OpenAI Launches a Data Agent for ChatGPT Work Without a Public Accuracy Benchmark

The new plugin is designed to turn questions across company data, documents and dashboards into analysis and approved follow-up work. The remaining test is whether enterprises trust its answers enough to widen those permissions.

By 3 min read
OpenAI Launches a Data Agent for ChatGPT Work Without a Public Accuracy Benchmark
OpenAI Launches a Data Agent for ChatGPT Work Without a Public Accuracy Benchmark

Listen to this story

The audio brief

About 1:25
0:001:25
Read transcript
OpenAI has added a Data agent to ChatGPT Work, giving employees one conversational way to investigate changes in business data, build interactive dashboards, and carry out approved follow-up actions. The product grew from an internal workflow OpenAI built around its own data, permissions, and reporting bottlenecks. OpenAI says nearly all of its product staff use data agents internally, along with more than two-thirds of its go-to-market organization. The pitch is straightforward: ask a business question in plain language instead of waiting for a report or learning a separate analytics tool. The agent can work across systems such as BigQuery, Databricks, Snowflake, Tableau, Power BI, Slack, Google Drive, and SharePoint. It uses an organization’s metric definitions, business terms, calculations, and data relationships, then lets users inspect the evidence, refine the analysis, and turn the result into a dashboard colleagues can edit, share, and refresh. Permissions remain with the connected systems, including table-, row-, and column-level restrictions. But there is a notable evidence gap. OpenAI has not published an external correctness or retrieval-accuracy benchmark for the Data agent. It cites an internal comparison, without releasing a result for the external product. So the immediate enterprise question is not just what the agent can connect to, but how much review customers require before its findings become recommendations, messages, or approved actions.

Story brief

3 key points

OpenAI has generalized an internal analytics workflow into a Data agent for ChatGPT Work, targeting questions that span business systems rather than replacing dedicated data platforms. The agent can investigate changes, expose supporting evidence, create collaborative dashboards, and execute approved actions through connected tools. Its adoption claims—nearly all product staff and over two-thirds of go-to-market...

  1. 01

    OpenAI reports nearly all product staff and more than two-thirds of go-to-market employees use data agents internally.

  2. 02

    Connections include Redshift, BigQuery, Databricks, Snowflake, Tableau, Power BI, Slack, Google Drive and SharePoint.

  3. 03

    Existing table-, row- and column-level permissions remain enforced by connected systems.

OpenAI has added a Data agent to ChatGPT Work that can investigate business-data changes, build interactive dashboards and carry out approved follow-up actions through connected tools. The launch brings more workplace systems into one conversational workflow, while leaving the product without a published external accuracy benchmark.

A tool built for OpenAI’s own data bottleneck

The external product grew from an internal tool OpenAI built around its own data, permissions and workflows, then generalized for other companies’ software. OpenAI says nearly all of its product team and more than two-thirds of its go-to-market organization use data agents in ChatGPT Work to analyze company data.

Its pitch is that employees should be able to ask a question in plain language rather than wait for a report or learn a separate analytics tool. The agent uses an organization’s business terms, metric definitions, custom calculations and data relationships to interpret results, according to OpenAI.

From a question to a shared dashboard

The agent connects to approved data sources and can bring Google Drive and SharePoint documents into its analysis. Users can examine evidence behind a finding, refine the work in conversation, and turn it into a dashboard that colleagues can edit, share and refresh.

Connected systems include

  • Data platforms including Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB and Snowflake.
  • Business-intelligence tools including Omni, Oracle BI, Power BI, Sigma, Tableau and ThoughtSpot.
  • Slack and email for sharing findings, plus connected tools for actions a user approves.

Permissions stay with the connected systems

Administrators decide which connections and roles are available. Queries enforce the connected account’s existing table, row and column restrictions. OpenAI says the agent does not create a permanent cross-tool context layer; it combines context from connected systems while it works.

OpenAI does not position the tool as a replacement for a company’s specialized data platform. Instead, it argues that ChatGPT Work is most useful when work crosses several systems. For customers, the next decision is how much review to require before an agent’s finding becomes a recommendation, message or approved action.

Editorial analysis

Our Read

OpenAI’s Data agent makes the workplace-agent promise more concrete: it is not only meant to retrieve information, but to carry a finding into a dashboard, message or approved action. That puts the practical burden on the data definitions and permissions companies already maintain. OpenAI’s lack of a public external accuracy benchmark does not establish poor performance, but it leaves buyers without a published measure of how the product handles the hard cases behind business decisions. The important next evidence will be whether OpenAI releases external evaluation results or customers disclose how they validate findings before acting on them.

Sources

  1. openai.comNow everyone can put data to work
  2. venturebeat.comOpenAI's new data agent skips a benchmark | VentureBeat

Loading discussion...