OpenAI’s Data agent for ChatGPT Work lets employees ask questions about company data, investigate changes, and turn analysis into interactive dashboards. It connects to approved data sources and files and can use an organization’s metric definitions and semantic layers. It is not an automatic replacement for analysts: connectors, definitions, and permissions determine whether its answers can be trusted. Read OpenAI’s Data agent announcement.
What data it can use
The announcement lists BigQuery, ClickHouse, Databricks, Snowflake, Amazon Redshift, Datadog, MongoDB, and other sources, as well as files from Google Drive and SharePoint. Context can come from dbt, GitHub, Snowflake Horizon, and BI dashboards to explain company metrics and relationships. Administrators choose available connectors and roles; queries are intended to honor the connected account’s table-, row-, and column-level access.
A sensible pilot workflow
- Have an administrator enable a small number of read-only connectors in a test workspace and confirm access to sensitive fields.
- Start with a specific question, such as “How much did renewal rate change this month, and which customer groups account for the change?”
- Ask the agent to identify the tables, date range, and metric definitions it used, then investigate unusual trends.
- Once the calculation is verified, turn it into an interactive dashboard and have the metric owner check the figures before sharing.
The agent can also work with tools such as Tableau, Power BI, Sigma, and ThoughtSpot, or take approved follow-up actions through connected services. Whether it can write to a particular system depends on the connector and permissions; the product name does not imply unrestricted access to every data source.
What determines answer quality
Natural-language questions are easier to ask, but they do not automatically standardize terms such as “active customer,” “revenue,” or “churn.” If the data model lacks semantic definitions, or departments use different definitions, the agent may produce a fluent answer with inconsistent logic. Prepare questions with known answers and compare them with existing reports. Check source references, date filters, aggregation, and row-level access.
Permissions and governance
Administrators control which connectors and roles are available, and queries inherit the connected account’s access. Before a pilot, decide how data is retained, how sensitive information is handled, what can be exported, and how activity is audited. The fact that the agent can read a company report does not mean it should be shared with every employee or an external recipient.