Inside OpenAI's In-House Data Agent

As data scale grew rapidly, OpenAI built a self-developed in-house AI data agent to reduce the time spent on data exploration and analysis. This tool is not intended for external release; it is an internal-only tool optimized for OpenAI's data structure, permissions, and workflows.

OpenAI currently holds over 600PB of data and 70,000 datasets, used by more than 3,500 internal users. Previously, a great deal of time was spent identifying the right tables or debugging logical errors in complex SQL queries.

The agent is based on GPT-5.2 and leverages OpenAI's core technologies including Codex, the Evals API, and the Embeddings API. Users can ask questions in natural language via Slack, a web interface, an IDE, or the Codex CLI.

The agent's core capabilities are end-to-end analysis and self-learning ability. It handles the entire process on its own, from understanding the question to data exploration, query execution, and result synthesis. In particular, through a closed-loop process in which it investigates the cause itself and revises its approach to retry when it determines an intermediate result is incorrect, it improves the accuracy of its analysis.