OpenAI is expanding ChatGPT Work with a Data agent designed to make enterprise data analysis accessible even to those who do not routinely use business intelligence tools, databases, or query languages. The concept is simple in form, yet more complex in execution: a user can ask a question in natural language about company performance, drill down into the answers within the same conversation, and arrive at a shareable interactive dashboard.
The new agent targets requests that, in many organizations, end up in data team backlogs: understanding why sales have slowed, identifying growing spending areas, or checking factors that could jeopardize key customer renewals. ChatGPT Work is intended to handle the initial exploration, building analyses based on authorized sources and the operational context defined by the company.
The new feature, therefore, is not simply a generic capability to upload a spreadsheet and get a chart. OpenAI places the Data agent inside already governed data environments, connecting it both to repositories and analytics platforms and to semantic layers that define what metrics, calculations, and relationships between datasets mean. It is this component that determines whether an answer becomes useful for the business or remains merely a plausible statistical interpretation.
From databases to the shared language of the enterprise
Among the sources cited by OpenAI are Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB, and Snowflake. The agent can also incorporate files and documents stored in Google Drive and SharePoint into its analysis. Access to diverse sources is crucial because the information needed for a decision rarely resides in a single place: sales data might live in a data warehouse, product events in an operational database, and the definition of an indicator in an internal document.
To find its bearings, the Data agent can draw on corporate context available through semantic layers and trusted sources. OpenAI cites, among others, Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon, and BI dashboards. In practice, the system shouldn't just see table and column names: it can rely on the definitions an organization uses to establish, for example, which revenues count toward a metric, how a renewal rate is calculated, or what relationships exist between customers, contracts, and products.
This is a critical step. In enterprise data, the same term can have different definitions depending on the department, market, or time frame considered. An answer generated without understanding those conventions might be well-presented on the surface, yet based on the wrong denominator, an inconsistent time window, or an unapproved segmentation. Integrating with semantic layers aims to shift ChatGPT from an assistant that writes queries to an interface for a data model already curated by the company.
A conversational workflow, but with existing controls in place
OpenAI describes the Data agent as a tool to launch an investigation, request clarification on results, and verify the evidence supporting each finding. After the analysis, users can turn it into a dashboard with integrated visualizations. The dashboard can then be edited, shared, and updated by the team.
The promise is to close the gap between question and initial answer, particularly for managers and operational teams who currently rely on dedicated analysts for ad hoc inquiries outside recurring reports. This does not eliminate the work of data teams: the quality of the output still depends on clean data, consistent definitions, and properly configured connections. However, it can shift a portion of exploratory requests toward direct interaction, leaving specialists to handle more complex cases, data modeling, and the verification of sensitive results.
Because the agent handles potentially confidential information, OpenAI emphasizes its connection management model. Enterprise administrators determine which sources to make available and which roles can access them. Queries also respect existing permissions on the connected account, including table-, row-, and column-level restrictions. Therefore, an employee should not be able to obtain data from the Data agent that they cannot already see in the source platform.
This setup is necessary, but it does not automatically resolve every governance challenge. Access controls define the boundaries of what a user can view; it remains to be seen how individual organizations will manage shared dashboards, aggregated results, source documents, and interpretations generated by the model. The more an answer is used to back commercial, financial, or operational decisions, the more crucial it will be to keep the trail that generated it traceable and review it when necessary.
The challenge is reliability, not just simplicity
Conversational data interfaces are nothing new in enterprise software. The difference OpenAI is trying to build lies in combining language models, connections to major data systems, and an enterprise's semantic context. The partners cited at launch—from AWS to ClickHouse, Databricks, Snowflake, MongoDB, and G2—signal an intention to integrate directly where companies already store, process, or enrich their information.
For enterprises, the decisive factor will not merely be the ability to quickly generate charts. A system of this kind must distinguish correlation from causation, make assumptions explicit, return accurate timeframes and segments, and allow users to trace back the underlying evidence. OpenAI states that users can inspect the evidence behind results: this will be a key element to watch in real-world adoption, alongside how accurately the agent interprets business metrics.
The Data agent thus enters a segment where generative artificial intelligence is evaluated less on conversation quality and more on the trust it manages to earn with internal data. For many enterprises, getting an answer in minutes is only useful if that answer is reproducible, compliant with access rules, and understandable even to those who must turn it into a decision.
OpenAI presents the product as part of ChatGPT Work and defines it as a new Data agent. Its actual impact will depend on the connections each organization enables, the maturity of its data assets, and the availability of an up-to-date semantic context. Where these foundations exist, conversation can become a more immediate gateway to analysis; where they are missing, no prompt can replace the preliminary work of data organization, quality, and governance.
