OpenAI developed an in-house AI data agent to streamline how employees access and analyze data. This custom tool harnesses internal data, permissions, and workflows, allowing teams across various functions—like Engineering, Data Science, and Finance—to get insights in minutes rather than days. The agent leverages OpenAI's tools, including Codex and GPT-5, to answer complex data questions using natural language, which increases efficiency and reduces reliance on specialized data teams. The agent’s design focuses on context, employing a multi-layered approach to ensure accuracy in its outputs. It utilizes schema metadata and historical queries to understand how to structure SQL queries effectively. Human annotations from domain experts provide further clarity on the meaning and purpose of data tables, while Codex enhances the agent's comprehension of data nuances. Additionally, it integrates institutional knowledge from platforms like Slack and Google Docs, which helps the agent grasp internal terminologies and company context. One of the agent’s key features is its self-learning capability. It evaluates its own results, identifies mistakes, and adjusts its methods in real time, allowing for continuous improvement in data analysis. If it encounters issues, like a zero-row result from a faulty query, it can troubleshoot and correct its approach without user intervention. This closed-loop learning process results in faster, more reliable analyses compared to traditional manual workflows. Overall, the agent represents a significant step forward in data accessibility and usability within OpenAI.