Technology
OpenAI Launches a Data Agent for Enterprise Analytics in ChatGPT Work
The new agent connects ChatGPT Work to warehouses, semantic layers and business-intelligence tools, letting employees investigate metrics and build dashboards while preserving source permissions.
By Patrick T ·
SAN FRANCISCO. OpenAI has launched a Data agent in ChatGPT Work that connects to company warehouses, documents, semantic layers and business-intelligence systems, allowing employees to investigate metrics, build interactive dashboards and propose follow-up actions through a conversation. The product moves ChatGPT deeper into work traditionally divided among analysts, data engineers and BI software.
The agent supports approved sources including Amazon Redshift, BigQuery, ClickHouse, Databricks, MongoDB and Snowflake, as well as files in Google Drive and SharePoint. OpenAI says queries enforce the connected account's table, row and column permissions. That detail is more important than the natural-language interface: an analytics agent becomes useful only when it can reach governed data without turning broad access into a shortcut around existing controls.
Analytics Moves From Answers to Investigation
The product is designed for follow-up questions rather than a single generated chart. A user can ask why sales slowed, inspect the evidence, refine the analysis and publish a dashboard. The agent can use business definitions and data relationships from sources such as dbt, Snowflake Horizon and Databricks Genie Ontology. Those semantic layers help distinguish revenue from bookings or active customers from registered accounts, terms that sound simple but vary across companies.
OpenAI says the agent can create or interact with dashboards in Power BI, Tableau, Sigma, Omni, Oracle BI and ThoughtSpot. It can then recommend next steps, identify stakeholders and, after approval, share findings through Slack or email. That sequence turns analysis into an operational workflow. It also makes approval design essential because a flawed interpretation can travel farther and faster than a mistaken private query.
The appeal is obvious. Business users often wait for an analyst because the underlying query requires technical knowledge and the metric requires institutional context. A conversational agent can shorten that queue. The danger is that speed disguises ambiguity. Two plausible filters can produce different retention rates, and an attractive visualization can make an uncertain conclusion look settled.
OpenAI says users can review evidence behind each finding. Enterprises should test what that means in practice. A credible result should identify the tables, time range, joins, filters, transformations and assumptions used. It should also preserve the generated query or equivalent execution plan so an analyst can reproduce the result outside the conversation.
Permission Inheritance Is Necessary but Not Sufficient
Respecting source permissions prevents an employee from directly querying rows they cannot access. It does not resolve inference risk. An agent may combine allowed aggregates, documents and prior answers to reveal something sensitive. It may also place restricted findings into a dashboard with broader sharing settings. Governance has to cover the full path from retrieval to generated artifact and outbound action.
OpenAI says nearly all of its product team and more than two-thirds of its go-to-market organization use the underlying capabilities internally. The company built shared definitions, access rules and safeguards for sensitive data. That experience is relevant, but customer environments will be less uniform. Acquisitions, regional systems and competing metric definitions can make the semantic layer the hardest part of deployment.
The competitive field is crowded. Data platforms and BI vendors already offer conversational analysis, often with deeper knowledge of their own metadata. OpenAI is betting that ChatGPT can become the neutral work surface across many systems. Partners gain distribution, while OpenAI gains a position above the warehouse and dashboard. Customers gain choice but must decide which layer owns audit logs and policy.
For data teams, the product should change priorities rather than eliminate work. Fewer routine dashboard requests can free analysts for experimental design and harder investigations. At the same time, more employees asking more questions will expose weak definitions and inconsistent pipelines. The agent can democratize access only to the quality that already exists underneath it.
The Data agent is a consequential release because it connects a general assistant to the numerical record used to run a business. Its success will not be measured by how quickly it draws a chart. It will be measured by whether teams can reproduce the reasoning, challenge the assumptions and act without losing control of the evidence.
Topics: OpenAI, ChatGPT Work, data analytics, business intelligence, AI agents