Once AI Agents Can Query Data, the Trouble Is Just Beginning

Databricks posted a demo, and the theme is direct: an AI Agent being able to query data does not mean it understands the data.

The interface in the video stops at the Sales domain. On the left are entries such as New chat, Agents, Discover, and My stuff; on the right are terms and definitions such as Average Order Value, ForgeDesk, and GMV. At the bottom there is a line: Genie Agents are conversational interfaces that are scoped to a single domain.

What this demo is trying to show is that Databricks puts Genie Ontology, Genie One, MCP, and Unity Gateway on one chain, so that agents can use trusted business context while retaining access guardrails. The official description is not complicated; the focus is on the two terms trusted business context and guardrails. Translated into everyday questions: the agent needs to know how GMV is calculated, whether AOV is measured by order or by user, and what ForgeDesk refers to in the business; it also needs to know which tables it can see and which data it cannot touch.

Many teams can already let agents connect to databases, generate SQL, and produce results. The real trouble is whether the results can be accepted by business and finance. Someone in the comments replied: Every company has three revenue columns and only one that finance trusts. That is a bit harsh, but very accurate. Once metric definitions get messy, the faster the agent is, the more consistently wrong it will be.

This Databricks demo does not stop at the level of being able to query data; it puts the semantic layer and the permission layer up front. For readers with a development background, MCP and Unity Gateway are the key terms for integration and governance; for non-technical readers, it can be understood as giving the agent an internal company dictionary plus an access card.

demo link: https://youtu.be/ZuLAAdIjeeE

If the agent does not know what your business is talking about, the numbers it queries are just strings.