OpenAI's Internal AI Agent Serves 4,000 Staff
OpenAI revealed its internal “AI data agent,” a tool for automating data workflows, has grown from a two-engineer project to a system that now serves 4,000 employees. The company claims the system is built on open principles and can be replicated by other organizations.
OpenAI's internal data agent operates on a staggering scale, serving over 3,500 employees by navigating 600 petabytes of data spread across 70,000 distinct datasets. This conversational AI, powered by GPT-5.2, allows teams in engineering, finance, research, and go-to-market to get answers to complex data questions in minutes instead of days. The agent's accuracy hinges on a sophisticated six-layer approach to context. This system integrates metadata from schemas, historical user query patterns, and human annotations that explain business-specific nuances. It also enriches its understanding by analyzing the underlying code that creates data tables, gleaning insights into how data is actually constructed. To grasp institutional knowledge, the agent can access and process information from internal resources like Slack, Google Docs, and Notion. This allows it to understand the context behind the data, such as product launches or internal codenames. A crucial memory layer enables the agent to learn from corrections, so if one analyst teaches it a specific data filter, that knowledge becomes available to everyone. The system is designed to function less like a rigid tool and more like a collaborative teammate. It is accessible through various platforms already used by employees, including Slack and developer IDEs. If a query is ambiguous, the agent asks clarifying questions rather than silently failing or providing a potentially incorrect answer. While OpenAI built the agent for its own specific workflows, it was constructed using the company's publicly available tools, such as the GPT-5.2 model, Codex, and various APIs. The agent's security model is built to be transparent, inheriting the existing permissions of the user, ensuring employees can only query data they are already authorized to access. This internal tool continuously improves through a process of "unit testing for AI reasoning." OpenAI maintains a set of "golden" questions with verified correct SQL queries and results. The agent's generated answers are regularly compared against these benchmarks to catch any regressions in its performance before they impact users.
