees teams start by listing attributes when they have not yet described the thing the attributes belong to. Her rule is the one people have given her for thirty years: explain it like I'm five. Machines need the same treatment.

That is also her answer to people who ask whether definitions are worth writing when a model can generate plausible ones. How an organization curates and presents its own knowledge is specific to it. Mullins compared organizations to baroque cathedrals, each built differently, which is why industry-standard data models never fit as well as promised.

You cannot shop for an ontology

Mullins had just returned from Big Data London, where several visitors to the Coginiti booth said they were shopping for an ontology. Talisman hears the same first question everywhere: what tool can I buy? Her counter-question is what happened to practice.

She listed what goes wrong when the tool comes first:

  • Sovereignty. Your knowledge can end up locked inside the product, with a dependence on that vendor.
  • Interoperability. Knowledge expands and contracts, and supporting exchange between systems is one of the core values of the field. A packaged solution may not.
  • Requirements. A system should be designed from functional requirements and competency questions, the questions it must be able to answer. A purchased solution arrives with someone else's answer to what it is for.

There is also a logical difference that tools do not remove. Relational databases work under a closed-world assumption: if a fact is not present, it is false. Most of the W3C stack around RDF works under an open-world assumption: if a fact is not present, it is unknown. Managing knowledge under that assumption means planning for change, such as a product launched next week or a new line of business, in a way that a schema migration does not cover.

What knowledge infrastructure consists of

Mullins asked what the knowledge equivalent of data infrastructure is. Talisman's list:

  • Defined language. Even a flat glossary of terms with meanings attached. It can serve as tags and navigation, and it states how the company describes itself.
  • Organized assets. An audit of what exists, by format and type: PDFs, documents, spreadsheets, and also databases, endpoints, and dashboards. People are already pasting these into context windows. You cannot build from material you have not mapped.
  • Documentation of the system itself. The audit trail, a management strategy, and the functional requirements.
  • A practice. Whether the function is centralized or distributed, the habits have to be built across the organization. Her example: when you post to Confluence, use a controlled tagging system.

Is a warehouse a knowledge management tool? Talisman argues no, not on its own. With a semantic layer and a data catalog around it, it can become part of the knowledge ecosystem.

She also noted that corporate knowledge management split from library science years ago and settled into managing repositories like SharePoint, without the principles the older discipline had developed over centuries.

Knowledge does not bake out of data

The DIKW pyramid, with data at the base and information, knowledge, and wisdom stacked above it, is widely used in data management. Mullins, who trained as an epistemologist, had not met it in philosophy and thought it was false on first reading. Talisman agrees. It looks like Maslow's hierarchy and implies emergence: collect data, build dashboards, and knowledge follows. No research supports that, and she thinks it has given the industry false hope.

Her example of knowledge management done without noticing is how people use AI assistants. Attaching artifacts, adding context, keeping a thread of work in one session, building memory: that is a small knowledge ecosystem, assembled by connecting things and preserving their lineage. The difference in an enterprise is that nobody has made it explicit.

Both spent time on sourcing. Talisman's training in history taught her to balance primary, secondary, and tertiary sources before taking a position, and she sees knowledge the same way: several perspectives that roll up to a common language. Mullins added the view from the epistemology of testimony. The usual accounts of why we can learn from what someone tells us rely on the speaker intending to inform and being reliable, and neither applies cleanly to a model. Whether grounding a model in curated knowledge makes it reliable enough to learn from is an open question. Talisman's practical version: asking a model to cite sources only helps if the sources are stable and repeatable, which means a constrained environment that you define.

The ontology pipeline

Talisman began applying the framework in 2018 at the Department of Justice, rebuilding grant review and reporting across eighteen program offices. It has five stages.

  1. Controlled vocabulary. Collect the terms used across the environment, clean them, deduplicate them, and define them.
  2. Metadata schemas. Decide how data and the things in your world are described.
  3. Taxonomy. Arrange the vocabulary into a hierarchy, informed by the schemas.
  4. Thesaurus. Add relationships between terms, along with alternative labels and aliases.
  5. Ontology. Design the formal model, informed by everything before it.

By the time a team reaches the last stage it has definitions, clean terms, and an audit, so it can state competency questions and functional requirements plainly.

Each stage also produces something usable, and this is what makes the approach practical:

  • A controlled vocabulary with definitions can be loaded into a company's AI agents as a skills file, even as plain text. Talisman has done this at several companies. It improves accuracy immediately and reduces prompting and token spend, because the agent is no longer reaching for a generic definition on each request.
  • Metadata schemas can be expressed as RDF, which gives you an RDF projection of the catalog that can be aligned with a semantic layer.
  • A taxonomy encoded in SKOS, the W3C standard for knowledge organization systems, is a concept graph. It supports basic inference, can be queried with SPARQL, and can back a RAG implementation.

Every concept carries a persistent identifier, and that identifier holds the label, definition, aliases, scope notes, examples, and links to external resources together. Feedback from one stage, such as a relationship that needs adjusting or a category that needs widening, shapes the next. Talisman added that several EU and Nordic countries require SKOS for government reporting.

Mullins compared it to agile software development: build incrementally on a vocabulary, as opposed to launching a waterfall project called "build the ontology."

Talisman has published a companion tool, Ontology Pipeline Studio, with two worked examples, one about music events and venues and one from the book. It lets readers add and remove things and find where the model breaks.

Seven roles, one hire

Asked to make the case for funding, Talisman started with what the work involves. Ontologies involve code and validation, and they are not easily improvised. Done properly, a program needs knowledge managers, an analyst, a product manager, a program manager, a taxonomist, an ontologist, and a knowledge graph engineer.

Most companies post one position, pay below what the skill set commands, and expect that person to do all seven jobs. At Adobe, her last in-house role, she built the vocabulary and the pipeline, designed the storage schemas, worked out the implementation, and acted as her own product manager. Her comparison for data teams: imagine one person responsible for the warehouse, every schema, every pipeline, and product management for all of it.

Mullins sees the same collapse in data, where the architect and modeler roles were folded into data engineer and organizations then wonder why nobody models.

The talent exists. There are PhD programs across Europe and Asia and a research community with decades of conferences behind it. Through Knowledge Graph Academy, Talisman and her colleagues have certified more than a hundred ontologists in the EU, Australia, and the US. The course is six weekly sessions of seven hours each, live and unrecorded, running from conceptual modeling and set theory through controlled vocabularies, SKOS taxonomies, the T-Box (classes), the A-Box (assertions), and SPARQL.

Two things she tells students: your first ontology will be bad and your second will be better, and the best ontologists are not always data engineers or architects. Content strategists and technical writers often do well. What matters is comfort with abstraction and systems thinking.

Amazon is the company she thinks handled this well. It hired her as a program manager, and she has come to think a sustained program needs someone coordinating all the moving parts.

Disagreement is something you model

Mullins asked whether an organization that cannot agree on definitions is not ready for an ontology. Talisman said no. Disambiguation is part of the method, and RDF gives you ways to structure it.

Her example is Adobe, where twenty-one products each had a feature called Journeys, and it was a different feature in every one. Nobody was going to agree on a single definition. The job was to model the term so each definition had its place and its context. Governance still matters, but the model can carry the nuance.

Semantic layers and ontologies

On where analytic semantic layers end and wider knowledge infrastructure begins, Talisman began with an opportunity she thinks is underused. At Pluralsight her team built a rich taxonomy and worked with analysts to turn it into a radar: what customers were searching for, where interest was growing, and how that lined up with industry trends. She calls the combination knowledge analytics.

Mullins described the pattern Coginiti sees with customers, which is to run the two as parallel systems behind a routing step. A question about fourth-quarter sales or the number of customers goes to the semantic layer. A question about how sales are calculated or how a customer is defined goes to the ontology. The hard part is keeping meaning in sync between them. Apache Ossie, the open-source project for exchanging semantics, has both an ontology working group and an analytics working group for this reason. (See our posts on Apache Ossie governance and the 0.2 release.)

Talisman has seen the failure from the other direction. Teams that deploy RDF taxonomies often validate against YAML configuration files, and the flattening loses much of the context the richer model held.

Who owns it

Talisman does not like answering "it depends," so she gave the conditions. A champion in leadership has to come first. After that, ownership follows how the organization already works. If governance is distributed and the company is committed to domain-driven design, knowledge infrastructure will be distributed too, with a core team able to manage it at scale. A less distributed company in a regulated industry needs a central team. In her current engagement the work is mostly taxonomies and metadata statements so far, under a program called data foundations, and much of it is teaching people the practice. A single enterprise knowledge graph is not there yet.

Watch the full conversation

The recording runs above. Talisman closed with what she is reading: The Book of Why by Judea Pearl and Dana Mackenzie, on the science of cause and effect. For related reading, see our post on the three meanings of "semantic layer" and the earlier webinar on semantic modeling with Christina Salmi.

See Semantic Intelligence in Action

Coginiti operationalizes business meaning across your entire data estate.