What Is Databricks Genie, and Should You Be Using It? (Webinar Recap)

TL;DR

• Genie isn't one Databricks feature. It's three separate products, Genie One, Genie Agents, and Genie Code, all sitting on top of a shared intelligence layer called Genie Ontology.
• It used to be free. As of last July, Databricks started charging for Genie Code, with Genie One and Genie Agents expected to follow eventually.
• The AI cost is easy to track. The compute cost Genie triggers is not, and that's usually the bigger number.
• Genie budgets can cap the AI side of the bill. Controlling the compute side takes a dedicated, properly tagged SQL warehouse.

Genie Is Three Products, Not One

Genie has become an umbrella term, and Databricks hasn't made that easy to follow. As Olivier put it, the naming has changed enough times that even people who've worked with Databricks for years find it confusing. Here's what actually sits under the Genie name today.

Genie One

The entry point for business users. It's a simplified chat interface, similar to using ChatGPT or Claude, except it's connected directly to your data: your tables, your metadata, and any agents you've built. It also works outside Databricks itself, it can be wired into Slack, Teams, or spreadsheets through Genie's API, so someone can ask a question like "how many sales did we have last week?" without ever opening Databricks.

Genie Agents

Genie has access to every table in your catalog by default, but that's not always enough to answer a specific question well. Genie Agents solve that by letting you configure a resource with real domain knowledge: which tables to use for a given topic, plain-language instructions, and example calculations or metrics. Genie One calls on these agents automatically when a question falls into their area, without you needing to specify which agent to use.

Agents can also be scoped to run under the credentials of whoever's asking the question, meaning existing row-level security and column masking still apply, or under a shared credential if you want the agent to have broader access.

Genie Code

The developer-facing piece. It runs SQL and Python, and can build actual Databricks resources: jobs, pipelines, dashboards, and even other Genie Agents. In the demo, Olivier didn't hand-write an agent's configuration, he used Genie Code to generate it, which asked him what information it needed and then formatted everything correctly.

The Layer Underneath: Genie Ontology

Genie Ontology isn't something you interact with directly, it's the intelligence layer that makes the other three products work. Built on top of Unity Catalog, it automatically extracts business definitions, maps relationships between tables, parses your query history to understand what's actually being used and how, and reads existing metadata, comments, and governance rules you've already set up. This is what lets Genie answer questions about your specific business instead of giving generic, disconnected answers.

A Quick Example: Teaching Genie About a Stock Portfolio

To show how these pieces connect, Olivier ran through a simple demo. He had a Unity Catalog table with historical stock prices, and asked Genie One a straightforward question: how did the value of my portfolio change over the last two years?

Genie One couldn't answer it. It could see the stock price table, but it had no idea what Olivier actually owned, that information doesn't exist in Databricks at all.

So he built a Genie Agent to fix that. The agent's instructions included the actual shares held (in the demo, roughly $100 of Apple and $200 of Microsoft, purchased two years earlier) and a plain-language explanation of how to calculate portfolio value from that. He also pointed the agent at the right source tables.

After that, he went back to Genie One and asked the exact same question again, word for word. This time, instead of saying it didn't have enough information, Genie produced a chart of portfolio value over time, total returns, gains, and key moments in the holdings. Genie One had automatically recognized which agent was relevant and called it, without anyone telling it to.

This same setup also supports things like unit conversion. Olivier gave an example of adding an instruction to convert prices from USD to CAD, or timestamps to UTC, directly in the agent's configuration, without changing anything in the underlying tables.

A few other things came up around this demo:

• Agents can hand off to each other. If a question spans multiple topics, Genie One can call more than one agent, wait for each answer, and combine them into a single response.
• Agent definitions can be exported as JSON or YAML and deployed through Databricks Asset Bundles, Databricks' simplified, Terraform-like deployment tool. Native Terraform support for agents specifically isn't available yet.
• Agents can be evaluated like code, using MLflow to define test questions, track expected versus actual answers over time, and monitor for drift once an agent is in production.

Olivier also shared a real example from his own work: a monitoring setup pulling from several agents (workflow failures, cost data, data quality checks) to generate daily reports automatically, plus the ability to ask follow-up questions like "why do I have duplicates?" and get a specific, useful answer back.

What Genie Actually Costs

Genie used to be free. That changed last July. Databricks originally planned to start charging for all three products at once, but pulled back to give teams time to adjust. As of this webinar, only Genie Code is billed, on a per-usage basis, though Genie One and Genie Agents are expected to be metered eventually too.

Every user gets 150 free DBUs per month (roughly $7 to $8, depending on cloud provider and region) before charges kick in. One exception worth knowing: service principals don't get this free allowance at all. Automated usage is billed from the very first request.

The cost itself splits into two very different pieces:

1. LLM usage. This is straightforward to track. It's metered in DBUs and shows up in your system tables, you can query it directly and filter by product type to see exactly what Genie's AI usage is costing you.
2. Compute. This is the hard part. Running Genie Code (or any agent) means running actual SQL and Python underneath, and that compute is billed as ordinary Databricks compute. There's no system table column that says "this warehouse activity was Genie." It's invisible unless you go out of your way to isolate it.

In the portfolio demo, the total cost came to around $5. The LLM portion was billed at $0 because it fell inside the free monthly allowance, though it would have cost roughly $4 on its own if it hadn't. Olivier was only able to separate LLM cost from compute cost because it was his personal workspace with no other activity running in that time window, in a shared production environment, that kind of clean separation isn't realistic without extra setup.