Databricks Genie: What It Is & How To Use It In 2026
Databricks Genie is the natural-language analytics feature in Databricks: people ask a business question in plain English, and Genie writes SQL against data in Unity Catalog, runs it on a SQL warehouse with the asker's own permissions, and returns a table, a chart and the query. Data teams curate it in Genie Agents (called Genie spaces, or AI/BI Genie, until July 2026), and business users reach it through Genie One (formerly Databricks One), dashboards, Slack, Microsoft Teams and an API. As of November 2026, people's use of Genie One and Genie Agents is free through January 31, 2027, but the SQL warehouse that runs its queries is billed as usual.
Databricks renamed most of this in 2026, so older guides use names the product no longer shows. Everything below is as of November 2026, with Databricks' documentation linked.
Genie One, Genie Agents, Genie Code: which Genie is which
"Genie" now names three products. Older pages say "AI/BI Genie" because AI/BI is Databricks' name for its BI line: AI/BI dashboards plus Genie.
Product | What it is | Former name | Who it's for |
|---|---|---|---|
Genie One | The business-user app for questions, dashboards and Genie Agents; Consumer access gives a view-only version (docs) | Databricks One; Genie One since June 2026 | Business users |
Genie Agents | A curated chat over one domain's Unity Catalog data (docs) | Genie spaces, AI/BI Genie; renamed July 2026 | Built by data teams, used by anyone with access |
Genie Code | An AI agent for notebooks, SQL, pipelines and dashboard authoring (release notes) | Databricks Assistant; renamed March 2026 | Data engineers, scientists and analysts |
This guide covers Genie One and Genie Agents. When someone says "Genie space", they mean a Genie Agent.
How a Genie Agent answers a question
Someone asks a question, Genie generates SQL from the agent's tables and curated context, runs it, and returns the SQL, a result table and a visualization (Databricks: Genie Agents). Each question runs with the asker's own credentials, so Unity Catalog grants, row filters and column masks decide what they see (Databricks: Genie in Slack; Genie One API).
For harder questions there is Agent mode, generally available since July 2026: it writes a research plan, runs several SQL queries, learns from each result and iterates, then returns a report with findings, visualizations and citations to its research steps. A Flash option (Beta) answers quick questions faster (Databricks: Agent mode; release notes).
How to set up a Genie Agent
An agent needs data in Unity Catalog and a SQL warehouse. Its accuracy comes from the context you add, which Databricks ranks in order: well-documented datasets, then SQL expressions for business semantics, then example SQL, with plain-text instructions as the last resort (Databricks: curate an effective Genie Agent).
Element | What it does | Rules and limits (November 2026) |
|---|---|---|
Data | The Unity Catalog tables, views and metric views the agent can query | One domain per agent; prejoin related tables into views or metric views rather than adding more (set-up docs) |
Descriptions | Table and column comments Genie reads to choose fields | First priority |
SQL expressions | Reusable definitions of metrics, filters and dimensions | Where "active customer" gets one meaning |
Example SQL queries | Complete queries for hard or multi-part questions, static or parameterized | Title each with the most typical phrasing of the question (tuning docs) |
Trusted assets | Example queries and SQL functions that give verified answers to recurring questions | Users need EXECUTE on any SQL function used (trusted assets docs) |
Text instructions | Plain-language guidance on terms and conventions | Last resort |
Benchmarks | Test questions with an optional SQL answer to compare against | Two to four phrasings of each question (benchmarks docs) |
If you define metrics as Unity Catalog metric views (measures and dimensions declared in YAML inside Unity Catalog), add them as data: the best-practices guide calls them particularly effective for Genie. Our semantic layer tools guide compares them with dbt, LookML and Malloy, the semantic layer we recommend.
Start with the few tables that answer most of a domain's questions, and write benchmarks before you invite users. A result counts as Good when Genie's SQL or its result set matches your SQL answer; extra columns or an empty result count as Bad. Benchmarks measure the agent but don't teach it, so re-run them after every change. Agents can also be managed as code with Declarative Automation Bundles (Databricks community blog).
Where people use Genie
Beyond the Genie One app, Genie reaches people in these places:
Surface | How it works | Status |
|---|---|---|
AI/BI dashboards | Viewers ask follow-up questions about a dashboard's data (release notes) | On by default for published dashboards since July 2026 |
Slack | DM the app or mention @Genie in a channel; it answers through Genie One or an agent the channel owner pinned (docs) | Public Preview since June 9, 2026 (June notes) |
Microsoft Teams | DMs and channel mentions; on Azure, Databricks picks a SQL warehouse the user can access (docs) | Beta on AWS since June 10, 2026; Public Preview on Azure |
Microsoft 365 Copilot, Copilot Studio, Foundry | Databricks recommends the Teams app for Microsoft 365 Copilot (docs); Copilot Studio adds Genie Agents as tools (docs), and Foundry can add Genie too (Microsoft) | Check preview settings in your tenant |
Genie Agents API | Start conversations and follow-ups with one agent from your own app (docs); the Agent mode API streams multi-step answers | Agent mode API in Beta |
Genie One API |
| Beta; an admin enables it |
MCP | Managed MCP servers let outside agents ask Genie One (docs) or one Genie Agent (docs) | Genie One server generally available; Genie Agent server in Public Preview; the older Beta Genie endpoint ends November 30, 2026 |
An account admin enables the Slack and Teams previews in the account console first. A Databricks App can also attach a Genie Agent as a resource (Databricks Apps docs). Our guide to AI data analysts for Slack and Teams compares Genie's chat apps with the alternatives.
Databricks Genie pricing
Genie is billed by usage in Databricks Units (DBUs), not per seat. Databricks moved Genie One, Genie Agents and Genie Code to pay-as-you-go billing on July 8, 2026, then paused billing for Genie One and Genie Agents on July 15. What applies as of November 2026, per the AI/BI and Genie One release notes:
Cost | What you pay (November 2026) | What to know |
|---|---|---|
Genie One and Genie Agents, used by people | Free through January 31, 2027 | Extended in August 2026 from July 31, 2026; budget controls don't apply during the promotion |
Genie used by service principals | Billed in DBUs | Excluded from the promotion, so an app calling the API as a service principal pays |
Genie Code | 150 free DBUs per user a month, then pay-as-you-go | 25% promotional discount above the allowance through January 31, 2027 (cost docs) |
SQL warehouse running Genie's queries | Billed separately at normal Databricks rates | Not covered by the promotion or by Genie budgets (budgets docs) |
The warehouse is the bill to plan for. A serverless SQL warehouse bills DBUs for every hour it runs, by size: the smallest, 2X-Small, uses 4 DBUs an hour and a Small 12 (Databricks: serverless DBU consumption). Running eight hours a workday, a 2X-Small uses about 670 DBUs a month (4 × 8 × 21), and a Small three times as much; multiply by the SQL Serverless rate for your cloud, region and tier on Databricks' SQL pricing page. Databricks advises starting serverless warehouses larger than you think and sizing down as you test (Databricks: warehouse sizing).
The notes don't say what applies after January 31, 2027, so budget for DBU billing to return. Genie spend shows up in system.billing.usage under billing_origin_product = 'GENIE' (Databricks community answer), and that table updates every few hours.
Databricks Genie limits
Limit | Value (November 2026) |
|---|---|
Data per Genie Agent | 50 tables, views or metric views (set-up docs); older pages and guides cite earlier limits of 25 or 30 |
Instructions per Genie Agent | 100; each example SQL query, each SQL function and the general instructions block counts as one (tuning docs) |
Benchmark questions | 500 per Genie Agent |
Conversations | 200,000 per Genie Agent and 10,000 messages per conversation (resource limits) |
Data sources | Tables in Unity Catalog; data in Snowflake, Postgres or BigQuery has to be federated in first (Lakehouse Federation) |
Chat apps | Slack and Teams apps are in preview; the Teams app doesn't support PrivateLink |
Accuracy is the curation. An agent knows only its tables and the context you wrote. Practitioners report plausible answers to questions an agent can't actually answer, built from a nearby metric (see the user reports in our Databricks Genie vs Dot comparison). Benchmarks catch that only if you add questions the agent should decline and review those answers by hand. Our text-to-SQL tools guide has a test plan for any tool.
Permissions follow the asker, including in channels. That is precise in a DM and a trade-off in a shared channel: when someone with broad access asks in Slack or Teams, everyone in the channel sees the rows. Decide which channels get the app before you install it.
When GetDot fits alongside or instead of Genie
GetDot (Dot) is an AI data analyst that answers in Slack, Microsoft Teams, email, its web app and MCP clients, with the SQL behind every number. It connects to Databricks with a service principal or personal access token and runs its queries on a SQL warehouse you choose (Databricks integration):
- It starts from your lakehouse, not a blank space. Dot reads Unity Catalog (or Hive metastore) table and column comments, learns from recent query history in
system.query.history, and sees only what its credential is granted. - It reaches past Unity Catalog. Dot connects 45+ sources, including Snowflake, BigQuery and Postgres, and answers through dbt, LookML, Power BI semantic models and Malloy, with no federation step.
- It trains against answers you trust. Train tests Dot on questions with trusted answers, investigates the misses, fixes notes, relationships and table docs in an isolated environment, and reports accuracy before and after (September 2026 changelog). Genie's benchmarks score an agent and leave the fixing to its author, with Genie Code's agent-maintenance skills (July 2026) to help.
- Channels have their own identity. Each Slack channel gets the groups and limits an admin grants, so it sees only that, whoever asks. Direct messages run as the person, with row-level security on the Team plan.
- Analysis beyond one query. Dot uses Python where SQL falls short, runs multi-step deep analysis, schedules reports and alerts with plain-English conditions, and exports PowerPoint.
Choco uses Dot as the front door to its data in Looker and Databricks and reports 3,300+ hours saved a year and an estimated 18x ROI (customer-reported).
Where Genie wins: it is part of your Databricks account, so there is no new vendor to review; Unity Catalog permissions apply per person with no setup; it runs inside your workspace, while Dot is a separate service that syncs metadata and samples; and people use it free through January 2027. Dot is SOC 2 Type II and GDPR compliant, its LLM providers run with zero data retention, and Enterprise can be self-hosted. Pricing (October 2026), unlimited users, billed on credits: Free $0 with 300 one-time credits; Pro $180 a month billed annually ($200 monthly) with 150 credits; Team $720 a month billed annually ($800 monthly) with 800 credits. A chat answer costs one credit.
Choose Genie when your data and your people live in Databricks, questions fit a few curated domains, and someone will own each agent and its benchmarks. Add or choose GetDot when questions cross Databricks and other sources or semantic layers, most people ask from Slack, Teams or email, or you want an analyst that trains against trusted answers. See the side-by-side in our Databricks Genie vs Dot comparison, or try Dot free on your Databricks SQL warehouse.
FAQ
A Genie space is the former name of a Genie Agent: a curated, natural-language chat over up to 50 Unity Catalog tables, views or metric views, with instructions, example SQL and benchmarks. Databricks renamed Genie spaces to Genie Agents in July 2026.
People's use of Genie One and Genie Agents is free through January 31, 2027. The SQL warehouse that runs Genie's queries is billed as usual, and so is Genie usage by service principals.
Databricks One, the simplified interface for business users, has been called Genie One since June 2026. It is where business users ask Genie questions and open dashboards and Genie Agents.
Yes. The Genie Agents API starts and continues conversations with one agent, the Agent mode API (Beta) streams multi-step answers, and the Genie One API (Beta) streams answers as server-sent events and runs as the calling user.
Up to 50 tables, views or metric views per Genie Agent as of November 2026. Guides that say 25 or 30 describe earlier limits.
Yes, in preview: the Slack app has been in Public Preview since June 2026, and the Teams app is in Beta on AWS and Public Preview on Azure. Answers run under each user's own Databricks credentials.
Sven Rudolph
Sven is a co-founder of Dot. He writes about the technical side of AI analytics: MCP servers, BI integrations, and how Dot connects to the data stack.