Databricks says governed context lifts Genie One accuracy to 95%

Databricks says Genie One reached 93, 95% accuracy across 125 questions when curated Genie Agents were combined with governed business context in Unity Catalog. The company’s study found that agents and governed data address different kinds of questions, making the combined setup more reliable than either approach alone.
- Genie One scored 93, 95% with curated agents and governed context.
- Tables alone scored 64% across the insurer test.
- Curated agents and the governed layer each scored 83% separately.
- Updating one Metric View and Page changed 12 related answers.
- Retailer accuracy rose from 54% to 82% with the governed layer.
Governed definitions and agents solve different problems
Databricks tested Genie One using a fictional commercial auto insurer and four agents for Finance, Reserving, Underwriting and Claims. All four used the same underlying tables, but each had its own team definitions, example queries and documents.
Using tables alone produced 64% accuracy. Curated agents reached 83%, as did the governed layer without agents. When both were used together, accuracy rose to 93, 95% across two runs. The combined configuration also scored 93% on loss-ratio questions and 97% on questions requiring both documents and structured data, according to Databricks.
The company says curated agents improved answers based on unstructured documents, while Metric Views, Domains, Subdomains and Pages improved answers that depended on business definitions. Neither was sufficient on its own to cover both needs.
Why the same business term can produce different answers
Databricks used loss ratio to illustrate the problem. In its fictional insurer, Finance reported an 8.2 percentage-point deterioration between 2024 and 2025, while Underwriting reported a 1.5 percentage-point improvement. Both answers were correct because the teams used different definitions.
Finance measured the calendar year net of reinsurance, so adverse reserve development on an older claim appeared in current figures. Underwriting used the policy year, gross of reinsurance and excluding catastrophes, because that measure was used to set rates. Reserving relied on a third metric, the accident-year ultimate loss ratio.
Databricks says SQL can be correct while still producing an answer that does not make sense to the person asking the question. The challenge is selecting the right business definition while keeping shared facts, such as total policies written, consistent.
What Genie Ontology adds to Genie One
Genie One, powered by Genie Ontology, combines knowledge learned from existing dashboards, queries, notebooks and semantic models with explicitly governed concepts in Unity Catalog. Databricks calls the learned material Ontology Snippets, while the modeled knowledge forms the “governed layer.”
That layer included certified Metric Views, Domains and Subdomains, and Pages containing short explanations of business terms and links to the tables and Metric Views they described. Databricks says teams can build on knowledge already captured in existing assets instead of modeling every concept upfront.
The study also tested whether a definition could be updated consistently. Updating one Metric View and its Page moved all 12 related Genie One answers to the new definition without any agent edits. Databricks repeated the comparison with a fictional retailer, where adding the governed layer to curated agents raised accuracy from 54% to 82%.
Why routing matters for enterprise users
Genie One does not automatically inherit an agent’s instructions unless it routes the question to that agent. If it answers using its own context, a well-curated agent may contribute little. Genie One can instead use governed Unity Catalog Semantics, route to a relevant Genie Agent, or search the workspace when it does not find a suitable match.
For users, Databricks’ findings suggest that reliable answers require both authoritative definitions and access to relevant documents. The study supports using governed business context for terms with multiple valid meanings, while curated agents remain important when answers depend on PDFs and other unstructured data.
Databricks: Genie One Accuracy: Why Curated Agents Need Governed Context


