Skip to content

Databricks AI/BI dashboards and Genie

Preetham ReddyUpdated 7 min read

Title graphic reading 'Business Intelligence Meets AI: Self-service analytics that finally work', with illustrated chat bubbles containing a chart.

Databricks AI/BI is the business intelligence built into the Databricks platform. It has two parts. AI/BI dashboards are low-code dashboards that run on a SQL warehouse against tables governed in Unity Catalog. Genie lets people ask questions in plain English and answers by writing and running SQL, showing the query alongside the result. Together they're Databricks' answer to self-service analytics: people across the business answering their own questions from the same governed data the data team uses, without exporting it to another tool first.

Updated by the TechFabric editors in October 2026. This post first ran in July 2025, shortly after Genie reached general availability on all clouds at Data + AI Summit in June 2025; dashboards had been GA on AWS and Azure since June 2024. The product details below were checked against Databricks' documentation on 11 October 2026.

What's changed since 2025

  • Genie spaces are now Genie Agents. Databricks announced the change in June 2026 and the docs switched over on 9 July 2026. The thing you configure, a set of tables plus instructions and examples that Genie answers from, is now an agent.
  • Genie One is the business-user front door. Databricks One was renamed Genie in April 2026 and Genie One in June, and it's where people outside the data team find dashboards and ask questions.
  • "AI/BI" now mostly means the dashboards. The docs still say AI/BI dashboards, and the release notes are titled "AI/BI and Genie One".
  • Metric views are generally available. Databricks announced Unity Catalog Business Semantics as GA on 2 April 2026, with metric views at its heart. They're the most useful addition for anyone trying to make Genie's numbers agree with finance's, and there's more on them below.
  • More room. A Genie Agent can now include up to 50 tables, views or metric views, raised from 30 on 10 September 2026.

Why this is different from a BI tool bolted on top

One copy of the logic, and one set of permissions.

In a traditional setup, data is modelled in the warehouse, extracted or connected into a BI tool, and the business logic gets rebuilt there. Two copies of the logic drift apart, and the governance in the warehouse doesn't follow the data into the tool.

AI/BI runs inside the platform, so Unity Catalog's permissions, row filters and column masks apply to what people see. For Genie, Databricks says "Unity Catalog evaluates data access as the end user's own identity, not the author's", while the query runs on a warehouse using the author's compute credentials. So a sales manager asking about revenue sees only the regions their row filter allows, without anyone building a separate security model in a BI tool.

Dashboards give the publisher a choice. "Share data permissions", the default, runs viewers' queries with the publisher's data access. "Individual data permissions" runs them with each viewer's own. If your tables carry row-level security, pick the second, or every viewer sees what the publisher sees.

Databricks AI/BI dashboard 'Sales Opportunity Overview' with a Launch Genie button, KPI tiles for days to close, close probability and opportunity size, a pie chart by stage, region bars and a scatter plot over time.

What makes Genie accurate

Genie's answers are only as good as the semantics it's given. Pointing it at fifty raw tables and inviting the company in is the fastest way to lose people's trust in week one. Databricks' own guidance is blunt: "Aim for five or fewer tables" to start, and "use text instructions only as a last resort." In order of how much they help:

  1. Clean, described tables. Column descriptions in Unity Catalog, plus synonyms set in the agent so "revenue", "sales" and "turnover" all land on the right column.
  2. Join relationships declared explicitly (many to one, one to many, one to one) so Genie doesn't guess at keys.
  3. SQL expressions for the business definitions people argue about. Databricks has three kinds, filters ("active customer"), measures ("net revenue") and fields.
  4. Example SQL queries, ideally parameterised, for the questions people ask most.
  5. Trusted assets. Databricks defines these as "example SQL queries and SQL functions that provide verified answers." When Genie answers with one, the response is marked as verified, which tells the reader a person signed off on the logic. That's the "certified answers" idea from the original version of this post, and it's what makes a CFO willing to use the answer.
  6. Prompt matching, on by default, which samples values from columns so a question mentioning "Phoenix" or "Q3" maps onto what's actually stored.

There are limits worth knowing. An agent's knowledge store holds 200 snippets. Instructions are capped at 100, and every example query and function counts against that cap. And Genie cancels any query execution that runs longer than 15 minutes, so a question over an unoptimised fact table can simply time out.

Metric views: one definition of revenue

A metric view defines measures and the fields you can slice them by once, in Unity Catalog, and dashboards, Genie Agents, notebooks, the SQL editor and alerts all read the same definition. Here's a minimal one, from Databricks' documentation, on the TPC-H sample data that ships with every workspace:

CREATE OR REPLACE VIEW orders_metric_view WITH METRICS LANGUAGE YAML AS
$$
  version: 1.1
  source: samples.tpch.orders
  fields:
    - name: Order Month
      expr: DATE_TRUNC('MONTH', source.o_orderdate)
  measures:
    - name: Total Revenue
      expr: SUM(source.o_totalprice)
$$

Metric views arrived in Databricks Runtime 16.4, but several features of the 1.1 YAML specification need newer runtimes. Agent metadata such as synonyms and display names needs 17.3, and one-to-many joins and window offsets need 18.1, so run definitions like this on a current runtime or SQL warehouse. You'll also see examples that use dimensions: instead of fields:. Databricks treats the two keywords as equivalent, and definitions written either way still work.

The reason this matters for Genie is that a measure is defined in one place. When someone asks "revenue by month for the West region", Genie aggregates Total Revenue the way finance defined it, rather than writing its own SUM over whichever column looked closest.

no metric view

metric view

Question: revenue by month for the West region

Genie

Writes its own SUM over the closest column

Aggregates Total Revenue as finance defined it

With a metric view, Genie aggregates Total Revenue the way finance defined it instead of writing its own SUM over whichever column looked closest.

Measuring whether it's right

Genie Agents have benchmarks, a set of test questions with expected answers, up to 500 per agent. Run them after every change to instructions or tables and you'll know whether the change helped before users find out it didn't. Write the questions people actually ask, including the ones with traps, such as fiscal versus calendar quarters.

In use, each answer asks "Is this correct?" with Yes, Fix it, or Request review, which flags the response for someone on the data team. The Monitor tab lists every question asked, filterable by user, rating and status. Feedback doesn't retrain anything. The docs say outright that an agent's "behavior does not change based on user feedback alone", so somebody has to read the reviews and turn them into instructions, examples or trusted assets. That loop is the real operating cost of self-service analytics.

What it costs

Genie Agents need a pro or serverless SQL warehouse, and Databricks recommends serverless. When Genie went GA in 2025 there was no separate licence for AI/BI. In 2026 the docs say Genie One and Genie Agents usage by users "is free through January 31, 2027", excluding service principals. The warehouse running the queries is billed as usual, so the cost to watch is warehouse time. A poorly modelled table that makes every question scan a year of data costs more than one with a metric view over a well-partitioned table.

Putting Genie inside other applications

The Conversation API lets an application start a conversation with a Genie Agent and poll for the answer. Start one with POST /api/2.0/genie/spaces/{space_id}/start-conversation, then fetch messages and query results. The path still says spaces despite the rename. Databricks' guidance is to poll every one to five seconds for up to ten minutes, with your own back-off. That's how Genie gets into Teams, Slack or a customer-facing portal.

A retail proof of value

Recently, we conducted a Proof of Value for a key client in the retail sector. By using Databricks AI/BI, we successfully demonstrated how their sales and operations teams could gain instant, natural-language answers to complex inventory and supply chain questions, drastically reducing the time spent on manual reporting and speeding up the decisions that follow. The client was impressed by the speed of insight and the intuitive interface that eliminated the need for specialized data knowledge.

Where to start

  1. Pick one subject area with a few well-understood tables, such as orders or inventory.
  2. Write ten real questions from the people who'll use it, with the right answer next to each. Those are your first benchmarks.
  3. Build a metric view for the two or three measures people argue about.
  4. Create a Genie Agent over five or fewer tables, add join relationships and example queries, and run the benchmarks.
  5. Open it to a small group, and read every "Request review" for the first month.

TechFabric is a Databricks partner, and making Genie answers trustworthy enough for a finance team is the focus of our Genie accuracy work. Our engineer Ivan Vydrin placed second in Databricks' Genie-Powered App Challenge in September 2026. If you'd like to run a proof of value on your own data, talk to us.

Frequently asked questions

What happens to a Genie answer when the question runs past the 15-minute limit?

Genie cancels the query, so the person asking gets no answer, not a slow one. That usually means a fact table the warehouse has to scan in full, and the fix sits in the tables, not in Genie. A metric view over a well-partitioned table is where we'd look first, and it's also cheaper to run.

What if I run out of room for instructions?

Start by cutting what the agent holds, since the guidance is five or fewer tables and text instructions only as a last resort. Every example query and function counts against the 100-instruction cap. Moving the definitions people argue about into a metric view defines each measure once in Unity Catalog, so Genie aggregates it the way finance defined it instead of needing instructions to steer it.

Why does my dashboard show viewers more than Genie shows the same people?

It's probably the sharing setting. A dashboard published with the default "Share data permissions" runs every viewer's queries with the publisher's access, while Genie evaluates access as the end user. If the tables carry row filters, switching the dashboard to "Individual data permissions" brings the two into line.

Sources

Checked 11 October 2026.