A few weeks after the Data + AI Summit in San Francisco wrapped up, the dust is starting to settle. 30,000 people were there. Dozens of announcements came out across five days. And now, most data teams are left figuring out what any of it actually means for the work they do on Monday morning.
That’s exactly what Alexandru Puiu, CTO at mindit.io, and Zhanna Pchela, Delivery Solution Architect at Databricks EMEA Central, spent 30 minutes working through in a live webinar. What follows is a summary of the conversation, the announcements they focused on, and what they recommended teams do about it.
Watch the full webinar recording:
The frame: four layers, not six announcements
Before going through the individual product updates, Alexandru offered a way to think about why they’re all coming at the same time. The announcements aren’t random. They’re filling in four layers that, together, make AI usable at the organizational level rather than just at the prototype level.
The first layer is live data. The second is context. The third is agents. The fourth is governance. Every major announcement from the Summit slots into one of these, and the argument is that none of them work without the others.
“Context is the missing piece. It’s what usually sits in people’s heads. It’s the thing that makes AI replies reliable or unreliable.” — Alexandru Puiu, CTO, mindit.io
With that frame in place, the specific announcements start to make more sense.
Lakehouse RT and LTAP: the end of the replication window
For anyone who has ever had to explain to a business user why the data in their dashboard is 24 hours old, Lakehouse RT is the announcement that matters most.
The pitch is sub-100 millisecond query latency directly against operational data, with Unity Catalog governance included. No separate real-time layer. No replication job running at midnight that breaks something every third Tuesday.
The companion announcement is LTAP, which makes Lakebase data available in the analytical interface without pipelines. Alexandru’s summary was direct: no pipelines, no monitoring, no latency. Everything is there right away.
The practical implication is that Lakebase, Databricks’ Postgres-compatible operational database, is now a more serious option for workloads that teams are currently running on standard Postgres. Not because Postgres is bad, but because the gap between operational and analytical has effectively closed.
For teams already running Postgres alongside Databricks, this is worth taking seriously. Migrating some of those workloads into Lakebase removes the 3am ETL monitoring, the five-platform coordination, and the replication window that ops teams have been complaining about for years.
Genie One and Genie Ontology: context as infrastructure
Genie One is the unified interface for everything Databricks has built around natural language access to data. Dashboards, agents, conversations, apps. One place, available on mobile.
Zhanna demonstrated it live during the webinar. The part worth paying attention to wasn’t the interface itself but what powers the answers it returns: Genie Ontology.
Ontology is a self-improving knowledge graph that connects all the data sources in your Databricks environment into a shared semantic layer. It uses a PageRank-style algorithm to determine which pieces of information are authoritative, then feeds that context into the agent loop when someone asks a question.
The result, in theory, is that when a business user asks “how did the EMEA marketing campaign perform last month,” Genie doesn’t have to guess what “EMEA,” “marketing campaign,” or “last month” mean in the context of your specific business. It already knows, because the ontology has connected the glossary definitions, the metric views in Unity Catalog, the table lineage, and the domain structure.
The glossary feature, going into private preview in August, lets domain stewards write business definitions of KPIs and entities. Those definitions are AI-generated, governed in Unity Catalog, and certified by domain owners. Zhanna noted that customers were signing up for the private preview in large numbers, which makes sense given how often semantic layer gaps are the real reason Genie returns inconsistent answers.
Genie Agents: from spaces to domain specialists
Genie Agents are what Genie Spaces have evolved into. They’re domain-specific, they can now work with unstructured documents as well as structured data, and they can be created directly from a Genie One conversation without going into the workspace settings.
The last point matters for adoption. Business users who get a useful result from Genie can now turn that conversation into a persistent agent without needing an engineer to do it.
The recommendation from both speakers was to start narrow. Pick five tables you know well, document them properly, and build from there. Trying to give a Genie Agent your entire data estate on day one is the fastest route to results that don’t hold up.
“You need to get a taste. You need to see how it behaves, where it fails, where you need to adjust. Then it becomes something you can show business users with confidence.” — Zhanna Pchela, Delivery Solution Architect, Databricks EMEA Central
Unity AI Gateway: governance for agent fleets
Unity AI Gateway extends the Unity Catalog governance model to cover AI assets. Model access policies, agent monitoring, guardrails before and after LLM calls, MCP server proxying, cost auditing. All managed in the same place where data governance already lives.
The practical reason this matters is that without it, every team that builds an agent sets its own rules. Or no rules. Unity AI Gateway is how you stop agents from going rogue and spending thousands of euros on tokens before anyone notices.
Zhanna’s recommendation: set up Unity AI Gateway before you need it, not after. The policies you put in place at the beginning are much easier to enforce than the ones you retrofit onto agents that are already in production.
The question everyone asked: how do you prevent data leaks?
During the Q&A, someone asked about data leaking when using AI. It’s the question every IT leader has, and it doesn’t have a one-size-fits-all answer. But the structure of the answer is consistent.
Genie is not an external tool that connects to your data from outside. It runs inside your Databricks environment, which means the security boundary is the same one you’ve already defined. Unity Catalog controls what data is visible to whom, and that applies to AI access the same way it applies to human access.
On top of that, Unity AI Gateway lets you add guardrails at the prompt level and at the response level. If you don’t want the LLM to return certain types of responses, you can enforce that. If you want to block certain inputs before they reach the model, you can do that too.
“Treat AI governance the way you treat data governance. Strict, from the beginning. Minimize the impact before it happens.” — Zhanna Pchela
What to do next
Alexandru ended the session with four specific recommendations for teams already running on Databricks.
1. Deploy Lakebase. It’s production-ready, enterprise-ready, governed in Unity Catalog. If you’re running Postgres workloads alongside Databricks, evaluate which ones make sense to migrate. The combination of Lakebase and LTAP removes a category of operational headache that most data teams have been managing for years.
2. Make your data AI-ready. This means cleaning up Unity Catalog metadata, applying domains and glossary entries, and creating the semantic structure that Ontology needs to work properly. The agents are only as good as the context they have access to.
3. Use Genie to ship agents faster. The tooling now makes it possible to build domain-specific agents without deep technical knowledge. The governance layer means you can do this without losing control. Start with a narrow scope, prove it works, then expand.
4. Extend your governance policies to cover AI. Unity AI Gateway is where this happens. Agent monitoring, guardrails, cost audits. Put the policies in place now, before the agent sprawl starts.
The direction Databricks is moving is clear. The platform is becoming the place where operational data, analytical data, AI agents, and business users all coexist in one governed environment. Whether that vision fully materializes is still being proven out in production deployments, but the pieces announced at the Summit close a lot of the gaps that were previously good reasons to keep things separate.
