Business Ontology & Semantic Model
AI that understands your business, not just your data
A managed ontology and semantic model gives your AI agents shared business terms, rules and governed metrics. We build it from the systems you already run, then maintain it as your business changes.
A business ontology is a map of the things your business runs on, such as clients, invoices and contracts, and how they relate. A semantic model binds that map to your data and defines your metrics once. Together they help AI agents give grounded, consistent answers. Tekscape, a Managed Intelligence Provider (MIP), builds both from the systems you already run and maintains them as a managed service.
Plain English
Ontology, semantic model, semantic layer: what each one does
These terms get used interchangeably. They are not the same thing. Here is how we use them.
| Term | What it is | Plain-English example |
|---|---|---|
| Ontology | The things your business works with, and how they relate to each other. | A Client has many Matters. A Matter has Documents. A Document can be Privileged. |
| Semantic model | The ontology bound to real data, with metrics and business rules defined once. | "Active client" means a client with an open matter or an unpaid invoice. Everyone, and every agent, uses that definition. |
| Semantic layer | The governed access point that tools and AI agents query instead of raw tables. | An agent asks for "revenue by client this quarter" and gets the approved number, not its own guess at a SQL join. |
Raw tables tell an agent what columns exist. An ontology and semantic model tell it what those columns mean to your business.
Why it matters
Why AI agents need a model of your business
A language model knows a lot about the world and nothing about your firm. Point it at raw databases and it guesses. Guesses look confident. That is the problem.
Grounding
Agents work from your entities and relationships, not from a best guess at what a column name means.
- Less room for invented joins and fields
- Answers trace back to named business objects
Shared definitions
One definition of "client", "open matter" or "overdue invoice" for people, reports and agents.
- Finance and operations work from one agreed definition
- New agents inherit the same vocabulary
Governed metrics
Metrics are defined, owned and versioned. Agents are set up to read the approved number, or say when there isn't one.
- Access rules live in the model, not in each prompt
- Changes are reviewed before agents use them
The evidence, with its caveat
One benchmark, one clear signal
In one published benchmark by researchers at data.world, a knowledge-graph vendor, GPT-4 answered questions over an enterprise SQL schema correctly 16% of the time. Over a knowledge-graph representation of the same database, accuracy rose to 54%.
Read that carefully. It is a single benchmark, in the insurance domain, from authors affiliated with a knowledge-graph vendor. Treat it as a vendor figure and as that paper's result, not a promise for your business. It points in a clear direction: models do better when they are given the meaning of the data, not just the tables.
Source: "A Benchmark to Understand the Role of Knowledge Graphs on Large Language Model's Accuracy for Question Answering on Enterprise SQL Databases" (arXiv), by researchers affiliated with data.world. Vendor figure; insurance domain only.
What it looks like
Your business as a graph, not a pile of tables
Each node is a business object. Each line is a relationship your people already understand. Agents follow the same lines.
Example: a Client has Contracts. Contracts generate Invoices. Invoices tie back to the work in Tickets. Ask an agent which clients have overdue invoices on an expiring contract, and it walks the graph instead of guessing at joins.
- Client signs Contract
- Contract bills through Invoice
- Client opens Ticket
- Ticket concerns Asset
- Contract covers Asset
- Vendor supplies Asset
- Employee role resolves Ticket
- Vendor is named in Contract
How Tekscape builds it
From your systems to a managed model
- Inventory
Inventory your systems
We list the systems that hold your business data: practice management, ERP, accounting, CRM, ticketing, file stores and existing reports. We note who owns each one and which numbers people already trust.
- Model
Model core entities and relationships
We name the objects your business runs on and how they connect. We start small, with the entities your first AI agents need, and grow from there.
- Bind
Bind the model to your data
Each entity and metric maps to real fields in your systems. Business rules and access controls are written into the semantic model, so agents inherit them.
- Validate
Validate with your team
Your people check definitions and test answers against numbers they already know. Nothing goes live on our word alone.
- Maintain
Maintain it as a managed service
Businesses change. New matter types, new products, new systems. We update the ontology and semantic model on an ongoing basis, review changes before agents use them, and keep a change history.
Start from what you have
Start from the Power BI models you already have
Most firms already own part of a semantic model. It lives in Power BI datasets, finance reports and the spreadsheets everyone trusts. We start there, not from a blank page.
Power BI semantic models
Your existing Power BI semantic models already define tables, relationships and measures. We reuse those definitions, reconcile conflicts and extend them for AI agents.
Microsoft Fabric IQ ontology (preview)
Microsoft describes the Fabric IQ ontology as "a shared, machine-understandable representation of your business" that can be generated from existing Power BI semantic models. It is in preview, and manual review and data bindings are still required. We use it where it fits and do not promise an auto-generated ontology.
Your own systems of record
Where no model exists yet, we build from the source systems directly. Where you choose private AI, the model runs alongside your data, on hardware you own or in your Azure subscription.
Examples by industry
What the core entities look like in practice
| Industry | Core entities | Questions an agent can answer once they are modeled |
|---|---|---|
| Legal | Matter, Client, Document, Privilege | Which documents on this matter are privileged? Which clients have open matters with no activity this month? |
| Healthcare | Patient, Encounter, Claim | Which encounters have no claim submitted? Which claims are waiting on documentation? |
| Finance | Account, Invoice, Contract | Which accounts have invoices outside contract terms? Which contracts are up for renewal with open balances? |
These are illustrations of structure, not client work or results.
A familiar idea, sized for the mid-market
The approach large enterprises use, run for you
The idea is not new. Palantir describes its Ontology as representing "the complex, interconnected decisions of an enterprise, not simply the data." Microsoft is bringing the same concept into Fabric.
We apply the same principle to mid-market firms in New York and New Jersey, without an enterprise platform program. Tekscape is not affiliated with Palantir; we mention it only as a comparison.
As a Managed Intelligence Provider (MIP), we treat the ontology as part of the service, not a one-off project. It sits under the managed agent harness and the private AI we run on your hardware or in Azure, and we keep it current as your business changes.
Frequently asked questions
Sources
- arXiv: A Benchmark to Understand the Role of Knowledge Graphs on LLM Accuracy for Question Answering on Enterprise SQL DatabasesVendor-affiliated authors (data.world); insurance domain only. Vendor figure; source of the 16% and 54% figures.
- Microsoft Learn: What is Fabric IQ (preview)Fabric IQ ontology is in preview and can be generated from existing Power BI semantic models; manual review and bindings still required.
- Palantir Foundry documentation: The Ontology systemUsed as an analogy only; no affiliation implied.
Find out what your business model looks like to an AI agent
We review the systems and reports you already have and show you where a managed ontology and semantic model could make your AI agents more reliable.
Talk to Tekscape