Databricks says Genie Ontology achieved a 91% pass rate on workspace tasks, compared with 45% and 58% for 2 frontier AI agents. The test used a fictional business whose reporting rules were scattered across tables, notebooks and documents.
Databricks tested its approach, built with Turing, in a live Databricks workspace representing fictional UK supplier Calderwick Energy. Each task ran 5 times, with all agents subject to Unity Catalog data-access controls.
Genie Ontology extracts business definitions from notebooks, dashboards, saved SQL queries and documents, then retrieves relevant snippets when Genie answers. In one task, it found the correct historical planning rate in a notebook rather than using the different rate shown in a table.
For companies building AI data tools, the result points to the importance of retrieving business context, not just reading tables. But the comparison has limits: the rival agents used a Databricks toolbelt rather than running natively, and Databricks says the performance gap narrows when workspace knowledge is sparse or tables and schemas are sufficient.
Existing notebooks and reporting documents may be useful inputs for AI systems that need to interpret company-specific business rules.
Builders should examine tool access and retrieval setup before treating an agent benchmark as evidence of a better underlying AI model.
Performance beyond this company-reported test remains unproven, especially in deployments with little stored business context.
Written automatically by the Made in 10 Newsroom from the source above.
BloombergNEF research finds AI data centres plan 181GW of on-site gas-fired power across the US, UK and mainland Europe. The proposals aim to bypass electricity grid connection delays, with 90% of the planned gas-fired power stations in the US.
Databricks has launched an updated Genie assistant and AI agents that continuously analyse company data. A separate Ontology feature maps information across an organisation so staff can retrieve it and run analyses.
Fractile has raised a $720 million Series C funding round to build chips for running AI models. Its specialised chips are planned for data centre deployment next year, so the promised performance gains remain a future test.