NVIDIA's open-source Auto Ontology has already received about 5,670 Stars on GitHub. Its premise is clear: when enterprises let AI query data, the biggest risk is not that the SQL will fail, but that it will run successfully and still produce the wrong numbers. Enterprises therefore need a layer of human-certified business semantics between models and databases.
What this is
Auto Ontology is a "semantic layer" connecting enterprise data with AI Agents—software agents capable of autonomously invoking tools to complete tasks. A semantic layer provides consistent definitions for business concepts such as revenue, active users, and customers, specifying field mappings, calculation formulas, and relationships between tables.
The project follows a "generate—certify—answer" workflow: it first extracts candidate terms and SQL from database metadata and query logs, after which people test, correct, and certify them using real business questions. When responding, an Agent retrieves these definitions, checks the query, and attaches the SQL and supporting evidence. When candidate information is insufficient, the system can ask follow-up questions or explicitly state that it cannot answer.
It can be integrated through MCP—an open connection standard that lets Agents invoke external tools—a REST API, or the command line, and supports 10 types of databases, including PostgreSQL, MySQL, Snowflake, and Databricks. One caveat: the project is not currently labeled with a stable release, and its data-ingestion interface lacks authentication in the default deployment. Security configurations must be completed before any public service is launched.
Industry view
We believe Auto Ontology addresses a problem that is easy to overlook: the more capable large language models become, the more readily enterprises package incorrect results as credible answers. Competition in enterprise data assistants is shifting from whether a system "knows how to write SQL" to whether it understands business definitions and can explain the evidence behind its answers.
This also makes it more pragmatic than general-purpose Text-to-SQL tools, which convert natural language into database queries: Auto Ontology embeds human validation into the workflow instead of assuming models can infer everything from field names alone.
The risks are equally clear. A poorly maintained semantic layer can institutionalize outdated definitions, while certification quality depends on continuous participation from business experts. The open-source repository does not accept external Pull Requests, and its README does not identify a stable release. In the short term, it is better suited to validation by technical teams than to replacing mature enterprise semantic models.
Impact on regular people
For enterprise IT: Teams with existing data warehouses and semantic layers can use it as a supplement. The priority should be to evaluate terminology certification, permission isolation, and operational costs—not build a competing set of business definitions from scratch.
For individual professionals: Sales, finance, and operations staff will increasingly see SQL, calculation steps, and "insufficient information" notices when using AI to query data. But whether the system can say "I don't know" still depends on whether the enterprise is willing to maintain reliable definitions.
For the consumer market: AI assistants will increasingly emphasize traceability when providing business figures, but calculations may still vary between companies because they use different definitions. Providing an answer alongside its evidence does not mean that answer is directly comparable across organizations.