Data Agent in ChatGPT Work: Connected Sources, Dashboards, and Alternatives (2026)
On September 10, 2026, OpenAI introduced a Data agent inside ChatGPT Work. The announcement is short on marketing and unusually specific about plumbing. It names which warehouses the agent reaches, which semantic layers it reads, and which BI tools it can write back into. It also names who decides all of that.
That specificity is the reason this release is worth reading closely. Most AI analytics announcements describe an outcome. This one describes a supply chain — from the warehouse, through the business definitions, to a dashboard someone can edit.
This article walks through what the announcement actually says and what it leaves undefined. It then covers how the same job gets done when your data does not live in a warehouse at all.
What OpenAI announced
The headline claim is a single sentence on the announcement page. The Data agent "connects to your company data, investigates what changed, and builds interactive dashboards you can share."
Three verbs are doing the work there. It connects, it investigates, and it builds. The third one is the part that separates this from a chat interface over a database. The output is not a paragraph of analysis. The output is an artifact your team can open again next week.
The page frames the problem in terms of waiting. "Getting those answers often means waiting for a report or asking someone else to run the analysis." The Data agent is positioned as the thing that removes the queue, not the thing that removes the analyst.
One more line sets the boundary of the experience. You direct and refine the analysis in one conversation, "without writing queries or learning a new analytics tool." The agent is meant to be the surface, not an additional tab.
Which data sources it connects to
This is the most concrete section of the announcement, and it is worth quoting rather than summarizing.
The Data agent "connects to approved data sources including Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB, Snowflake, and more." Alongside those, it "can also bring files and documents from Google Drive and SharePoint into the analysis."
So the input surface has two halves. One half is the warehouse and the operational databases. The other half is the pile of documents that never made it into a warehouse. That means the quarterly deck, the signed contract, and the spreadsheet a regional manager maintains by hand.
Redis and G2 also appear as platform partners on the page. That widens the picture past analytics storage into two other kinds of input. Redis covers real-time state, the numbers that change while you are looking at them. G2 covers third-party market data, which is context a company does not generate itself.
The combination is worth noting. This is an agent that can read the warehouse, the document store, the live cache and an outside data provider. It is being positioned as a single front door to everything a business knows.
The semantic layer is the interesting part
A warehouse connection alone does not tell an agent what "active customer" means at your company. The announcement addresses this directly.
The agent "uses your organization's business terms, metric definitions, custom calculations, and data relationships to interpret the data." That context, per the page, "comes from semantic layers and trusted sources such as Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon, and BI dashboards."
Two things follow from that list. First, GitHub sitting next to dbt implies that metric definitions living in version-controlled code are treated as a source of truth, not just documentation. Second, the value of this feature scales with how much definitional work a company has already done. A team with a mature dbt project gets a very different experience than a team without one.
This is the quiet prerequisite in the whole announcement. OpenAI is explicit about this when describing its own rollout. Its data team "made this possible by creating shared business definitions, setting access rules, and putting safeguards in place for sensitive data."
What it produces
The stated output is an interactive dashboard, not a static answer.
Per the page, you "turn the analysis into an interactive dashboard with built-in visualizations," and "your team can edit, share, and refresh it as needed." Brand guidelines can be supplied so outputs match an organization's look and feel.
There is also a write-back path into tools that already exist. The Data agent "can also build and interact with dashboards in Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot." That design choice matters more than it first appears. It means the artifact does not have to live inside ChatGPT Work to be useful. That lowers the cost of trying the feature in a company that has already standardized on a BI tool.
Findings can be pushed onward as well. The agent can share results "through Slack or email and carry out the actions you approve through connected tools."
That last clause deserves attention. Approval sits between the recommendation and the action, which means the described flow stops short of letting the agent act unattended. The page also says you can "ask ChatGPT Work to recommend next steps and identify who needs to be involved." Routing is therefore treated as part of the analysis rather than as a separate step.
Who controls access
The announcement is unusually direct here, and this section is the one most worth reading if you are evaluating rollout risk.
"Enterprise administrators choose which data connections are available and which roles can use them." Queries then "enforce the connected account's existing permissions, including table, row, and column restrictions."
That is two separate layers. Administrators gate which connectors exist at all. The underlying warehouse permissions still apply on every query. A person who could not see a salary column yesterday does not see it through the agent today.
How to turn it on
The setup path described on the page runs through the plugin system rather than a global toggle.
You find the Data agent "listed as Data in the Plugins directory in ChatGPT Work." Administrators "can make it available or install it for their teams through Workspace settings > Plugins." They can also "enable and configure the relevant data-source plugins, such as Databricks and Snowflake, and manage who can use them."
If it is not already installed, you select Install plugin from the directory, complete the account-connection steps, then start a conversation with @Data.
What the announcement does not cover
Worth stating plainly, because the gaps shape who this is for.
The page does not describe pricing for the Data agent as a separate line item. It also does not list which ChatGPT Work seat tiers include it. No general availability date is stated. Named Alpha program participants are described as already using it, among them NTT DATA, Thermo Fisher Scientific, ServiceTitan and Zipline.
There is also no described path for someone whose data is not in a connected system. Every documented entry point assumes either a warehouse connection or a Google Drive or SharePoint account that an administrator has enabled.
Where this leaves smaller teams
Read the prerequisites together and a pattern appears. You benefit most if you have a warehouse, a semantic layer, an administrator who can approve connectors, and a BI tool to write back into.
That describes a real and large category of company. It does not describe the team of six that runs on exports. Their inputs are a CSV from the billing system, a spreadsheet the operations lead maintains, and a PDF the vendor sends every month.
For that second group the bottleneck was never query access. The bottleneck is that the data arrives as files, and files have to be reconciled before any question can be asked of them.
Alternatives when your data lives in files
If the input is a folder rather than a warehouse, the tool has to start somewhere else. It starts at upload rather than at connection.
Powerdrill Bloom is built for that starting point. You upload spreadsheets, CSVs, PDFs and documents directly. Ask the question in natural language, and charts, tables and exports come back with the underlying rows attached. Its homepage describes the behavior as an analyst that "shows its work." Every number "comes back with the page, the row and the figure behind it."
That traceability requirement is the same one the warehouse-connected approach solves through the semantic layer. When you cannot point the agent at a governed definition, the next best guarantee is showing which row produced which number.
Other routes exist depending on where the friction is. If your data is already in Google Sheets and the complaint is about scale, the best Google Sheets alternatives cover tools built for larger grids. If you want the broader category definition first, what is an AI data agent sets out what these systems have in common.
FAQs
What is the Data agent in ChatGPT Work? It is a plugin announced by OpenAI on September 10, 2026. It connects to approved company data sources, investigates changes, and builds interactive dashboards that a team can edit and share.
Which databases does it connect to? The announcement names Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB and Snowflake, and says more are supported. It can also pull files from Google Drive and SharePoint.
Can it create dashboards in Power BI or Tableau? Yes. The page states it can build and interact with dashboards in Omni, Oracle BI, Power BI, Sigma, Tableau and ThoughtSpot.
Who decides which data the agent can reach? Enterprise administrators choose which connections are available and which roles may use them. Each query also enforces the connected account's existing table, row and column permissions.
What if my data is in spreadsheets rather than a warehouse? Every documented setup path for this feature assumes a connected system or an enabled Drive or SharePoint account. File-first tools take uploads directly instead, which suits teams whose inputs arrive as exports.
Conclusion
The Data agent announcement is best read as a statement about prerequisites. Connectors, semantic layers, administrator approval and a BI destination are all named explicitly. The value of the feature rises with how much of that groundwork already exists.
That makes it a strong fit for companies that have invested in governed data. It is a weaker fit for teams whose numbers still arrive as attachments. If your inputs are files rather than tables, start where the files are. Try Powerdrill Bloom free and upload one export to see the path from raw rows to a shareable answer.
Source: OpenAI, "Now everyone can put data to work," September 10, 2026. https://openai.com/index/put-data-to-work