OpenAI built a custom AI data agent to help its 3,500+ internal employees quickly navigate a massive data platform with over 600 petabytes across 70,000 datasets. The agent, powered by GPT‑5.2, allows teams in Engineering, Data Science, Finance, and Research to ask questions in natural language and get answers in minutes instead of days. It combines Codex for table-level knowledge with organizational context, and uses a self-learning memory system that improves with every interaction. The tool addresses common pain points like finding the right tables, avoiding SQL bugs, and handling complex joins. OpenAI built it using the same tools available to external developers, including Codex, the Evals API, and the Embeddings API. Why it matters: This shows how even the creators of frontier models need to solve real data bottlenecks internally, proving that AI’s biggest impact often comes from reducing friction in everyday knowledge work.