OpenAI's AI data agent, built by two engineers, now serves 4,000 employees — and the company says anyone can replicate it

OpenAI has developed an internal AI data agent that allows employees to analyze data using plain language prompts. This agent, built in three months with 70% AI-generated code, provides access to 600 petabytes of data across 70,000 datasets. It saves analysts significant time by generating charts and reports in minutes, improving access to insights. The agent is built on GPT-5.2 and is used by over 4,000 employees daily across various departments. Codex, OpenAI's AI coding tool, plays a key role in mapping data tables and generating code for the agent. The agent uses multiple context layers, including schema metadata and institutional knowledge, to enhance its performance. Prompt engineering is used to encourage the agent to validate data sources, preventing overconfidence. Safety is ensured through access controls, user feedback, and model self-evaluation, with no plans to commercialize the tool. OpenAI's strategy is to provide building blocks for enterprises to create their own AI agents. Good data governance and clean, annotated data are crucial for the effectiveness of such data agents. Data agents serve as a new, more autonomous and accessible entry point for data intelligence.