Wink Pings

Copilot Can Read Your Email, but It Can't Recognize Your Revenue Table

Databricks put Genie into Copilot Cowork, Teams, and Excel. What enterprise AI has always lacked is not model intelligence, but the semantic layer that knows which table actually counts.

Databricks published a blog post on November 18 and put it plainly: your AI assistant can read email, dig through files, and look at your calendar, but it doesn't know which revenue table is the certified one, and it doesn't know how the finance department actually defines net retention.

The result is that every department gets different numbers, and after using them twice, no one dares to use them again.

This time their answer is the Genie plugin, which connects Genie on Azure Databricks into Microsoft Copilot Cowork and Copilot Chat. The same set of capabilities has also been extended to Teams, Excel, and Copilot Studio.

The key issue is not the model

Databricks repeatedly emphasizes something called Genie Ontology. According to the official description, it blends business semantics explicitly defined by teams (such as metric views and certified definitions) with context learned from daily work to form a shared layer of business understanding. This layer is now generally available.

In plain English: previously, when Copilot saw a pile of table names, it could only guess. Now it can first ask Genie which definition Finance uses to calculate net retention; Genie provides the certified definition, and the answer comes with sources.

The value here is not that the answer is smarter, but that the answer is reproducible. Auditors can check it, and Finance accepts it.

Where exactly it has been plugged in

Teams. The Databricks Genie app is now generally available. You can DM it directly or @Databricks Genie in a channel. It replies with citations. It supports scheduled analyses; for example, if you say 'Every Monday at 9 a.m. send me this analysis,' the result will appear in your DMs. You can also drag files in for analysis—PDF, Word, and CSV all work—and the returned charts are posted directly in the conversation.

One example from the blog is quite specific: a sales manager asks in a channel which leads over $100,000 have no one following up. Previously, someone had to pull reports, cross-check the CRM, dig through meeting notes, and confirm with each sales rep one by one. Now you just @ it.

Copilot Cowork. Open the Azure Databricks Genie plugin, click Connect, and sign in with your Databricks credentials. The official post specifically notes that you do not need to register an OAuth app or modify the manifest. This detail matters to people doing integrations: one less process step means a lot fewer support tickets.

Excel. In the add-in, you can ask questions directly in natural language, and answers come with citations; you can follow up and continue from the previous session. You can also build live pivot tables on Unity Catalog metric views and tables—drag fields and numbers appear, with no refresh button. It also supports writing inputs such as forecast assumptions back to Databricks tables.

Copilot Studio. Through the Azure Databricks connector for Power Platform, you can mount Genie Agent as a tool. This is suitable for scenarios where you need to combine Genie with other non-Databricks tools.

The governance part is fairly concrete

Every question is executed under the identity of the person asking; Unity Catalog and source-system permissions determine what that person can see. Questions are processed within the Databricks workspace, and the message routing service between Teams and Databricks does not store questions or answers.

On the admin side, Unity Gateway manages MCP policies, cost control, and invocation tracking. MCP policies determine which servers and tools Genie can use, who can access them, and which operations require human approval.

Frankly, where enterprise AI assistant rollouts get stuck, eight times out of ten it's here. It's not that it can't answer; it's that no one dares to let it answer recklessly.

A signal worth pondering

There are many players doing semantic layers in the Microsoft ecosystem. Databricks chose to move its own layer into Microsoft's tools rather than asking users to move to Databricks. The direction is clear: Microsoft owns the entry point, while Databricks keeps the definition authority.

The blog also left a teaser: Genie will later connect to Work IQ through MCP, bringing in email, meetings, documents, and people context from Microsoft 365. Each request will run under the signed-in user's Entra ID identity, and Microsoft's permissions and sensitivity labels will continue to apply.

If this gets fully connected, when an enterprise asks in Teams, 'How is this customer doing lately?', what gets strung together behind the scenes is two governance systems. It sounds complex, but that is indeed what a real enterprise should look like.

The documentation is on the Databricks website, and the original blog is here.

发布时间: 2026-11-21 07:51