OpenAI's Internal AI Agent Serves 4,000 Staff
OpenAI revealed its internal AI data agent, built by just two engineers, now serves 4,000 employees by automating data access and analytics. The company claims the system's architecture is replicable, offering a potential blueprint for other enterprises to build their own powerful internal AI assistants.
The internal data agent at OpenAI, detailed in a report by engineers Bonnie Zu, Aravvin Suresh, and Emma Tang, manages an immense data environment. It serves over 3,500 internal users, providing insights from a staggering 600 petabytes of data spread across 70,000 datasets. This scale of data previously made finding the right information a significant bottleneck for analysts. Built on OpenAI's own technology, the agent is powered by GPT-5.2 and was constructed using tools available to developers, including Codex, the Evals API, and the Embeddings API. This allows the agent to understand natural language queries and reason over OpenAI's vast data platform. The agent is designed to function like a human teammate, engaging in conversational back-and-forth to refine questions and correct its own mistakes mid-analysis without user intervention. If the agent heads down an incorrect path, employees can interrupt and redirect it, allowing for iterative exploration of data. Its accuracy is grounded in six distinct layers of context that provide a deep understanding of the data. These layers include metadata, human annotations, insights from source code via Codex, institutional knowledge from documents, a memory of past corrections, and live validation. To ensure reliability, OpenAI continuously evaluates the agent's performance using a system of curated "golden" SQL queries. These automated tests compare the AI-generated results against verified benchmarks, allowing the system to catch regressions before they impact users. The tool is not a standalone application but is integrated directly into existing employee workflows. Staff across engineering, finance, and research can access the agent via Slack, a web interface, integrated development environments (IDEs), and directly within OpenAI's internal ChatGPT application.
