OpenAI Reveals Its Replicable Internal AI Agent
OpenAI has revealed that an internal "AI data agent," built by just two engineers, now serves 4,000 of its employees by automating data retrieval and analysis. The company claims the system is modular and can be replicated by other organizations, showcasing a model for rapidly scaling internal AI-first operations.
OpenAI's internal data agent, powered by GPT-5.2, operates across a massive 600 petabytes of data spread over 70,000 datasets. It serves over 3,500 internal users across various departments, including Engineering, Finance, and Research, transforming how they interact with data. The system is designed to function like a teammate, engaging in conversational back-and-forth to refine questions and carry context across multiple interactions. This allows employees to ask complex, open-ended questions in natural language and receive insights in minutes instead of days. To ensure accuracy, the agent is grounded in six layers of organizational context. This includes understanding table relationships, human-written annotations for business logic, and even institutional knowledge scraped from internal resources like Slack, Docs, and Notion. The agent's intelligence is further enhanced by using Codex, another OpenAI tool, to analyze the actual code that creates data pipelines. This provides a deeper semantic understanding of the data's origin and intended use that metadata alone cannot offer. For accessibility, the tool is integrated directly into existing employee workflows. It can be accessed through Slack, a web interface, IDEs, and OpenAI's internal ChatGPT application, meeting users where they already work. OpenAI treats this internal agent like a production-grade piece of software, complete with a continuous evaluation system. They use curated question-and-answer pairs with "golden" SQL queries to automatically catch regressions and ensure the agent's responses remain accurate over time.
