Complete your Databricks User Groups profile!

Fill out a few details about yourself so the community can get to know you.
Genie Community

When we started, the instinct was one big Genie Room. I steered us somewhere more deliberate. 🧞

Summary: Valeria Kozel shared their experience of designing Genie Rooms for a finance function, emphasizing the need for separate, purpose-driven rooms: narrow rooms for trusted answers and a blended room for exploration. Lingeshwaran Kanniappan suggested using a 'system of record vs system of discovery' approach and highlighted the importance of setting expectations for each room. Grazia Frontoso related this structure to Data Mesh principles, advocating for verifying data in trusted rooms before exploring. Matthias Mueller questioned whether this approach indicates a preference against holistic platforms like Databricks One.
AI Summary

I was helping a UK enterprise Finance function stand up Genie. The easy path was one room, all the data, ask it anything. Instead we designed for two different jobs:

Narrow rooms for trusted answers. Two tightly-scoped rooms — one operational, one financial — each doing one thing exceptionally well. There's a number behind this: a Genie Room performs best at around 7 tables. It'll cope with 30. It won't work well with 100. Fewer, well-chosen tables mean sharper answers people actually believe — and trust is the whole game.

One blended room for exploration. Then, deliberately, a room that brings finance, operations and HR together — SAP, SAP Analytics Cloud and Workday in one place — to answer questions no single source can. Think cost per head against budget by function, or whether headcount growth is actually tracking financial performance. That's the beauty of the platform: the data was never designed to sit together, and here it just does.

The lesson: scope isn't one rule, it's a design choice per question. Narrow when people need to trust the answer. Broad when they need to explore what's possible. Same platform, two very different jobs.

And access shaped all of it — HR and financial data don't share an audience, so room boundaries were drawn around who should see what from day one, not retrofitted later.

Curious how others balance the tight, trusted rooms against the broad, exploratory ones? 👇

#Databricks #Genie #DataGovernance

6 comments
Lingeshwaran Kanniappan
1 month ago

Great article @Valeria Kozel. I’d balance them by making the narrow rooms the system of record, and the broad room the system of discovery.

We can also start by giving different expectations to different rooms. For example, in Trusted rooms, the expectations could be "This should return consistent answers to known business questions”, and when it is an exploratory room, "This is where we discover relationships and new questions worth productizing."

Valeria Kozel
4 weeks, 1 day ago


Lingeshwaran, "system of record vs system of discovery" is a much cleaner way to say what I am circling around — I'm definitely going to use it in future 🙂

The expectation-setting point is the crucial one. We found that when users don't know which kind of room they're in — they treat a discovery answer as gospel, or distrust a record room because a discovery room gave them something fuzzy before. So we started explicitly labelling them with detailed introduction message to signal the room's "job."

Grazia Frontoso
1 month ago

Thanks Valeria for sharing your thoughts!

Your approach sounds like applying Data Mesh principles to AI rooms. By treating data as a product with distinct domains (operational vs. financial), you ensure ownership and high fidelity.

We’ve seen success using the narrow rooms as the foundational "building blocks." Once those narrow domain rooms are stable and trusted, we treat the blended rooms somehow as an aggregation layer. Essentially, you don't let anyone into the exploratory playground until they've verified the data in the trusted sandbox first. This is important to ensure the first experience is at high standards.

Valeria Kozel
4 weeks, 1 day ago

Thank you, Grazia — this is the kind of comment that makes posting worth it. I hadn't consciously connected it to Data Mesh, but you're absolutely right: the whole thing comes down to treating each room as a product with a clear owner, and the table-count discipline is just a symptom of that.

I really like "verify in the trusted sandbox first" — it keeps the discovery room useful without letting it become a source of truth by accident. We got there informally; making it an explicit habit for users is the bit I was missing.

Matthias Mueller
3 weeks ago

Valeria, sounds like you are not a fan of Databricks One because of its holistic approach?

Valeria Kozel
3 weeks ago

Hi Matthias,

Quite the opposite — overall fan of Databricks One.

However, holistic access and scoped rooms aren't opposites:

  • Databricks One is about the holistic experience — one place, simple access, everyone in.

  • Genie Room (or now Genie Agent) scoping is about what sits behind each answer.

You can definitely have a unified front door (Databricks One) and still have well-scoped, well-owned rooms feeding it. Holistic access, disciplined design — they are not competing, they compliment and reinforce each other.

Valeria Kozel
1 month ago