Analysing and serving · Chapter 21

Databricks Genie

Genie is the part of Databricks that business people actually see. They type a question in English, it writes the SQL, runs it against governed data and answers with a number, a table or a chart. Everyone wants it after the first demo. Whether they still use it a month later depends on work that happens before anybody asks it anything, and that work is what this chapter is about.

What Genie is now #

The names moved at the Data + AI Summit on 16 June 2026, and most of what you'll read online predates that. If an article says "Genie space", it means what the documentation now calls a Genie Agent.

Genie One is the front door for business users: one place to ask data questions, open AI/BI dashboards and run Databricks Apps. It also lives inside Slack and Microsoft Teams, has iOS and Android apps, and can run schedules and alerts. Databricks calls it "the data-smart AI coworker for business users".

Genie Agents are the subject-area environments behind it. A data team configures one per domain (sales, supply chain, claims) with the tables, metric definitions and business rules for that domain, and Genie One draws its answers from them. This is the thing you build, and the rest of this chapter is about building it.

Genie Code is the assistant for developers and analysts inside the workspace, writing notebook code and SQL. It's a different audience with a different bill, covered under cost below.

Genie Ontology is newer and works in the background. It reads tables, queries, dashboards and pipelines and pulls out how your company defines things, weighting each definition by where it came from. Databricks compares the approach to PageRank. Treat it as a helpful starting point, not as a replacement for the definitions you write down deliberately.

The renames are collected with the others in What Everything Used to Be Called.

How a Genie Agent answers a question #

When somebody asks a question, Genie assembles context from five places and writes SQL from it: the Unity Catalog metadata for the agent's tables, the example queries you have supplied, the knowledge store, your instructions, and the conversation so far. The SQL then runs on the agent's SQL warehouse.

Two things in that picture decide quality.

Trusted assets are the answers you have already checked. They are parameterised example queries and SQL functions whose logic an author has verified, and when Genie uses one, the documentation says "the answer comes from this verified logic." For the ten questions people ask most, you want Genie using a trusted asset rather than improvising.

Metric views carry the definitions. A Genie Agent can read tables, views and metric views, and a metric view is where revenue, margin and churn get one definition that dashboards, SQL and Genie all share. Reporting, Semantics, and Power BI makes the case for putting definitions there. Genie is the strongest argument for it, because a language model asked to calculate gross margin from raw tables will produce a plausible formula, and plausible is not the same as yours.

Who sees what #

The governance model is good, and it has one detail worth understanding before you share anything.

Data access is "always evaluated using each end user's own Unity Catalog permissions." The row filters and column masks you built in Unity Catalog apply automatically, so a regional manager asking about revenue sees their own region's revenue, from the same Genie Agent their colleague in another region uses.

Compute is different. When an author picks the agent's warehouse, their compute credentials are embedded so that every user can run queries without their own warehouse permission. Users borrow the author's compute and keep their own data access. If a different author changes the warehouse later, that person's credentials are embedded instead. In practice this means the warehouse belongs to whoever last saved it, so pick it deliberately and give the agent a warehouse whose cost is tagged to the team that owns the domain.

Building one people trust #

Databricks' own best-practice guide and our experience agree on the order, and the order matters more than any single step.

Start with one narrow subject. Databricks advises picking "a single, high-value use case, such as a specific sales dashboard or an operational report." An agent can hold up to 50 tables, views or metric views. That is a ceiling, not a target. Five well described tables beat forty undocumented ones every time.

Write the test questions before you configure anything. Collect thirty to fifty real questions from the people who will use it, each with the SQL that produces the agreed answer. These become benchmarks (below). Builders write questions the agent can answer. Users write the ones with an ambiguous date range, the one about data that isn't there, and the one that's really two questions.

Describe the columns. Table and column comments in Unity Catalog are the cheapest improvement available. A column called amt_2 with no comment will be guessed at.

Put the definitions in metric views and the hard queries in trusted assets. Anything with a business rule in it, such as how refunds net against revenue or which orders count as shipped, belongs in a metric view or a verified function, so Genie reuses it instead of reinventing it.

Curate the knowledge store. It holds agent-level descriptions, synonyms, join relationships, SQL expressions and prompt-matching settings, and changing it doesn't alter your Unity Catalog metadata. Synonyms are where most of the early wins come from: your users say "bookings" and the column says order_value.

Use instructions last. Databricks' guidance is blunt: "Use general instructions only when none of the specific tools apply." A paragraph of instructions is the least reliable lever, because it's prose a model interprets, where a trusted asset is SQL it runs.

Then run the benchmarks, look at what failed, fix the context, and run them again. Databricks' guide says to expect failures at first, and they're useful, because each one points at missing context.

Testing it, and keeping it honest #

Benchmarks are built in. Each Genie Agent holds up to 500 benchmark questions. In chat mode, Genie's answer counts as good when its SQL returns the same result as yours, including the same rows in a different order, or numbers that agree to four significant digits. Up to 5,000 rows of each result are compared. In agent mode an LLM judge grades the answer instead, optionally guided by your notes.

After launch, every answer carries an "Is this correct?" prompt with three responses: Yes, Fix it, and Request review. The monitoring tab lists every question asked, filterable by rating and user. Read it weekly for the first month. The questions people ask that you didn't anticipate are your next benchmarks, and the ones marked "Fix it" are your next trusted assets.

Agent mode, for questions that need more than one query #

Standard chat answers a question with a query. Agent mode is for exploratory questions, such as why margin fell in the Midwest last quarter. It "creates and refines a research plan, runs multiple SQL queries, learns from each result, and iterates" until it can write a report with citations, charts and supporting tables. It can also read files that authors attach from Unity Catalog volumes, so a pricing memo or a contract can sit beside the tables.

It's generally available in the Americas, Europe, Australia, New Zealand and Japan. It is also slower and runs more SQL per question, so it costs more warehouse time per answer. Use it for the analysis a person would otherwise spend an afternoon on, not for "what were sales yesterday".

What it costs #

Three separate things, and people usually only think about the first.

Genie use by people. Genie One and Genie Agents are free for users through 31 January 2027. Service principals are excluded from that promotion and are charged, which matters the moment you call Genie from an application or a pipeline rather than from a person.

Genie Code. Since 8 July 2026 it bills pay-as-you-go, with a free monthly allowance per user, so a developer who uses it lightly may never see a charge.

The warehouse. Every query Genie writes runs on the agent's SQL warehouse, and that bills exactly as it would for an analyst. Genie Agents need a pro or serverless warehouse, and Databricks recommends serverless. The smallest serverless warehouse uses 4 DBUs an hour, about $2.80 at the AWS Premium list price, so a busy agent on a small warehouse that shuts down when idle is cheap. One left running all day is the expensive mistake What It Actually Costs describes.

A Genie Agent that answers four hundred questions a day instead of ten is a success and also a larger warehouse bill. Budget for the success.

Putting Genie in other places #

The Conversation API lets your own application, chatbot or agent ask a Genie Agent questions, authenticated as a user through OAuth or as a service principal. Answers are asynchronous, so the client polls every one to five seconds and backs off on failures.

A Databricks App can declare a Genie Agent as a resource, which is the cleanest way to put conversational analytics inside an internal tool. Agent Bricks' Supervisor Agent can use Genie Agents as subagents alongside other tools, covered in Orchestrating Agents. Each of these is a way to reach the same governed agent, so the curation and benchmarks you did once carry everywhere it's used.

How it compares #

The real difference is where the definitions live. Every tool on that row needs curated meaning to answer well. On Databricks the meaning sits in Unity Catalog, where dashboards, SQL, notebooks and agents read the same definitions. In the others it sits inside the analytics tool, and only that tool understands it.

When not to use it #

When the numbers aren't settled. If finance and sales disagree on revenue, Genie will pick one of their definitions and answer confidently with it. Fix the definition first, in a metric view, and add Genie afterwards.

When the question is the same every day. A dashboard answers "what were sales yesterday" faster, more cheaply and with no chance of a different interpretation tomorrow.

When nobody owns it. A Genie Agent needs somebody who reads the feedback, adds benchmarks and fixes what users flag. Without that owner the agent drifts as the tables change, and people quietly stop asking it things.

If you'd like a second pair of eyes on a Genie rollout, working with TechFabric describes how we help.

Table of Contents