OpenAI's Internal Data Agent Serves 4,000 Staff
OpenAI revealed that a custom AI data agent, built by just two engineers, now serves 4,000 of its own employees. The company claims the agent is so generalizable that 'anyone can replicate it,' hinting at a future of scalable, off-the-shelf AI tools for internal corporate data analysis.
The internal data agent is designed to navigate OpenAI's massive data platform, which serves over 3,500 employees and manages 600 petabytes of data across 70,000 datasets. This vast scale previously made it time-consuming for employees to simply locate the correct data table for analysis. The tool allows staff in engineering, finance, and research to get insights from this data in minutes instead of days. Powered by GPT-5.2, the agent utilizes natural language, allowing employees to ask complex questions without writing manual SQL queries. It is accessible through existing workflows, including as a Slack agent, a web interface, and directly within OpenAI's internal ChatGPT application. This integration is key to its adoption, meeting users in the interfaces they already use. A key feature of the agent is its ability to self-correct and learn. It runs in an observable loop, interpreting questions, generating and executing queries, detecting errors, and then repairing its own mistakes without human intervention. The system also has a memory function that stores corrections to avoid repeating errors. To ensure accuracy, the agent is grounded in six distinct layers of context. These layers include historical query patterns, human annotations for business meaning, knowledge extracted from internal documentation on platforms like Slack and Notion, and a memory system that learns from past corrections. This data agent is part of a broader suite of internal AI tools at OpenAI. Other custom agents include a "GTM Assistant" for sales teams, trained on the techniques of a top-performing salesperson, and a "Support Agent" that has improved customer support ticket deflection rates.
