共有:

ChatGPT Work · Data Agent

Ask your company data a question, get a dashboard back.

No analyst required — ask a plain-language question about internal data and it builds the dashboard for you. ChatGPT Work's new "Data agent" positions itself as a layer in front of the BI tools you already run.

AI Navigate Editorial2026.09.127 min read

"Why did churn rise last month?" DATA AGENT Snowflake / BigQuery Databricks / MongoDB Drive / SharePoint Datadog APPROVED SOURCES Tableau / ThoughtSpot
01
The Announcement

Skip the analyst,
ask the data directly

Instead of writing a query, you just ask what you want to know.

On September 10, 2026, OpenAI announced a "Data agent" for ChatGPT Work. Per the official post, Now everyone can put data to work, it's a plugin that connects to a company's approved data sources, investigates what changed in a business metric, and automatically builds interactive dashboards employees can share. Questions that used to require a round trip through a BI tool and an analyst — why did sales slow last month, where is spend rising, which large accounts are at risk of churning — can now be worked through inside a single conversation.

Supported data sources span Amazon Redshift, Google BigQuery, ClickHouse, Databricks, MongoDB, Snowflake, and Datadog, and the agent can also pull in internal documents from Google Drive and SharePoint alongside the numbers. The full integration list is laid out on OpenAI's "ChatGPT Work for Data Teams" page.

BeforeWith the Data agent
Required BI-tool skills or SQL knowledgeA plain-language conversation is enough
Dashboards meant filing a request with an analystThe requester builds and refines it inline
Numeric data and internal docs checked separatelyWarehouse data and Drive / SharePoint docs cross-referenced together
Output reproduced into BI tools by handWritten directly into Tableau, Power BI, ThoughtSpot, etc.

02
Coverage

Breadth of integration is the adoption barrier, lowered

7
supported data platforms (Redshift, BigQuery, etc.)
6
connected BI tools (Tableau, Power BI, etc.)
2
document integrations (Drive, SharePoint)

The Data agent can build and operate dashboards directly inside Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot. Rather than replacing those BI tools, it layers a natural-language front door in front of them — which means companies already running several BI stacks in parallel get the benefit without a migration cost.

This breadth lines up with ThoughtSpot's own "Spotter 3" news from September 11, 2026 (an MCP-host update that searches Slack and Salesforce alongside the warehouse). BI vendors are moving to pull in outside data; OpenAI is pushing in the opposite direction — writing from an outside conversational agent into the BI tools themselves.

03
In Practice

Who actually benefits

PMs and IT

Building an internal dashboard no longer means waiting on an analyst's queue. Smaller teams without a dedicated analytics function stand to gain the most.

Engineers

Less time spent building lightweight internal analytics tools — but a new design question appears: exactly which data-platform permissions should this agent be handed.

Everyday users

This is a ChatGPT Work (business plan) feature only, so it has no direct bearing on personal-plan users.


Once a question turns straight into a dashboard,
the real question becomes who vouches for the numbers.


04
Risk & Rollout

Assumptions worth checking before rollout

The first step, before adopting this at all, is confirming your own data platforms are actually on the supported list. Second, because the agent can now cross-reference internal documents from Google Drive and SharePoint, document-level access permissions need to be tighter than before — a loose permission on the document side becomes a leak path the moment an analysis touches it. Third, before a full rollout, it's worth asking a business question you already know the answer to and checking whether the resulting numbers and dashboard actually match.

The other concern is that the ambiguity of natural language becomes ambiguity in the analysis itself. Does "last month" mean the calendar month or the trailing 30 days? What exactly counts as "churn risk"? A human analyst would ask a clarifying question; an agent may just pick an interpretation and hand back a dashboard that reads as definitive. Treating this as a first-pass screening tool — with any number that actually drives a decision double-checked by a person — is the more realistic way to run it.