OpenAI's new Data agent for ChatGPT Work queries databases and edits dashboards

OpenAI added a Data agent to ChatGPT Work on Thursday. The pitch, in OpenAI's own words, is that it lets someone "turn your company's data into answers, interactive dashboards, and action, just by asking". The interesting part is the list of things it connects to, because a chat-over-your-data feature is only useful when it reads the warehouse the company already runs.

The connector list is the main news

The Data agent connects to Snowflake, BigQuery, Databricks, Redshift, MongoDB, Google Drive, and SharePoint, and it works with Power BI, Tableau, Sigma, and ThoughtSpot on the dashboard side. OpenAI's page says a dashboard the agent creates "can be edited, shared, and refreshed" in those BI tools, so the analyst who reviews it keeps working in the tool the analytics team already uses.

A user activates it through the ChatGPT Work Plugins directory, connects a data source, and then addresses the agent with @Data inside a chat. Results can be posted to Slack or sent as email, and the agent can "perform approved actions through connected tools", which OpenAI does not describe in more detail on the announcement page.

How permissions are handled

The one line that decides whether a company can adopt this reads: the agent "respects access controls. Admins manage connections and availability; existing table, row, and column permissions apply". If the warehouse blocks an analyst from seeing customer PII in a Snowflake row, the agent asked from that analyst's ChatGPT seat is blocked too. That prevents the problem that has kept most companies from putting a natural-language layer over a warehouse: the model reads and quotes whatever it can access, and permission checks in the model layer are the wrong place to enforce them.

OpenAI also says the agent "incorporates established metrics, calculations, and data relationships to interpret results", which is the missing part of self-serve analytics. The reason a business user asking a warehouse for "revenue last quarter" gets different answers from three tools is that each one guesses the join and the filter. The agent using the metric layer the analytics team already defined at least removes that guess.

What the announcement page does not say

OpenAI does not name the price for the Data agent, does not name the ChatGPT Work tier it is available on, and does not name the models it uses. It does not describe the write path in enough detail to know which "approved actions" it can perform in a connected tool. It also does not give any accuracy figure for the SQL the agent generates, which is the number every analytics team will ask about first.

Those gaps matter because the agent will be connected to the warehouse an analytics team spent years cleaning up. It is worth asking, before installing the plugin, whether an admin can review the queries the agent runs and whether a bad answer is auditable back to the row that produced it. The permission model is right. The rest is still to be measured.

Source

OpenAI: Now everyone can put data to work, and OpenAI developer community: Introducing the Data Agent for ChatGPT Work.