OpenAI launches Data agent in ChatGPT Work: BigQuery, Snowflake, Databricks, and interactive dashboards
OpenAI has launched a new Data agent in ChatGPT Work, giving employees a conversational interface for querying company data, investigating changes in business metrics, and building interactive dashboards without having to move manually between data warehouses, BI tools, spreadsheets, and internal documentation.
The agent connects to approved enterprise sources including Google BigQuery, Snowflake, Databricks, Amazon Redshift, ClickHouse, MongoDB, and Datadog, while documents and unstructured context can also be incorporated from services such as Google Drive and SharePoint.
Rather than treating every request as an isolated natural-language query, Data can use an organization’s existing metric definitions, custom calculations, dataset relationships, semantic layers, and trusted dashboards to determine what a term such as revenue, retention, active customer, or qualified pipeline actually means inside that company.
Users can investigate an initial result through follow-up questions and then convert the analysis into an interactive dashboard that can be edited, refreshed, shared, and customized to corporate visual guidelines.
OpenAI has also designed the agent to work with existing BI environments rather than replace them outright, with integrations spanning Power BI, Tableau, Sigma, ThoughtSpot, Omni, and Oracle BI, while access to underlying data continues to follow the permissions of the connected enterprise account.
The result is a product that sits between the company’s data infrastructure and the employees making operational decisions, with the analytical workflow moving from manually requesting reports toward continuously interrogating governed data through the same ChatGPT conversation.
··········
THE DATA AGENT CONNECTS CHATGPT TO THE COMPANY’S EXISTING DATA STACK.
OpenAI is not asking companies to migrate their warehouses into a new proprietary analytics database, because the agent is designed to operate across the data systems organizations already use.
A typical enterprise analytics environment is fragmented across several layers that were originally built for different users and different purposes.
Transactional and analytical data may sit in Snowflake, BigQuery, Databricks, Redshift, MongoDB, or ClickHouse, while business logic exists separately inside dbt models, semantic layers, internal documentation, dashboards, and finance or operations definitions.
Employees who need an answer often do not have direct access to every layer, and even when they do, they may need SQL knowledge, familiarity with table structures, and an understanding of how the organization defines each metric before they can produce a reliable result.
The Data agent introduces a conversational layer over those systems, allowing the user to ask a business question while the agent retrieves data and context from approved sources.
The architecture matters because the model is not expected to infer every business definition independently from raw tables.
OpenAI specifically recommends connecting a semantic layer containing authoritative definitions and queries, so that the agent can distinguish company-specific concepts rather than relying only on statistical patterns or column names.
Supported context can come from systems such as Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon, and existing BI dashboards, creating a bridge between the underlying data and the definitions employees already trust.
........
Data Studios mapping of the Data agent’s enterprise information layers.
Enterprise layer | Examples explicitly supported | Role in the analysis |
Data warehouses and databases | BigQuery, Snowflake, Databricks, Redshift, ClickHouse, MongoDB | Structured company data |
Operational analytics | Datadog and other connected analytics services | Real-time or operational metrics |
Documents and files | Google Drive, SharePoint | Unstructured business context |
Semantic and definition layer | dbt, Databricks Genie Ontology, Snowflake Horizon, GitHub | Metric definitions, calculations, relationships |
BI and visualization | Power BI, Tableau, Sigma, ThoughtSpot, Omni, Oracle BI | Existing dashboards and visualization environments |
ChatGPT output layer | Interactive Data dashboards and reports | Investigation, presentation, sharing and refresh |
........
The significant change is therefore not simply that ChatGPT can generate SQL.
The product attempts to connect data, definitions, visualization, and business context inside the same analytical session, which addresses a much broader part of the analytics workflow than query generation alone.
··········
USERS CAN INVESTIGATE A METRIC CHANGE WITHOUT WRITING THE QUERY THEMSELVES.
The agent translates a business question into an analytical workflow, then lets the user continue investigating the result through natural-language follow-ups.
A user can begin with a question such as why weekly active users changed, why spending increased, which customer accounts present the greatest renewal risk, or why sales performance diverged from the previous quarter.
The first answer is not intended to end the analysis.
The user can ask Data to compare periods, segment the result, test another explanation, inspect the evidence behind a chart, or verify which metric definition and filters were used.
This iterative structure is important because business analytics rarely consists of a single query whose output immediately explains the underlying cause.
A decline in revenue, for example, can require decomposition by geography, product, customer cohort, acquisition channel, contract size, pricing, or timing before the relevant driver becomes visible.
Traditional self-service BI already supports much of that exploration, but it generally requires the user to understand the interface and available dimensions beforehand.
The Data agent shifts more of the navigation into the conversation, where the employee describes the analytical direction while the system determines how to interrogate the available data.
OpenAI still advises users to check the source, time period, filters, and metric definition before relying on a result, particularly when an answer differs from an existing company report.
That requirement remains important because natural-language access reduces the effort required to run an analysis but does not remove the need to validate whether the correct dataset and business definition were used.
··········
INTERACTIVE DASHBOARDS CAN BE CREATED AND REFINED INSIDE THE SAME CONVERSATION.
An investigation can be converted into a persistent visual output without requiring the employee to rebuild the result manually inside a separate dashboarding application.
After completing an analysis, the user can ask Data to create an interactive dashboard containing selected metrics, comparisons, breakdowns, filters, and visualizations.
The dashboard can then be modified conversationally, including changes to charts, layout, metrics, or presentation style.
Companies can provide brand guidelines so that generated outputs follow the organization’s visual identity rather than using a generic reporting format.
The generated dashboard is also designed to remain operational after the first analysis because users can edit, share, and refresh it, and the Data plugin can help configure cloud automation to keep a published dashboard updated.
OpenAI Sites provides another publishing layer for organizations that want to distribute these outputs internally.
There is an important governance implication when a dashboard is published through Sites: the data used in the analysis is copied into the published site, so organizations need to consider the permissions of the recipients rather than assume that the original database permissions automatically govern every downstream artifact.
Data can also work directly with established BI platforms, including Power BI, Tableau, Sigma, ThoughtSpot, Omni, and Oracle BI.
This creates two possible output paths: a dashboard generated within the ChatGPT environment or an analysis that interacts with the organization’s existing BI tooling.
........
Data Studios workflow comparison: conventional business analysis versus the Data agent model.
Stage | Conventional workflow | Data agent workflow |
Business question | Sent to analyst or BI team | Asked directly in ChatGPT Work |
Data discovery | Analyst identifies tables and sources | Agent works across approved connected sources |
Metric interpretation | Analyst checks internal definitions | Semantic layer supplies business definitions |
Query construction | SQL or BI interface | Natural-language request |
Iteration | New query/report cycle | Follow-up in same conversation |
Visualization | Rebuilt in BI or spreadsheet | Interactive dashboard can be generated directly |
Distribution | Export, presentation, BI share | ChatGPT Sites or connected BI tools |
Refresh | Scheduled BI pipeline or manual update | Refresh plus cloud automation where configured |
........
This does not eliminate the underlying data warehouse, semantic model, or BI platform.
Instead, it adds an analytical control layer above them, reducing the amount of interface switching between the question, query, interpretation, visualization, and reporting stages.
··········
EXISTING TABLE, ROW, AND COLUMN PERMISSIONS STILL DEFINE WHAT THE AGENT CAN ACCESS.
Connecting a data source does not give ChatGPT broader access than the underlying account, which makes identity and workspace configuration part of the analytical architecture.
Workspace administrators determine which data connections are available and which roles or groups can use them.
When the Data agent queries a connected source, OpenAI says the request follows the existing permissions of the connected account, including table-level, row-level, and column-level restrictions where those controls are supported by the underlying system.
This distinction prevents installation of the Data plugin from functioning as an automatic bypass around warehouse governance.
An employee whose connected account cannot access a protected dataset should not gain that access simply because the query is expressed through ChatGPT.
Plugin availability and source authorization are also separate controls.
A workspace can have Data installed while a Snowflake, Databricks, or other required data-source connection still needs authorization or administrative configuration.
Some integrations require administrators to configure app templates before individual workspace members can use them, adding another policy layer between ChatGPT and the company’s systems.
The same principle applies when Data performs actions rather than analysis.
Findings can be prepared for distribution through channels such as Slack or email, while other connected tools can expose additional actions, but those operations depend on the capabilities of the relevant plugin, the user’s permissions, workspace policy, and any approval requirement attached to the action.
The Data agent is therefore not structured as an unrestricted autonomous process across enterprise systems.
Its effective capabilities depend on the intersection of available plugins, connected accounts, source permissions, workspace roles, and action approvals.
··········
DATA STUDIOS COUNTS AT LEAST FOUR DISTINCT ENTERPRISE LAYERS BETWEEN A QUESTION AND A SHAREABLE DASHBOARD.
The launch is easier to understand as an orchestration product than as a standalone analytics model, because the agent coordinates systems that previously required separate interfaces and permissions.
Based on the integrations explicitly described by OpenAI, the analytical path can involve at least four independent infrastructure layers before a result reaches another employee.
The first is the data layer, where structured records reside in systems such as Snowflake, BigQuery, Databricks, Redshift, ClickHouse, or MongoDB.
The second is the semantic layer, which gives meaning to those records through approved metric definitions, relationships, calculations, documentation, and organizational terminology.
The third is the analysis and visualization layer, where Data investigates changes and can create its own interactive output or interact with a connected BI platform.
The fourth is the distribution and action layer, where the result can be published through Sites, shared through communication tools, refreshed through automation, or used as the basis for an approved downstream action.
This four-layer mapping explains why the Data agent is materially different from a text-to-SQL feature.
Text-to-SQL addresses the translation between natural language and a database query.
The Data agent is intended to coordinate the broader chain from business question to governed data, company definition, analysis, dashboard, and ultimately distribution or action.
That expansion also increases the number of places where incorrect configuration can affect the result.
A perfectly generated query can still be misleading if it points to the wrong dataset, uses an outdated definition, inherits inappropriate permissions, or presents an analysis without sufficient context.
The quality of the system therefore depends as much on the company’s data architecture and semantic governance as on the reasoning capability of the underlying model.
··········
OPENAI IS USING THE SAME DATA-AGENT APPROACH INTERNALLY ACROSS PRODUCT AND GTM TEAMS.
The external launch follows broad internal adoption, giving OpenAI a production environment in which the workflow has already been used for everyday business analysis.
OpenAI says that nearly all of its product organization and more than two-thirds of its go-to-market organization use data agents in ChatGPT Work to analyze company data themselves.
That internal deployment required the company’s data team to create shared definitions, establish access rules, and implement safeguards for sensitive information, which reinforces the point that conversational analytics still depends on disciplined data governance underneath the interface.
The external alpha program has included organizations such as NTT DATA, Thermo Fisher Scientific, ServiceTitan, Zipline, Empower, CookUnity, Turing, and micro1, with reported uses covering sales analysis, spending, operational performance, reporting errors, customer behavior, staffing, and dashboard creation.
These examples indicate that OpenAI is positioning Data primarily as an enterprise productivity layer rather than as a specialist tool restricted to data scientists.
Employees in finance, sales, operations, product, and management can use the same interface while the underlying permissions and company definitions continue to control the data they receive.
The strongest productivity gain is likely to appear where organizations already have well-maintained warehouses, semantic layers, and access policies but still face a bottleneck between employees asking questions and analysts having time to investigate them.
Organizations with fragmented definitions or poor data quality will face a different constraint because making analytics conversational does not automatically reconcile conflicting metrics or repair unreliable source systems.
··········
DATA IS INSTALLED AS A PLUGIN IN CHATGPT WORK AND CAN ALSO BE USED FROM CODEX.
Administrators control deployment, while individual users can invoke the agent explicitly with @Data or allow ChatGPT to trigger it when the request requires data analysis.
The Data agent appears as Data in the ChatGPT Work Plugins directory.
Administrators can make it available to selected roles or groups, pre-install it for users, and separately configure the data-source plugins required to reach systems such as Snowflake or Databricks.
Once the required sources are connected, users can start a conversation with @Data and describe the business question, preferably including the relevant source, metric, period, and comparison when those details are already known.
The plugin can also be triggered implicitly, so users do not necessarily need to mention @Data on every request once the environment has been configured.
OpenAI’s current help documentation additionally states that the Data plugin can operate in Codex, extending the same data-analysis capabilities beyond the main ChatGPT Work interface.
OpenAI has not announced a separate standalone price for the Data agent in the launch material.
Practical availability instead depends on whether Data and the required source plugins are enabled for the user’s ChatGPT Work or Codex account and whether the organization has authorized the relevant enterprise connections.
That deployment model makes the product less comparable to a conventional per-seat BI application and more comparable to an analytical capability added to an existing AI workspace.
The strategic question for enterprises will therefore be whether this conversational layer can reduce the distance between company data and the employees making decisions without weakening the definitions, permission boundaries, and validation processes that make those data trustworthy in the first place.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
