01
Overview
OpenAI's Data agent is a business-data agent for ChatGPT Work and Codex that can connect approved company data, investigate business questions, explain changes and turn findings into interactive reports and dashboards. OpenAI launched it on September 10, 2026 as an OpenAI-built plugin rather than as a new foundation model.
02
What is OpenAI Data agent?
Data is designed to sit between a user's business question and the systems where the underlying evidence lives. Instead of manually exporting warehouse tables into a spreadsheet or pasting dashboard screenshots into ChatGPT, a permitted user can ask a question in natural language and let the agent work across connected data sources, business definitions and supported BI tools.
OpenAI's launch material positions Data for jobs such as understanding why a metric moved, investigating a funnel problem, comparing cohorts, finding patterns across structured and unstructured sources, and building a dashboard from the analysis. These are vendor-described capabilities. They should not be read as a guarantee that every query will be correct or that Data replaces a company's analytics review process.
03
Where Data works
OpenAI currently documents the Data plugin for ChatGPT Work and Codex. A workspace administrator may need to make the plugin and the required underlying source plugins available before a user can install and use it.
Access can therefore depend on the workspace, role, source configuration and individual permissions. OpenAI's launch pages do not provide one universal entitlement table or standalone Data-agent price, so availability should be checked in the live workspace rather than inferred from a generic ChatGPT plan name.
04
Which data sources can it connect to?
OpenAI's current documentation lists Amazon Redshift, ClickHouse, Databricks, Google BigQuery, MongoDB and Snowflake among the supported structured data sources. Documents and files can also come from Google Drive and SharePoint when those sources are connected and permitted.
The important boundary is that Data does not bypass the original system's permissions. OpenAI says existing table, row and column restrictions continue to apply. If a user cannot access a source record through the connected system, the Data agent should not be treated as a way around that restriction.
05
Business context and semantic layers
A warehouse contains data, but it does not automatically explain what a company's metrics mean. OpenAI therefore emphasizes business context: metric definitions, custom calculations, trusted dimensions, data relationships and semantic layers.
Its documentation cites sources such as Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon and existing BI dashboards as possible places from which Data can obtain that context. This matters because a request like “why did conversion fall?” is only useful if the agent understands which conversion definition, time window, segment and source of truth the company actually uses.
Users should still verify the source, time period, filters and metric definition behind important conclusions. Natural-language access reduces friction; it does not remove the need for measurement discipline.
06
How the analysis workflow works
A typical Data-agent workflow has four parts. First, the user asks a business question. Second, Data determines which approved source or sources are relevant and investigates the evidence. Third, it explains the result and can iterate when the user asks follow-up questions. Fourth, the findings can be turned into an interactive report or dashboard when that format is useful.
For a marketing team, that could mean investigating an acquisition drop by channel, comparing landing-page cohorts, examining customer-retention patterns or reconciling campaign data with a business-defined metric. For sales or operations, the same pattern could apply to pipeline movement, account health or recurring operational metrics.
07
Interactive dashboards and BI tools
OpenAI says Data can create and interact with dashboards and can work with existing BI systems. Its current documentation names tools including Omni, Oracle BI, Microsoft Power BI, Sigma, Tableau and ThoughtSpot.
This makes Data different from a one-off chat with a CSV. The intended workflow can continue into a reusable reporting surface: build a dashboard, refine it conversationally, share it through supported tools and return to it as the underlying question evolves.
The exact actions available depend on the connected BI product, workspace configuration and user permissions. A supported integration name should not be interpreted as universal permission to create, publish or modify every artifact in that product.
08
Publishing and keeping dashboards updated
OpenAI's Help Center says Data can work with OpenAI Sites when that plugin is installed. A dashboard can be published and shared there, and cloud automation can be used to keep a dashboard up to date.
There is an important data-governance implication: OpenAI warns that data used in an analysis is copied into a published Site. Teams should therefore review the audience, source permissions and sensitivity of the included data before publishing an internal dashboard more broadly.
09
Can Data take actions?
Data can be part of workflows that lead to actions through connected tools. OpenAI gives examples involving communication and business systems, but action capability depends on the specific tool, permissions and any approval requirements. A report that recommends contacting an account is not the same as an agent having unrestricted authority to message that account.
For production business workflows, the safer design is to separate analysis from consequential action unless the downstream tool provides appropriate permissions and review controls.
10
Permissions and security boundaries
OpenAI's documentation says Data respects existing source permissions, including table-, row- and column-level restrictions. Workspace administrators control plugin availability and users connect sources through the approved integration flow.
That does not eliminate governance work. Teams should still define which systems may be connected, which semantic definitions are authoritative, what kinds of reports may be shared and which actions require human approval. Published Sites deserve particular review because source data used in the analysis may become part of the published artifact.
11
Pricing and availability
No standalone Data-agent price was verified in the primary sources reviewed for this article, and OpenAI does not present one universal access matrix for every workspace and source combination. Do not treat the feature as universally included or free.
Before adopting it, verify three things in the live product: whether the Data plugin is available to the workspace, whether the required source plugins are available and permitted, and whether the connected systems introduce their own licensing or usage costs.
12
What Data does not prove on its own
OpenAI's launch includes customer examples and descriptions of faster analysis workflows, but those examples are not independent performance benchmarks. The presence of an AI agent also does not make an ambiguous metric definition correct, fix poor warehouse modeling or guarantee causal explanations for a change in a business metric.
A useful operating rule is to use Data to accelerate investigation and synthesis while keeping the underlying data model, metric definitions and high-impact conclusions reviewable by people who own the business context.
13
Who is it most useful for?
Data is most relevant to teams that already have meaningful business data but lose time moving between a warehouse, documents, BI dashboards and manual analysis. Marketing, growth, finance, product and operations teams are obvious fits because their questions often require evidence from several systems and business-specific definitions.
Smaller businesses without a warehouse or structured analytics setup may get less value from the full workflow. The agent can reduce the interface cost of analytics, but it cannot create high-quality underlying data that does not exist.
14
Frequently asked questions
Is OpenAI Data agent the same as uploading a spreadsheet to ChatGPT?
No. File analysis can answer questions about uploaded data, while Data is designed to work across approved company sources, semantic context and supported business tools as a reusable workflow.
Does it work with Snowflake and BigQuery?
OpenAI currently lists Snowflake and Google BigQuery among supported data sources, alongside Databricks, Redshift, ClickHouse and MongoDB.
Can it build Power BI or Tableau dashboards?
OpenAI documents interoperability with Power BI, Tableau and several other BI tools. The exact create, edit, publish and sharing actions depend on each integration and the user's permissions.
Does it ignore row-level permissions?
OpenAI says existing source permissions, including row- and column-level controls, continue to apply.
Is there a separate price for Data?
No standalone Data-agent price was verified in the primary launch documentation used here. Workspace access and connected-tool costs should be checked directly before purchasing or deploying.
15
Bottom line
OpenAI Data agent is an attempt to turn business analytics from a sequence of tool handoffs into an agentic workflow: ask a question, retrieve approved evidence, apply company-specific metric context, investigate the result and turn it into a reusable dashboard or action-ready report. The strongest opportunity is not that it makes analytics automatic, but that it can reduce the distance between a business question and the systems needed to answer it while preserving source permissions and human review.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.