Databricks ships Genie Ontology: agents answer 91% of benchmark runs correctly, versus 58% for the best frontier agent

Once enterprises deploy many agents across every function, each one needs the same accurate and complete understanding of how the business works, and that context is far larger than any team can write down or model by hand. Enterprise ontologies were previously built by specialists and maintained manually, capturing only a narrow slice of the business and quickly becoming outdated as the business evolved. Databricks says thousands of customers already run Genie Agents in production with hand-curated context, including 7-Eleven, Virgin Atlantic and General Motors.

Genie Ontology splits knowledge into automatically learned and explicitly modeled parts, and an authority-aware ranking algorithm picks the most trustworthy context for each question. The learned part comes from the Databricks dashboards, SQL queries, notebooks and semantic models in tools such as Power BI that teams already use, captured as reusable ontology snippets that refresh as the underlying assets change. The modeled part is governed through Unity Catalog semantics, including Metric Views holding certified KPIs, Domains organizing knowledge by business area and Pages recording authoritative definitions; OntoRank weighs source authority, certification, the expertise and relevance of the people behind the knowledge, popularity and recency to resolve conflicting definitions.

On a benchmark built with Turing, Genie powered by Genie Ontology answered more questions correctly than two frontier agents while using fewer tokens and finishing faster. The benchmark simulates a live, realistic Databricks workspace for a fictional UK energy supplier, motivated by existing data agent benchmarks focusing mostly on tables, frontier agents scoring similarly on them, and real enterprise questions depending on business context fragmented across notebooks, dashboards and documents. Every agent answered each task five times and answers were graded against a rubric defined before the runs: Genie passed 91% of runs at 100 seconds and 432K tokens per run on average, Agent A 45% at 435 seconds and 889K tokens, and Agent B 58% at 307 seconds and 1.05M tokens.

The demonstrated scope is a benchmark on a Databricks workspace for a fictional UK energy supplier, plus the thousands of customers Databricks says already run Genie Agents in production with hand-curated context; the results reflect an established deployment where teams have accumulated that knowledge. Databricks says account teams can help run the same evaluation in a customer's own workspace.

Databricks states the results reflect an established deployment where teams have accumulated the knowledge, so performance on a fresh workspace is not covered by the benchmark. External knowledge sources are in Beta and support only Power BI today, with Tableau and Sigma support and lineage support still to come; Ontology Manager is still to arrive in Beta.