OpenAI releases an agent that queries company data in plain language

The surface it plugs into is wide. Approved data sources include Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB and Snowflake, and files and documents in Google Drive and SharePoint can be pulled into the analysis as well. The output has more than one home, too. Dashboards can be built and manipulated inside Omni, Oracle BI, Power BI, Sigma, Tableau and ThoughtSpot.

Understanding vocabulary that only means something inside one company is where tools like this are won or lost. OpenAI said the agent reads an organization's business terms, metric definitions, custom calculations and the relationships between data to interpret what it sees. That context comes from semantic layers already accumulated in places like Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon and BI dashboards. A phrase such as active customer means something different at every company, and rather than take that definition down again, the agent reads wherever it has already been settled.

Permissions are inherited rather than rebuilt. Enterprise administrators choose which data connections are open and which roles can use them, and queries enforce the permissions already attached to the connected account. Restrictions hold at the row and column level, not only the table. A number you could not see before does not become visible because it passed through a chat window.

One notable feature of this launch is that the companies who build data warehouses lined up behind it at once. Pablo Stern, chief product officer for AI and emerging products at MongoDB, said that "our customers' most important data lives in Atlas, updating in real time as their business runs. Now the Data agent in ChatGPT Work can reach that data directly, so instead of waiting on a report built from yesterday's numbers, anyone can ask a plain-language question and get an answer grounded in what's happening right now." Southard Jones, executive vice president and chief product officer at Tableau, explained that connecting Tableau "brings trusted business semantics into a place employees already work, so the answers they get are grounded in the same data model their teams rely on."

The examples from early users cluster around who gets to build a dashboard. Yuji Shono, head of the global AI office at NTT DATA Group, said that "at NTT DATA, licensing costs, effort, and technical expertise have made it difficult to expand the use of dashboards across our organization," and that many non-engineers, particularly in sales and corporate functions, now build and update their own dashboards in plain language. Ankur Bhatt, vice president of engineering for agent OS and applied AI at ServiceTitan, said a dashboard built with the tool showed that users of Atlas, the company's AI sidekick, launched campaigns at roughly three times the rate of non-users, and that the finding is now the basis for simplifying onboarding.

OpenAI said it ran the tool inside its own walls first. Nearly its entire product organization, and more than two-thirds of the go-to-market organization that covers sales and marketing, analyze company data with data agents themselves. That did not happen on its own, though. The company wrote that it was possible because its data team created shared business definitions, set access rules and put safeguards around sensitive data. There is groundwork to finish before the tool gets switched on.

All eleven companies cited as examples in the announcement were alpha participants or evaluation partners. Carlos Arroyo, strategic project lead at micro1, described himself as an OpenAI evaluation partner while saying his operations team rebuilt its performance-tracking dashboards in half an hour and caught errors in the originals along the way. CookUnity wrote that it checked the results against internal reports. METAL read the document, and the case that caught errors and the admission that cross-checking was still needed sit side by side in it.

Taken whole, this is not a story about a new model but a story about position. The companies that build data warehouses and BI tools have accepted the ChatGPT window as a new front end for their products, and in exchange OpenAI is borrowing the semantic layers and permission systems each of them has already built. Neither side replaces the other in that trade. METAL has reported on ChatGPT connecting to Epic EHR and nine other kinds of healthcare data, and this release moves that list one step further toward the enterprise data warehouse.

Getting to it is simple. It appears as Data in the ChatGPT Work plugin directory, and an administrator installs it or opens it to a team from workspace settings. Data-source plugins such as Databricks and Snowflake are enabled from the same place, along with who may use them. Once account connections are complete, you call Data in a conversation and ask your business question. The answer to who does the analysis is shifting from the people who know how to analyze to the people who have something to ask.