What is an AI Ontology and Why Enterprises Need it

Shelf’s Aristotle AI

What you will learn

You will learn how to use AI ontologies to build ontologies 80 times faster, to improve the accuracy of the applications that run on your ontologies to 99%, and to maintain the entire system with several people instead of a team, for an enterprise of any size.

BeforeAfter

Setup takes months per project

Setup in a week

Requires a project team

Target: 1 person

Experts are bottleneck

Train AI Ontologist to be business expert for unlimited scale

Never complete, always chasing greater coverage

Target: 97%+ use case coverage in weeks

Difficult to maintain & constantly going out of date

AI maintenance assistant proactively helps manage ontology

1

What is an ontology for AI?

A machine-readable explanation of how a business understands its world.

An ontology defines concepts—such as customer, product, contract, and entitlement—their properties, relationships, and logical restrictions. An AI ontology is an ontology used to ground AI in those explicit business definitions.

customerproductcontractentitlementpropertiesrelationshipslogical restrictions

It defines meaning beyond a list of terms or connections.

A taxonomy organizes categories; an ontology also specifies relationships and their semantics. A knowledge graph can populate that model with actual customers, purchases, and entitlements.

Take a product table. It lists each product, with a tag such as “Pro”. The tag alone says little.

An ontology gives the tag its meaning. “Pro” is a tier. The tiers have an order: Basic, Plus, Pro. And each tier is for one kind of customer: a small, a medium, or a large business.

product tableBasicProPlusa tag alone says littleontology: “Pro” is a tierBasicsmall businessPlusmedium businessProlarge businessthe tiers have an order

It gives AI an explicit reference for interpreting enterprise information.

Documents and databases can be mapped to common concepts, allowing applications to interpret different terminology consistently. The ontology must be connected to retrieval, queries, or reasoning tools to affect an AI system’s behavior; merely creating it is insufficient.

The key is the link. On your computer, you put the files of one project into one folder. An enterprise does the same. But an enterprise has millions of projects and millions of files. Its folders, tags, categories, and metadata become too complex to manage. An ontology is the system that manages them.

For example, a refund procedure is one thing. The ontology links that one thing to its documents and to its conversation topics. The same links then serve topic modeling, the analytics of chats, business KPIs, and more.

linked todocumentsconversation topicsrefund procedureone thingtopic modelinganalytics of chatsbusiness KPIsthe same links serve these, and more

It comes from documents, not from the tables of a data warehouse.

Some tools build an ontology on top of a data warehouse. That data has structure before the work starts: tables, columns, and types.

Much of what a business knows is not in tables. It is in documents: policies, procedures, manuals, price lists, slides, videos, and recorded calls. These have no structure.

An AI ontology for that knowledge comes from the documents themselves. That is much harder work than to read the columns of a table.

DocumentsReturn Policy.pdfPrice list.xlsxService manual.docxTraining video.mp4Recorded call.mp3no structureAI ontologyData warehousehas structure
2

Why are AI ontologies important for enterprise AI?

They support answers that depend on relationships and business context.

Such answers require connected facts and defined relationships. An ontology-and-mapping layer improves the accuracy of these answers. For example: “Which customers qualify for this service under their contracts?”

facts on their owncustomercontractservicethe answer misses the business contextwith an ontology: defined relationshipscustomercontractservicehascoversthe answer uses the business context??

They reduce ambiguity across departments and systems.

“Customer,” “active account,” or “eligible product” can mean different things in different workflows. Ontologies make those definitions and distinctions explicit, helping applications combine information without silently treating unlike concepts as equivalent.

the same wordSalesSupportontologycustomer2 meaningsactive account2 meaningseligible product2 meaningsunlike concepts are never treated as one≠≠≠

They make parts of AI behavior inspectable and testable.

Explicit models enable logical consistency checks; complementary SHACLThe Shapes Constraint Language: the W3C standard to check the data of a graph against rules. constraints validate data; provenance models record sources and derivations.

AI answerlogical consistency checkson the modelSHACL constraintsvalidate the dataprovenancerecords sources and derivations

A customer asks your AI assistant: “How many days do I have to return my shoes?”

The assistant finds three files: Return Policy.pdf, Refund Policy.pdf, and Supplier Returns.pdf.

The Return Policy and the Refund Policy have different names. But they are about one thing, a customer return: the customer sends a product back and gets the money back. The customer has 30 days.

The Supplier Returns file has the same word in its name. But it is about a different thing, a supplier return: the shop sends products back to its supplier. The shop has 90 days.

The words do not tell what a file is about. An assistant that goes by the words can answer “90 days”. An ontology says what each file is about. Then the assistant answers “30 days”.

How many days do I have to return my shoes?By the wordsReturn Policy.pdfRefund Policy.pdfSupplier Returns.pdf“return”90 daysWith an ontologyReturn Policy.pdfRefund Policy.pdfSupplier Returns.pdfcustomer returnsupplier return30 days
3

Why building ontologies the old way does not scale

Expert knowledge becomes a production bottleneck.

Teams must translate business requirements into concepts, relationships, hierarchies, and formal definitions. This requires repeated collaboration between domain specialists and ontology engineersSpecialists who build an ontology and keep it correct. They write the concepts, relationships, and rules of a business in a formal language that software can read.—work that must be repeated as the scope expands.

business requirementsdomain specialistsontology engineersBottleneckformal definitionsconceptsrelationshipshierarchies

Ontology Modeling is labor intensive.

Teams must resolve competing interpretations and test whether the model answers the questions it was built for. This evaluation work is labor-intensive and error-prone even for experts.

competing interpretationscustomer≠questions it was built forchecked by expertsQQQerror-proneQQ

Maintaining Ontology is never ending and very challenging.

Ontology development is inherently iterative. Changes to business definitions require reviewing the model, mappings, and dependent applications. A project that funds initial authoring without ongoing ownership will struggle to remain useful.

business definitions!modelmappingsdependent applications

The Semantic relationships are difficult to automate.

LLMs can identify terminology more reliably than they construct correct property definitions, hierarchies, and logical axioms. A plausible-looking ontology may still encode incorrect meaning.

LLMscustomercontractproductentitlement?property definitions?hierarchies?logical axiomsincorrect meaningplausible-looking, still wrong

Bringing new information does not produce an accurate graph.

Document extraction, source mapping, entity resolution, and validation are distinct implementation tasks. An ontology does not automatically produce a complete, accurate knowledge graph

new informationknowledge graphmissing1Document extraction2source mapping3entity resolution4validation

Quality assurance is a black box and incomplete.

Matching terminology or producing valid syntax does not prove that an ontology represents the domain adequately. Quality assurance for ontologies is not an established practice. An ontology is metadata, so a mistake in it costs more than a mistake in a document: one wrong document gives one wrong answer, and one wrong definition gives a wrong answer each time it is used. And no tool shows the impact of a bad ontology. That is why its quality is a black box.

Document1 mistake1 wrong answerOntology1 wrong definition?no tool shows the impactretrievalwrong answerquerieswrong answerreasoning toolswrong answerAI assistantwrong answerwrong answers, each time it is used
4

The New way of automating the building of Ontologies

You can build each part yourself, and an AI assistant helps with each step. Each step has an easy start. Each step also has a hard part.

1Transform documents into a structured, AI readable format

Today you can give a document to ChatGPT, Claude, or Gemini. Ask the assistant to convert the document into a format that AI can read.

This method has two limits. You convert one document at a time. And you must say what “AI readable” means. The assistant does not know which format you want.

A good first definition is AI-ready content. It keeps the headings, the lists, and the tables of the document. For each picture, one line says what the picture shows.

Return Policy.pdfAI-ready contenta map of the documentTabledays, dollarsPicturechart → dataStruck30 daysremovedLink→ Exceptions PolicySourcep. 2one document at a time?you define “AI readable”

2Mine ontology entities

Ontology mining is a process of parsing, extracting, and building new structures and relationships from organizational documents & files.

Ask the same assistant to read each document. Tell it to list the concepts, their properties, and the relationships between them.

Then reconcile the lists. This step is the difficult part. Two documents can use different names for one thing. And one name can mean two different things. Merge the duplicates. Then decide which thing each name refers to.

The assistant can propose each merge. A person who knows the business makes the decision.

Document ARefundCustomerReturnDocument BMoney backClientReturnrefund2 → 1customer2 → 1customer returnsupplier returnbusiness expertApprove merge

3Build agentic conversational AI layer to help manage, change, update ontologies

Now you have a lot of data. Each change to the data needs a decision. Make a place to review the changes, like an inbox. Each change stays there until a person accepts it or rejects it.

Ask your AI assistant for each change in plain words. The assistant puts each change into the inbox. It must not change the ontology directly.

To try this yourself, a spreadsheet is enough. We started with one too.

ontology-starter.xlsx

But a spreadsheet is for the start only. An enterprise will have millions of entities to manage: each customer, each product, each order, each document. Then the ontology needs a graph database, such as Neo4j, Amazon Neptune, JanusGraph, or QLever.

A graph database is a system of its own. It needs its own servers, its own query language, and people who know both. And an ontology is not one more table of data. It is data about your data. So it needs its own owners, reviews, versions, and checks.

ontology-starter.xlsxOntologyworkbookHomeConceptsPropertiesRelationshipsRulesInboxQualityYour ontologyBuild your ontology here, and keep it correct.Concepts5things of yourbusinessOpen →Properties5facts about theconceptsOpen →Relationships4links betweenthe conceptsOpen →Rules4things that arealways trueOpen →Inbox2changes wait fora decisionOpen →Quality1problems torepairOpen →Start withthe 10 mostimportantthings of yourbusiness.From the guide “What is an AI Ontology and Why Enterprises Need it”, shelf.ioHomeConceptsPropertiesRelationshipsRulesInboxQuality

4Build quality assurance application to monitor & identify issues, gaps, and inaccuracies in ontology

Check the ontology after each change. Simple rules find the first problems. A concept has no definition. Two concepts have the same name. A relationship goes to a concept that is not in the list.

Then let AI read each definition and say if it is clear. A person reads each definition that is not clear, and repairs it.

These checks find what is missing. They do not find what is wrong. A definition can be clear and still not be true for your business. That check needs a person who knows the business, after each change.

In a spreadsheet, one formula asks the question for each row. Google Sheets has an AI function, for accounts with an eligible Google Workspace or Google AI plan: =AI("Is this definition clear to a new employee? Answer Yes or No, and give the reason.", E11)

fx=AI("Is this definition clear…", E11)ConceptDefinitionChecksAIcustomer returnsupplier returnrefundrefundorder→ invoiceNo definitionSame nameMissing conceptYesNo · reason

All of this works for a first try. At the size of an enterprise, each step becomes a system of its own. Somebody must build it, run it, and keep it correct.

5

AI Ontology Architecture

Here is how you would build your AI ontology architecture

Input
organizational documents & files
  • Documents
  • databases
AI Ontology Infrastructure
1
ontology templates
  • public vocabularies
  • ontologies of your industry
  • pieces that fit
2
structured, AI readable format
  • plain text
  • headings
  • lists
  • tables
3
Parse, Extract, and build
  • read each document
  • reconcile the lists
  • Merge the duplicates
ontology
  • concepts
  • properties
  • relationships
  • logical restrictions
4
agentic conversational AI layermanage, change, update
  • inbox
  • AI assistant
  • review the changes
5
quality assurance applicationmonitor & identify issues, gaps, and inaccuracies
  • simple rules
  • AI reads each definition
  • a person who knows the business
Output
AI applications
  • retrieval
  • queries
  • reasoning tools
Detail
6

Before/ After

BeforeAfter

Setup takes months per project

Setup in a week

Requires a project team

Target: 1 person

Experts are bottleneck

Train AI Ontologist to be business expert for unlimited scale

Never complete, always chasing greater coverage

Target: 97%+ use case coverage in weeks

Difficult to maintain & constantly going out of date

AI maintenance assistant proactively helps manage ontology

Quality monitoring is a blackbox

Always on, transparent AI quality monitoring

7

Build in months or get it out of the box right now

Two years ago, we started to build ontologies for enterprise AI. We had no tools for this work. We started with a spreadsheet too. Then we tried graph databases, one after another.

Graph databases are slow, expensive, unreliable.

  • We started on a property graph database, AWS Neptune.
  • It became slow when many requests ran at the same time.
  • We tried many others. Neo4j. GraphDB. Qlever. Oxigraph. JanusGraph. And more.
  • All of them fell short in some way.
  • Property graphs lack semantic layer and OWL reasoning.
  • Graph databases in general are not as performant and reliable as traditional OLTP databases. So they cannot be trusted as source of truth.
  • So we still need an OLTP database in front of the graph database. And the moment we add it, we own a data sync problem between the two.
  • So we had to build our own because nothing existed close to that on the market.

It's hard to collaborate on managing ontologies with the team.

  • Many people write the definitions, and each one defines the same entity in a different way.
  • Should intents be hierarchical or flat?
  • There is no collaborative editing software. Each person's contributions have to be reconciled, agreed upon, etc.

It's slow and expensive to process many documents at scale

  • To build ontologies you have to continuously process documents.
  • Ontologies require higher precision of document understanding and require processing them with expensive models.
  • Take 100,000 documents and an LLM run that costs $5,000 to mine them.
  • If 1% of the documents change each day, that is $50 more each day.

Versioning and trying out different branches.

  • A document has one version. An ontology is a snapshot of the entire business.
  • How do you try a new one before it goes live? How do you know that it is better than the old one?
  • First we went with spreadsheets.
    • One file for each team.
    • Nothing checks that a link points to something that exists.
    • A new version is a copy of the file.

We know what does not work. So we built what was missing. It is Shelf Cortex, and its assistant is Aristotle AI.

You tell Aristotle in plain words what you need. Aristotle does the work: it mines the ontology from your documents, finds what is wrong, explains it, and offers the fix. It runs at the size of an enterprise. That is millions of documents, thousands of updates an hour, and tens of thousands of people in one ontology.

4Aristotle AI assistantthe conversational interface on top of everything3Cortex AI knowledge studiothe key workspace110,000's of ontology templatescountriespackages & goodsimport / export rulesa preset, e.g. for a logistics company2Cortex Knowledge data processorcrunches incoming dataorganizational documents & filesdocumentsdatabases5Cortex Sentinelchecks before publishingrulesupdatesCopilots& AI appsconsumers
FeaturesBasic Description
  • 110,000's of Ontology templatesOntologies for an industry, a department, or a use case. You start from one that fits your business, not from scratch.
  • 2Cortex Knowledge data processorIt reads each document, also the most difficult ones: scans, complex tables, slides, audio, and video. It mines the concepts and the relationships automatically.
  • 3Cortex AI knowledge studioOne place where your people read, change, and approve the knowledge. Each change has an owner and a record.
  • 4Aristotle AI assistantYou ask in plain words. Aristotle does the work in the ontology: it finds what is wrong, explains it, and offers the fix.
  • 5Cortex SentinelIt checks each document against the ontology, all the time. It finds contradictions before a customer gets a wrong answer.
8

Impact of using Shelf as your AI ontology infrastructure

  1. Save time

    Weeks, not months to production

  2. Save money

    One person to run and maintain versus a team

  3. Improve ANY AI application accuracy

    New benchmark - 99% accuracy for any use case

  4. Automate more, faster

    quickly expand coverage to new use cases

  5. Infrastructure that can scale across organization

    same infrastructure can support any department and AI initiative

If you want to learn more, let us know.

Talk to an expert