GILD verdict
Pilot Databricks Genie if your data lives in Unity Catalog and a data analyst will curate and benchmark each domain's agent.
Databricks reports that Genie answered 84.5% of 28 internal test questions correctly on the first try in June 2026, against 52.4% for the strongest general-purpose coding agent. Chat is free until February 2027, so analyst time is the main cost of a pilot.
Databricks Genie
AI Data Analysis by Databricks
+36 PositiveFit and limits
Who it fits
- Your governed data is registered in Unity Catalog, and you want answers to follow each user's own grants, row filters, and column masks.
- Business users wait on the data team for ad hoc questions that recur, and a data analyst can own one domain's agent.
- You already define metrics in Unity Catalog metric views or can start, so Genie and dashboards share the same definitions.
- You want to measure demand while chat is free: named users of Genie One and Genie Agents pay nothing for LLM usage through January 31, 2027.
Where it falls short
- Uncurated agents return believable wrong numbers. Databricks tells curators to start with five or fewer tables and to prefer SQL expressions and example SQL over text instructions.
- Each agent holds up to 50 tables, views, or metric views, and Chat mode answers from structured data only; files in volumes work in Agent mode as a Beta.
- Data outside Databricks needs to be registered in Unity Catalog first, so mixed-platform estates fit a tool that reaches several databases.
- From February 2027, each user's free allowance is 150 DBUs a month. Service principals pay for all usage, and warehouse compute sits outside Genie budgets.
Implementation reality
Curate one agent first
Assign a data analyst to one domain with five or fewer tables or a metric view. Write table and column descriptions in Unity Catalog. Define business terms as SQL expressions, then add example SQL for frequent questions.
Prove it with benchmarks
Add benchmark questions with SQL answers, in two to four phrasings each, and run them before launch. The Genie Agents API documentation recommends at least five tested example queries and five benchmark questions before integrating.
Open it to a pilot group
Share the agent with a few business users through Genie One. Ask them to rate answers and request reviews. Read the Monitor tab each week, then add verified, parameterized queries for recurring questions.
Set budgets and ship as code
Create a Genie budget in Unity Gateway with an alert-only shared threshold during discovery, then per-user limits. Databricks reports that a small fraction of users drive most spend, so set tiers from observed usage. Move the agent into a Declarative Automation Bundle so changes are versioned and benchmarks rerun before each promotion.
What to measure in your pilot
- GILD found no customer measurement of Genie One routing across many domain agents; run each domain's benchmark through Genie One and through the agent alone, and launch company-wide only if scores match.
- Log each pilot user's DBUs under the GENIE_FREE_USAGE SKU in system.billing.usage weekly; anyone above 150 a month will pay US$0.070 per extra DBU from February 2027.
- GILD found no accuracy test of Agent mode reports from Databricks or customers; have an analyst grade 20 reports against hand-built answers, and keep review on until at least 18 pass.
Sources
- Official documentation
- 16
- Databricks docs: Genie
- Databricks docs: How Genie Agents work
- Databricks docs: Create and manage a Genie Agent
- Databricks docs: Curate an effective Genie Agent
- Databricks docs: Tune Genie Agent quality
- Databricks docs: Test and monitor a Genie Agent
- Databricks docs: Genie Agents API
- Databricks docs: Genie One
- Databricks docs: Manage budgets and cost controls for Genie
- Databricks docs: AI assistive features trust and safety
- Databricks docs: Partner-powered AI features
- Databricks docs: Analyze files in volumes with a Genie Agent
- Databricks: Genie pricing
- Databricks: Lakehouse (Databricks SQL) pricing
- Databricks docs: Genie consumption guide
- Databricks blog: Introducing Genie One, Genie Agents, and Genie Ontology (Jun 16, 2026)