What Is Databricks Genie and How Do Its Agents Work

The Genie family in one place

Genie is now the name for a family of AI products across the Databricks platform. Databricks expanded the family at its Data + AI Summit in June 2026, and each product serves a different type of user.

The products are designed to work together. Genie One acts as the main entry point for business users, and it can pull answers from the Genie Agents that analysts have already set up. Genie Code sits on the technical side of the platform. It also opens automatically when someone creates a new Genie Agent, so analysts can use it to help configure the agent.

The rest of this guide focuses mainly on Genie Agents. They are where analysts decide what data Genie can use and how it should interpret business questions.

What a Genie Agent can do

A Genie Agent covers more than question-and-answer chat. Analysts can extend it to handle deeper analysis, outside documents and connections to other applications. The main capabilities are outlined below.

1. Answer data questions in chat

A Genie Agent answers plain-English questions with a SQL query, a results table and, where useful, a chart. Users can ask follow-up questions in the same conversation, and Genie keeps the earlier context in mind. Each answer shows the SQL behind it. This gives analysts a quick way to confirm that Genie used the right tables and filters.

2. Run multi-step analysis in Agent mode

Agent mode handles complex, exploratory questions that a single query cannot answer. Genie breaks the question into steps, runs several queries and tests possible explanations before it responds. A question like “why did renewals drop last quarter” is a good example. Agent mode suits this kind of investigation better than standard chat.

Databricks also offers Flash, a faster option within Agent mode. Flash is meant for quick questions where speed matters more than depth, and it is currently in beta.

3. Work with files and outside sources

Genie Agents can also use data that sits outside Unity Catalog tables.

• Uploaded files: Users can upload a local CSV or Excel file and combine it with Unity Catalog data in their questions.

• Files in volumes: Analysts can attach Unity Catalog volumes so Genie can answer questions about unstructured files alongside structured data.

• External tools: Analysts can connect MCP connectors for sources such as Google Drive and Slack. These connectors work in Agent mode.

4. Connect to other applications and agents

Teams can bring a Genie Agent into their own products through the Conversation API. A Genie Agent can also be embedded directly in an external application. For AI agent workflows, Databricks provides a managed MCP server that lets other agents send questions to a Genie Agent. This server is read-only, and it does not pass conversation history between turns.

How to use Databricks Genie

Most accuracy problems with Genie come from how the agent was set up. An agent built on poorly described tables or too many columns will misread questions, and business users often have no way of spotting the error. The steps below follow the order Databricks recommends, starting with what needs to be in place first.

Before you start

A Genie Agent has a few technical requirements.

• Your data must be registered in Unity Catalog.

• You need a pro or serverless SQL warehouse, with at least CAN USE permission on it.

• Partner-powered AI features must be enabled in your workspace.

Each agent can hold up to 50 tables, views or metric views. Databricks still advises keeping agents much smaller than that limit.

1. Create the agent

Open Genie Agents from the workspace sidebar and create a new agent. Genie Code opens automatically during this step. You can describe what the agent is for in plain language, and Genie Code helps with the initial configuration.

2. Attach a focused set of data

Add only the tables a business team actually needs for its questions. Genie queries the sources attached to the agent and nothing else, even if the instructions mention other tables. A narrow scope with fewer columns gives Genie less room to pick the wrong field. Metric views work especially well here because they predefine metrics, dimensions and aggregations.

A very narrow agent has a trade-off. Users will hit questions it cannot answer, so some teams end up building several small agents for different domains.

3. Improve the metadata

Write clear table and column descriptions, and add synonyms for the terms your teams use. These edits stay within the agent and do not overwrite the metadata stored in Unity Catalog. If you use AI-generated descriptions, review them before keeping them. Prompt matching, which Genie turns on automatically, also helps it map misspelled or informal terms to the right values.

4. Add example SQL and instructions

Example SQL queries are the most reliable way to teach Genie how your business calculates things. Databricks recommends relying on them over long text instructions. SQL expressions can define business terms like “active customer” or “net revenue” once, so Genie applies them the same way each time. Examples work best for questions that come up repeatedly. They help less with one-off questions that look nothing like the examples.

5. Test with benchmarks

Benchmarks are a set of test questions with known correct answers. Run them to measure how often the agent gets things right, then adjust metadata and examples where it fails. Curation is ongoing work. Many teams underestimate how much analyst time an agent needs after launch, especially as tables and business definitions change.

6. Share the agent

Share the agent with the users or groups who need it. The agent runs queries using the warehouse credentials of its creator, while data access still follows each user’s Unity Catalog permissions.

How accurate is Databricks Genie?

Databricks Genie is accurate on well-prepared data, and Databricks’ own testing puts its first-answer accuracy at 84.5%. That figure comes from an internal benchmark Databricks ran in June 2026 on 28 real enterprise data analysis tasks. In the same test, the strongest general-purpose coding agent scored 52.4% and the weakest scored 25%. Databricks also reported that Genie answered about twice as fast as the strongest of those agents.

These numbers need some context. The benchmark was designed and run by Databricks, and it covered a small set of tasks. It has not been independently audited. Databricks’ own documentation also describes Genie as nondeterministic, which means the same question can sometimes produce a different query or answer.

Accuracy in practice depends mostly on how the agent was set up. The factors that make the biggest difference are covered in the setup steps above.

• Metadata quality: Clear table and column descriptions help Genie choose the right fields.

• Scope: Agents with fewer tables and columns make fewer wrong picks.

• Example SQL: Genie selects relevant examples from the agent’s query library when it writes new queries.

• Business definitions: Terms defined once as SQL expressions are applied consistently across answers.

The most reliable way to judge Genie is to test it on your own data. Benchmarks let you run a fixed set of questions and track the accuracy score as you refine the agent. Databricks also offers an Inspect feature, currently in public preview, that helps analysts review and improve individual responses. Teams that skip this testing step usually find accuracy problems only after business users have started relying on the answers.

Databricks Genie pricing

Databricks Genie uses pay-as-you-go pricing, billed in Databricks Units (DBUs) on top of a free monthly allowance for each user. Databricks started charging for usage above that allowance in July 2026. One shared price covers Genie One, Genie Agents and Genie Code, and usage through the UI and the API draws from the same pool.

Each user receives 150 free DBUs per month, which resets on the first of each month. Databricks is also running promotions on top of this allowance until January 31, 2027.

At US East rates, 150 DBUs works out to roughly $10.50 per user each month. The dollar value of a DBU varies by cloud, region and tier, so the same usage can cost different amounts across workspaces.

Two details catch many teams off guard when they plan a Genie budget.

• Compute is billed separately. The DBU charge covers only the AI usage. Every query Genie runs also uses a SQL warehouse, which is billed under standard Databricks compute pricing.

• Service principals get no free allowance. Scheduled jobs, apps and automated agents that call Genie through a service principal are billed from the first DBU. Teams that embed Genie in other applications should account for this early.

Account admins can set budgets for Genie usage and track consumption through the Databricks billing system tables. These tables show free and billed usage separately for each user. Genie is often a small part of the overall bill, and teams estimating a full move to the platform will find the larger cost drivers in our breakdown of Databricks migration cost.

The promotional dates have already been extended once. Check the Databricks pricing page for current rates before you finalize a budget.

Databricks Genie vs Snowflake Cortex

Databricks Genie and Snowflake Cortex Analyst do the same core job, and the right choice usually depends on where your data already lives. Both turn plain-English questions into governed SQL. Genie works on data in the Databricks lakehouse under Unity Catalog. Cortex Analyst works on data stored in Snowflake. Neither tool is built to query the other platform’s data directly.

The setup work looks different on each platform, but the effort is similar. Genie asks analysts to curate each agent with descriptions and examples. Cortex Analyst asks them to define metrics, dimensions and relationships in a semantic model upfront. In both cases, answer quality depends far more on this preparation than on the underlying model.

The pricing models differ in how predictable they are. Cortex Analyst charges a set amount per message when used on its own, so a repeated question costs the same each time. Genie’s DBU consumption varies with the work each question requires. That makes Genie harder to forecast, although the free allowance covers light usage for most individual users.

Snowflake’s own direction is also worth noting. Snowflake now recommends that teams move toward Cortex Agents, which wrap Cortex Analyst inside a broader agent framework. Teams evaluating Cortex today should look at both.

For organizations that run both platforms, using Genie for lakehouse data and Cortex Analyst for Snowflake data is a common arrangement. For everyone else, the choice between these two tools follows the broader Databricks vs Snowflake decision, since each one only works on its own platform’s data.

Conclusion

Databricks Genie is a practical way to give business users self-service access to data that already sits in a Databricks lakehouse. Genie Agents handle most of that work, and their answers are only as reliable as the tables, descriptions and examples analysts give them. The pricing is manageable for most teams while the promotional period runs, though compute and service principal usage need planning from the start.

Teams on Snowflake can get a similar result from Cortex Analyst, provided they invest in the semantic model. For Databricks teams, the deciding factor is the state of their data. Incomplete Unity Catalog metadata and undocumented tables will show up quickly in Genie’s answers. Addressing those gaps through AI readiness and data foundation work before launch gives your first Genie Agent a much better chance of earning business users’ trust.