Why Data Giants Are Racing to Build Business Ontologies for AI Agents

The Problem: Correct SQL, Wrong Business Meaning

The article opens with a concrete example from Snowflake: a user asks "What was net revenue last quarter?" Finance defines net revenue as recognized revenue minus refunds and discounts; Sales defines it as contracted order value minus discounts. Both use the same term "net revenue" but answer different business questions. An AI agent that generates syntactically correct SQL can still return the wrong number if it picks the wrong definition at the outset. The database can validate query executability, but cannot verify that the chosen definition matches the user's intent.

This class of error is harder to catch than a syntax error because the agent may produce a seemingly rigorous report with real data sources, reproducible queries, and full explanations. Reviewers who only check arithmetic can miss the initial definition mismatch. Stronger models help discover ambiguity and retrieve materials, but they cannot decide which department's definition takes precedence or treat a frequently used rule as formally approved.

Consequently, enterprise questioning now demands an extra step: besides finding data and computing, the system must explain which definition was used and why it applies to the current question. Ontology gains attention because this knowledge has long been scattered across reports, documents, and employee experience, making it difficult for agents to reuse reliably.

Ontology Must Connect Business Meaning to Actual Computation

In enterprise data scenarios, an ontology is a set of interrelated business definitions: what customers, orders, and metrics exist, their relationships, applicable rules, and ownership. Unlike a standalone glossary, it must link concepts to actual data assets so an agent can move from "what does this mean in business" to "how is it implemented in data."

Taking net revenue as an example, documenting the definition in a wiki is insufficient. The system must know whether the concept belongs to Finance or Sales, which orders and policies it relates to, and which metric object ultimately calculates it. Snowflake's implementation organizes concepts as nodes connected by typed relationships, then maps them to tables, views, or metrics. The mapping locates the implementation but does not execute the calculation itself.

This separation explains Snowflake's three-layer architecture:

Cortex Sense infers context from existing assets and usage logs.

Business Ontology stores approved business meanings (currently in preview).

Semantic Views execute metric calculations.

Auto-discovery helps teams find candidate knowledge; the ontology decides which definitions are enterprise-approved; semantic views materialize the corresponding calculation rules. The three layers can be combined as needed and need not be built sequentially.

Governance extends to the definitions themselves. Finance and Sales can each retain a "net revenue" concept distinguished by business domain. Changing a formal definition requires a proposal approved by the business owner; until approval, agents continue using the approved version. Technical admin rights do not automatically grant authority to approve business definitions. This design preserves legitimate internal differences — agents must recognize each definition's applicable scope and, when information is insufficient, surface the existence of multiple approved definitions and ask the user to clarify.

Vendors Reorganize Existing Enterprise Knowledge

If every new agent required rebuilding business definitions from scratch, ontology would become a new construction burden. Databricks' November 5 update addresses this by letting Genie read existing Power BI semantic models. In this beta, Power BI measures, calculated columns, and relationships are converted into ontology knowledge fragments that feed Genie's context inference. Genie Ontology also combines actively defined/governed context with context automatically extracted from queries, dashboards, and other assets; the latter receives authority scores based on source, usage frequency, and freshness.

Both context types serve different purposes: manual modeling expresses explicitly approved rules; auto-extraction recovers knowledge scattered in daily work. However, frequent usage only provides reference — it cannot replace business sign-off. A widely copied legacy report may already be outdated. Automation expands coverage, but enterprises must still judge which pieces can serve as current business evidence.

Microsoft Fabric IQ follows a similar path: it supports generating ontologies from Power BI semantic models, linking business entities, relationships, and rules to data; related capabilities remain in preview. Across Databricks, Snowflake, and Microsoft, data platforms are attempting to reorganize business knowledge originally built for reports and analysts into context that agents can consume.

The shift means business knowledge gains an independent maintenance lifecycle separate from any single Q&A interaction. Previously, teams might stuff rules into prompts or documents into retrieval stores for one agent. When multiple agents share the same knowledge, questions of who maintains definitions and how changes propagate demand unified management. The more enterprises want to reuse existing experience, the more they need to reduce each agent's freedom to reinterpret the same business concept differently.

Who Holds Business Definitions Holds the Agent Entry Point

This dynamic turns ontology into a way for data platforms to compete for the enterprise agent entry point. Regardless of which model an agent calls, analyzing revenue, judging customer status, or understanding order relationships requires the enterprise's own business evidence. A platform that simultaneously hosts data, metric definitions, and business relationships gains a stronger foothold in the core of these tasks. Competition extends to: where will enterprises organize and maintain this knowledge, and through what interfaces will it be served to different agents?

"Business interpretation authority" draws a clear line: business owners decide the rules; platforms store, link, and serve them. Vendors vie to become the knowledge host, not to set financial policies for customers. Models can be swapped continuously, but the enterprise's years of accumulated definitions, rules, and usage experience must persist and be reusable by the next generation of applications.

However, becoming a unified entry point must not recreate silos. Databricks reusing Power BI definitions demonstrates that business knowledge needs to flow across tools; Snowflake has also published open-format exchange capabilities for business concepts and semantic definitions. For enterprises, beyond evaluating how easily agents connect, they should assess whether definitions are traceable to sources, whether changes can be continuously maintained, and whether existing knowledge can be ported when switching tools.

Ontology ultimately faces validation by business outcomes: does the agent pick the wrong definition less often for same-named metrics? When rules change, do dependent applications detect it promptly? Do multiple teams reduce duplicate interpretation and maintenance? These matter more than graph size. Even with a perfectly correct finance definition, unrecorded refunds still affect query results — data quality issues remain a data engineering responsibility. Clear definitions, correct calculations, and complete data must all hold together.

As agents begin to analyze and act on behalf of the enterprise, who confirms the rules they rely on and who maintains them can no longer remain ambiguous. SQL enables machines to query data; ontology attempts to make queries follow enterprise-approved business meaning. Stronger models help systems discover ambiguities and understand complex relationships, but enterprises must supply explicit, continuously updated evidence. Rules the enterprise itself has not clarified cannot be expected to be guessed correctly by an agent every time.

References

Snowflake Engineering, "Business ontology: giving AI agents business context", November 2026

Snowflake, "Business ontology: context for enterprise AI", November 2026

Databricks on AWS, "Genie Ontology", updated November 4, 2026

Databricks on AWS, "External knowledge sources", updated November 4, 2026

Databricks on AWS, "November 2026 product release notes", November 2026

Microsoft Learn, "Fabric IQ overview"