How Ontology Powers AI Analytics: Making Companies AI-Ready

Updated October 2026

An ontology powers AI analytics by telling the AI what your data means. It defines entities like Customer and Contract, how they relate, and the rules behind your metrics, so a question in plain English maps to the right data. In 2026 studies, adding explicit business semantics raised answer accuracy by 17 to 23 points, more than switching models did.

What is an ontology?

An ontology is a formal model of what your business data means. It names the entity types that matter, such as customers, accounts, products, and contracts, the relationships between them, and the rules that apply. A schema tells a database that one column joins another. An ontology tells software that a customer owns accounts, an account holds contracts, and a customer counts as active only under certain conditions.

Salesforce frames this well in a November 2025 post on the ontologies agents need. A descriptive ontology works like a business dictionary of meaning and rules. A structural ontology works like an atlas that ties each concept to where the data actually lives. AI analytics needs both.

How does an ontology power AI analytics?

The ontology sits between the question and the data, and does the translation an LLM cannot do reliably on its own. Here is what happens when someone asks which enterprise customers are at risk of churning.

1. Resolve the words. The ontology maps enterprise customer and churn risk to defined concepts, with the exact criteria for each.

2. Resolve the entities. It finds every record for each customer across CRM, billing, and support, and treats them as one entity.

3. Follow the relationships. It walks from customer to accounts, contracts, usage, and open tickets, joins a model would otherwise have to guess.

4. Apply the rules. It uses the governed definition of churn risk rather than whatever the model infers on the spot.

5. Check the result. Validation rules catch queries that break the model before a wrong answer reaches the user.

Without those steps, the LLM writes SQL against raw tables and hopes the column names mean what they appear to mean. That is why the same model gives sharply better answers with an ontology in front of the data.

What does the evidence show?

Every credible study points the same way. Semantics matter more than the model.

The 2023 data.world benchmark found GPT-4 answered 16 percent of enterprise insurance questions correctly from raw SQL, and 54 percent from a knowledge graph of the same data. A 2024 follow-up used the ontology to check and repair generated queries and reached 72 percent, with errors falling to about 20 percent.

The newest result is the most useful. In an April 2026 study of 100 questions over a retail dataset, adding explicit business semantics lifted accuracy by 17 to 23 points across Claude Opus 4.7, Claude Sonnet 4.6, and GPT-5.4. The gap between models was far smaller than the gap the semantics created. dbt Labs found a similar pattern in its own vendor benchmark, counting only questions its semantic layer covered.

Why is every major vendor building an ontology?

Because whoever owns the meaning of your data owns how your AI behaves. In the past year every large platform shipped its own version.

Salesforce now positions Data 360 as the semantic layer its agents reason over, and it closed its Informatica acquisition in November 2025. SAP presented a Knowledge Graph covering about 7.3 million ERP data fields, as Futurum reported.

Snowflake made Semantic View Autopilot generally available to generate governed models automatically. Fabric IQ can generate an ontology straight from existing Power BI models, and Databricks added a business semantics layer to Unity Catalog.

Each of these is good inside its own platform. The problem is that most companies run several. A customer lives in Salesforce, their orders in SAP, and their usage in Snowflake, and each vendor's ontology sees mainly its own slice.

Why should you own your ontology?

Your ontology is the written-down version of how your company works, and it is one of the most valuable assets you have. If it lives inside one vendor's semantic model, that vendor decides what your AI can see across the rest of your stack, and what it costs to leave.

Owning it means three things. The model is written in open standards like OWL, SHACL, and SKOS, or in code in your own repository. It spans every system, not one vendor's slice. And it feeds every tool you use, including the vendor semantic layers above, rather than being trapped in one of them.

Galaxy works with ambitious teams to build that foundation on the stack they already run. We use internal tooling for ontology design, semantic modeling, and entity resolution to get a first domain working fast, and the model stays in your environment in open formats. We build in the open, starting with Filament for moving data between systems, and plan to open source more of the semantic tooling.

What are the ways to get an ontology for AI analytics?

There are four realistic paths, and they combine well.

The ontology platforms worth knowing are Palantir Foundry for operational decisions, Stardog for query-time reasoning over virtualized sources, and Timbr for SQL-first ontologies. Timbr added Salesforce ontologies in December 2025. Also consider metaphactory for visual modeling, and Graphwise for standards-based GraphRAG. Our knowledge graph platform comparison covers each in depth.

How do Salesforce, SAP, and Snowflake fit into one ontology?

Most enterprises need one model that spans all three, and each brings its own challenge.

Salesforce

Salesforce objects like Account, Opportunity, and Contact carry relationships that do not map one to one onto enterprise concepts like Customer or Revenue Event. Mapping means defining which Salesforce fields and objects correspond to which ontology classes, and resolving accounts that also appear in billing and support. The Data 360 object model already unifies identities inside Salesforce, which is a useful input.

SAP

SAP S/4HANA schemas are deep, with cryptic table names that encode decades of process. The work is abstracting tables into business entities like Purchase Order, Cost Center, and Vendor. SAP's own HANA Cloud Knowledge Graph Engine stores RDF and answers SPARQL, so a standards-based ontology can live close to SAP data.

Snowflake

Snowflake is often where Salesforce and SAP data meet. Semantic views define metrics and relationships natively, and Cortex Analyst uses them to answer questions more accurately. An enterprise ontology can generate or feed those views, so the definitions Snowflake serves match the ones every other tool uses.

How do you build your first ontology?

Start small and prove it on real questions. These are the steps we follow.

1. Pick one domain and one consumer. Customers feeding a revenue or churn agent is a strong first choice.

2. Write the questions first. List the 20 to 50 questions the consumer must answer, because they define what the ontology needs.

3. Model only what the questions need. Define the entities, relationships, and rules in open standards, and reuse public ontologies like FIBO or schema.org where they fit.

4. Map and resolve. Map each source system to the model and merge duplicate records into single entities.

5. Validate against the questions. Run the question set, check every answer against a known result, and fix the model where it fails.

6. Expose it and expand. Serve the ontology to your AI tools through APIs or MCP, then add the next domain.

LLMs speed up the drafting. A 2025 evaluation found models like o1-preview produced ontologies that met many requirements set by ontology engineers, though quality varied. Use them for first drafts and keep people responsible for meaning. For the mapping step, our guide to ontology mapping goes deeper, and our comparison of RAG, knowledge graphs, and semantic layers shows where the ontology fits in a full AI stack.

Frequently asked questions

What is an ontology in AI analytics?

An ontology is a formal model of what your business data means. It defines entity types such as Customer and Contract, how they relate, and the rules that apply, such as when an account counts as active. AI analytics tools use it to translate a plain question into the right data, joins, and definitions instead of guessing from column names.

How does an ontology improve AI accuracy?

It removes guessing. In a 2026 study on retail data, adding explicit business semantics raised accuracy by 17 to 23 points across three current models. Earlier work found GPT-4 answered 16 percent of insurance questions from raw SQL and 54 percent from a knowledge graph, and an ontology check pushed that to 72 percent.

What is the difference between an ontology and a semantic layer?

A semantic layer defines metrics and dimensions so numbers are calculated one way, such as net revenue. An ontology defines the entities and relationships behind those numbers, such as which customer owns which account and contract. Many 2026 products combine both, including Microsoft Fabric IQ and Salesforce Data 360, and most AI analytics needs both.

What is the difference between an ontology and a knowledge graph?

The ontology is the model and the knowledge graph is the data shaped by it. The ontology says customers own accounts and accounts hold contracts. The knowledge graph holds the actual customers, accounts, and contracts, linked according to those rules. You design the ontology once and the graph fills and updates from your source systems.

Do Salesforce, SAP, and Snowflake have ontologies?

Each has moved that way. Salesforce describes structural and descriptive ontologies for agents and a Semantic Data Model in Data 360. SAP introduced a Knowledge Graph that grounds Joule across its ERP data. Snowflake offers semantic views with Semantic View Autopilot, generally available since February 2026. Each covers its own platform best.

Can an LLM generate an ontology automatically?

It can draft one, not finish one. Research in 2025 found models like o1-preview can produce ontologies that meet many engineer requirements, but quality varies and business meaning still needs expert review. Tools like Snowflake Autopilot and Fabric IQ generate starting models from existing schemas, which speeds the first draft considerably.

Which open standards are used for enterprise ontologies?

The core W3C standards are RDF for data, OWL for the ontology itself, SHACL for validation rules, and SKOS for taxonomies. Industry ontologies build on them, such as FIBO for financial concepts. Using these standards keeps the ontology portable between tools, which matters more as vendors build their own proprietary semantic models.

Should we build our own ontology or use a vendor's?

Use a vendor's when one platform holds nearly all your data and you accept its lock in. Build your own in open standards when data spans several systems, such as Salesforce, SAP, and a warehouse, because each vendor's ontology covers its own platform best. Many teams build with a partner so the model is theirs from day one.

How do you start building an ontology for AI analytics?

Pick one domain, such as customers, and one consumer, such as a revenue agent. List the questions it must answer, define the entities and relationships those questions need, map your source systems to them, resolve duplicate records, and validate with real questions. Then add domains one at a time rather than modeling the whole company first.

What makes a company AI-ready?

Clean data is not enough. AI-ready means the business meaning behind the data is written down where software can use it. That includes resolved entities, governed metric definitions, documented relationships, and access rules, all kept current from source systems. The ontology is the part most companies are missing.