OpenAI announced a Data agent in ChatGPT Work on September 10, 2026. The company says it can connect to approved enterprise data sources and business context, answer questions in natural language, create interactive dashboards and carry out user-approved actions through connected tools.

The announcement establishes the product’s described capabilities, but it does not provide independent benchmarks for analytical accuracy, calculation correctness or the quality of its recommendations. It also does not specify pricing, country-by-country availability or whether every listed integration is generally available.

Contents

What OpenAI announced

OpenAI is positioning the Data agent as a natural-language interface for business analysis. Users can ask questions such as what changed in a metric, investigate possible drivers and continue with follow-up questions without writing database queries or learning a separate analytics application.

The agent can connect to approved data sources including:

  • Amazon Redshift
  • Datadog
  • Google BigQuery
  • ClickHouse
  • Databricks
  • MongoDB
  • Snowflake

OpenAI says additional sources are supported. The agent can also bring files and documents from Google Drive and SharePoint into an analysis.

This is not presented as a replacement for an organisation’s existing data infrastructure. Instead, the announced workflow is designed to sit across data warehouses, databases, documents, semantic systems and business-intelligence platforms.

The Data agent is listed as Data in the Plugins directory in ChatGPT Work. Administrators can make it available or install it through Workspace settings, then configure the relevant data-source plugins.

How the agent interprets business data

A central part of the product is its use of organisational business context. OpenAI says the Data agent can use metric definitions, custom calculations, business terms and relationships supplied through semantic layers and trusted sources.

A semantic layer is a governed representation of business concepts—such as metrics, dimensions, relationships and calculations—that helps analytics systems interpret underlying data consistently. It provides context for how an organisation defines and connects its business information.

OpenAI lists Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon and existing business-intelligence dashboards as examples of sources for this context.

The announcement presents the Data agent as combining natural-language interaction with organisation-specific metric definitions, calculations, relationships, semantic context and access rules.

OpenAI says administrators choose which connections are available and which user roles can use them. Queries are also described as enforcing the connected account’s existing permissions, including table-, row- and column-level restrictions.

Table-level permissions control access to whole datasets. Row-level permissions limit which records a user can see, while column-level permissions can restrict access to particular fields. These controls are intended to ensure that a conversational request does not automatically bypass the access rules already applied to the underlying account.

The announcement describes the controls, but does not provide a detailed security, privacy or governance specification.

From questions to dashboards and actions

After investigating a question, users can inspect the evidence behind findings and turn the analysis into an interactive dashboard. OpenAI says dashboards can be edited, shared and refreshed, and can be adapted to an organisation’s visual identity using supplied brand guidelines.

An interactive dashboard combines visualisations with functions such as filtering, exploration and data refresh. That allows users to examine a trend from different angles rather than receiving only a static report or a single generated answer.

The Data agent can also build and interact with dashboards in:

  • Omni
  • Oracle BI
  • Power BI
  • Sigma
  • Tableau
  • ThoughtSpot

OpenAI describes the workflow as extending beyond analysis. The agent can recommend next steps, identify people who should be involved, share findings through Slack or email, and carry out actions approved by the user through connected tools.

This is the main change in the announced workflow: the product is intended to connect a business question to investigation, visualisation, communication and an approved follow-up action. The evidence packet does not establish how reliably those steps work in practice, or how the system handles ambiguous metric definitions and incorrect interpretations.

OpenAI provides example prompts for diagnosing changes in a metric, designing a key-performance-indicator framework and preparing leadership updates containing actuals, comparisons, drivers, caveats and recommended actions.

Controls and limitations

The product’s capabilities are currently supported by an OpenAI announcement rather than an independent evaluation.

The company also says organisations in its Alpha programme—including NTT DATA, Thermo Fisher and ServicePiston—are using the Data agent for tasks such as analysing sales and spending, identifying reporting errors and evaluating opportunities. These examples describe reported use cases, not controlled comparisons with conventional business-intelligence workflows.

Several practical details remain unspecified:

  • the complete list of supported connectors;
  • data-volume limits and expected latency;
  • refresh behaviour;
  • failure-handling procedures;
  • the availability of each integration across plans or programmes;
  • pricing and country-by-country availability;
  • how sensitive data is retained and isolated across connected systems.

The announcement also does not quantify the accuracy of analyses, dashboards, recommendations or generated actions. Permission enforcement can restrict which data a user queries, but it does not by itself establish that the resulting interpretation is correct. Organisations would still need to validate metric definitions, calculations, evidence and proposed actions before relying on them for consequential decisions.

In particular, the product description does not explain how users should verify analyses, resolve ambiguous metric definitions or detect a plausible but incorrect causal explanation for a change in a business metric. A system may identify a correlation or produce a coherent narrative without establishing that one factor caused another.

What to watch

The most important next evidence will be independent testing on real enterprise datasets. Useful evaluations would measure factual accuracy, calculation correctness, evidence quality, hallucination rates, permission handling and the reliability of generated actions.

More information is also needed about data governance, retention, isolation and handling across the connected services. Those details will matter to organisations managing sensitive data.

OpenAI’s announcement presents the Data agent as a way to make existing business data more accessible without replacing an organisation’s warehouses or BI systems. Whether it reduces reporting bottlenecks in practice will depend on the quality of those underlying systems, the clarity of business definitions and how carefully organisations validate the agent’s conclusions.

Sources