OpenAI is finally bridging the gap between natural language queries and enterprise data warehouses. The new Data agent in ChatGPT Work allows employees to ask plain-English questions about company metrics and receive interactive dashboards as answers. Instead of waiting days for a business intelligence team to generate a report, users can directly query approved sources like Amazon Redshift, Google BigQuery, Snowflake, and Databricks. This shift moves analytics from a specialized skill set to a conversational interface, potentially unlocking insights for non-technical staff across sales, marketing, and operations.
The system doesn’t just pull raw numbers; it interprets them through your organization’s specific semantic layer. By integrating with tools like dbt, GitHub, and Snowflake Horizon, the agent understands custom business definitions and metric relationships. This context awareness is critical. Without it, AI models often hallucinate or misinterpret column headers. OpenAI claims this setup lets users diagnose why weekly active users dropped or identify spending spikes without writing a single line of SQL. The output isn’t static, either. Users can refine the analysis in real-time and export results to existing BI platforms like Tableau, Power BI, and ThoughtSpot.
Security remains the primary hurdle for enterprise AI adoption, and OpenAI has structured this release to address IT concerns. Administrators retain full control over which data connections are active and which roles can access them. Crucially, the agent enforces existing row-level and column-level permissions from the connected databases. If a user doesn’t have access to sensitive salary data in Snowflake, the Data agent won’t reveal it. This permission mirroring ensures that the AI acts as a conduit for existing governance policies rather than bypassing them. Early adopters in the Alpha program, including NTT Data and Thermo Fisher, report using the tool to catch reporting errors and prioritize sales opportunities.
Internal usage at OpenAI suggests strong product-market fit within technical organizations. Nearly all product team members and over two-thirds of the go-to-market staff now use these agents daily. This internal dogfooding validates the workflow before wider release. However, the real test will be adoption among less data-literate teams. If marketing managers can genuinely trust the agent’s diagnosis of a churn spike without double-checking the underlying query logic, this could redefine how companies operate. The friction of data access just dropped significantly, but the responsibility for interpreting those answers still rests with humans.
