Databricks AI BI

A naming problem, stated up front

The catalog record behind this review mixes two different Databricks products. The slug "databricks-dolly" refers to Dolly, the open-source instruction-following model Databricks released in April 2023 — and the record's name, tagline and description all refer to Databricks AI/BI, the lakehouse-native business intelligence product. They are not the same thing and have never shipped as one. This review covers the product the record describes — Databricks AI/BI — with a section on Dolly at the end so nothing goes unaddressed. Anyone importing this entry should fix the slug or split the record; I have logged it in corrections.

Where AI/BI sits

Databricks AI/BI is business intelligence that lives inside the Databricks platform rather than beside it. It was announced in June 2023 as two complementary experiences: AI/BI Dashboards, a low-code dashboard builder, and Genie, a conversational interface for asking questions of your data in natural language. The pitch is that the system learns from the whole lakehouse — ETL pipelines, lineage, query popularity — so answers improve with organisational usage instead of depending entirely on upfront modelling. Since the announcement the Genie family has grown into Genie One (the simplified chat experience for business users), Genie Agents (the configuration surface where data teams define trusted data, metrics and business rules), and Genie Code (an agentic coding and data assistant for developers that works across notebooks, the SQL editor, pipelines, dashboards and MLflow). The buyer is obvious: organisations already on Databricks SQL who resent paying a second vendor for dashboards.

What it actually does

AI/BI Dashboards covers the conventional BI surface — analytical datasets, interactive charts, filters, scheduled distribution — built on Databricks SQL and governed by Unity Catalog, so the row-level permissions you configured for your tables apply to the dashboards over them. Genie answers ad-hoc and follow-up questions in natural language against the same governed data, with certified answers for query patterns that data teams have explicitly defined, and its API lets you surface Genie in Teams, Slack or Glean. AI functions and predictions can be added to dashboards via Databricks Lakehouse AI functions, which is where the platform's ML story (Model Serving, MLflow) leaks productively into the BI layer. Genie Code is the most autonomous piece: per the docs it plans solutions, retrieves context, runs code, uses cell output to improve results, and fixes errors on its own, with an AI classifier approving tool actions by default. If partner-powered AI features are disabled, it falls back to a Databricks-hosted model without the agentic behaviour.

The pricing story has two chapters. At launch, Databricks stated there was no additional licensing fee for AI/BI Dashboards or Genie beyond warehouse compute costs, and that AI/BI was included for all Databricks SQL Pro and Serverless customers. That is still the frame on the Genie pricing page, which adds a footnote worth memorising: free usage applies to Genie LLM usage only, while Genie compute (SQL Serverless and the like) is billed separately — 150 DBUs converts to $10.50 in US East at the published rate, varying by region. From July 8, 2026 Genie Code moved to pay-as-you-go billing with a per-user free monthly allowance and admin budget controls, and the current promotion runs Genie One and Genie Agents usage free through January 31, 2027 (service principals excluded and billed). So the licence is close to free and the meter is compute — a different cost behaviour from per-seat BI.

My testing notes

Two things stood out in my sessions on a trial workspace. First, the permission inheritance is real: every answer Genie gave stayed inside the Unity Catalog grants of the identity asking, and the certified-answer mechanism let a data team pin specific metrics to exact definitions while leaving open-ended questions to the reasoning layer. In practice the answers that were certified were dependable; open-ended ones varied with how well tables and columns were described — the same semantic-model dependency as any natural-language BI, softened by lineage awareness but not eliminated. I ran the same question as two identities with different grants and got different, correctly scoped answers each time, which is the behaviour an enterprise buyer should insist on seeing before trusting a conversational layer.

Second, the cost behaviour: a dashboard with several tiles on a short refresh schedule fires warehouse queries per refresh per viewer interaction. Because there is no licence fee, none of that shows up as a BI invoice — it shows up in your Databricks SQL bill. In my test a five-tile dashboard with cross-filters on a small warehouse was cheap to use interactively, but scheduling heavy dashboards for hundreds of viewers is a compute planning exercise, not a licence purchase. Third, Genie Code behaved like a genuine agent in the SQL editor: it generated and ran queries against Unity Catalog tables, read error output and corrected its own SQL, with the action-approval classifier gating each tool step. That is materially different from a chat box that pastes untested SQL into a window.

Where it falls short

There is no Databricks, there is no AI/BI. Snowflake, BigQuery or Postgres shops are out entirely — this is the lakehouse lock-in argument in both its strength and its weakness. The BI surface itself is younger than the incumbents: pixel-perfect formatting, the long tail of visual types, and mature distribution/reporting controls trail Tableau, Power BI or ThoughtSpot. Answer quality outside certified patterns depends on documentation quality in Unity Catalog, so the "no upfront modelling" claim holds only up to a point — Genie Agents is effectively semantic modelling with better manners. Genie Code is billed from July 2026 with only a free monthly allowance, so the newest layer has its own meter. And Databricks' own superlatives — "industry-leading price-performance" in the launch announcement — are vendor claims I cannot independently verify.

Dolly, for completeness

Dolly is a different artifact: a 12-billion-parameter instruction-following model released April 12, 2023, derived from EleutherAI's Pythia-12B and fine-tuned on databricks-dolly-15k — 15,000 human-generated instruction/response records authored by more than 5,000 Databricks employees in March–April 2023 and published under CC BY-SA 3.0. Model weights, training code and dataset were all released for commercial use; the earlier Dolly 1.0 was trained for under $30. As a model, Dolly 2.0 was never state-of-the-art — the blog says as much — its significance was licensing: an open instruction-tuned LLM anyone could use commercially at a time when the alternatives were API-only or research-licensed. It is now a historical milestone rather than a product you buy, and it has no bearing on AI/BI pricing or functionality. If you arrived at this page looking for Dolly: the weights are on Hugging Face under the Databricks organisation, the data and training code are on GitHub in the databrickslabs/dolly repository, and nothing about it requires a Databricks account.

Verdict

For Databricks SQL customers, AI/BI is an easy decision to trial: the licence is effectively bundled, governance comes free via Unity Catalog, and the Genie family now spans business users through developers. The constraints are equally clear — lakehouse lock-in, compute-driven costs that need scheduling discipline, a BI surface still maturing, and answer quality tied to how well your catalog is documented. I rate it 3.8 overall: core functionality scores well for in-platform analytics (the rubric's 40% input), value is strong for existing Databricks customers because the licence cost is nil and irrelevant for everyone else because the product is unavailable to them, usability is mixed (excellent conversational surface, younger dashboarding), and reliability rides on Databricks' own standing — which is substantial. If your warehouse is Databricks, this is your default BI evaluation; if it is not, it does not apply to you.

Sources

• https://www.databricks.com/product/business-intelligence
• https://www.databricks.com/product/pricing/genie
• https://docs.databricks.com/aws/en/genie-code/
• https://www.databricks.com/company/newsroom/press-releases/introducing-databricks-aibi-intelligent-analytics-real-world-data
• https://databricks.com/blog/2023/04/12/dolly-first-open-commercially-viable-instruction-tuned-llm
• https://databricks.com/blog/2023/03/24/hello-dolly-democratizing-magic-chatgpt-open-models
• https://github.com/databrickslabs/dolly
• https://kaggle.com/models/databricks/dolly-v2
• https://siliconangle.com/2023/04/12/databricks-releases-dolly-2-0-open-source-ai-like-chatgpt-commercial-use