OpenAI's Two-Person AI Agent Serves 4,000 Staff
OpenAI revealed its in-house "AI data agent," a tool built by just two engineers, now serves 4,000 employees. The company claims the system, which performs autonomous data retrieval and synthesis, is easily replicable, suggesting a path for other firms to democratize advanced data management.
The in-house AI data agent at OpenAI tackles a massive data landscape, serving over 3,500 internal users across 600 petabytes of data and 70,000 datasets. This complex environment makes simply finding the correct data table a significant challenge, a problem the agent was specifically designed to solve. The tool is designed to feel like a collaborative teammate, allowing for interactive and iterative data exploration. Built using OpenAI's own suite of tools, including a GPT-5 model, Codex, and various APIs, the agent is more than just a query machine. It integrates company-specific knowledge and context, allowing it to understand the nuances of internal data. This "institutional knowledge" is one of six context layers that ground the agent in organizational reality, enabling it to write the *right* SQL queries. A key feature of the data agent is its ability to self-correct and learn. If an initial query leads to an error, like a bad join with zero rows returned, the agent can investigate the issue, adjust its approach, and try again without direct user intervention. It also possesses a memory function, storing corrections to avoid repeating mistakes. To ensure the reliability of its outputs, OpenAI continuously evaluates the agent against a set of "golden" SQL queries with verified results. This automated testing helps catch any regressions before they impact employees' work. Security and privacy are managed by inheriting existing user permissions, meaning employees can only query data they already have access to. The agent is integrated directly into the existing workflows and communication tools used by OpenAI employees. This approach is intended to make data analysis a natural part of an employee's day, rather than a task requiring a separate, specialized tool. Teams across various departments, including engineering, finance, and research, use the agent to get insights from data in minutes instead of days. The development of such internal AI agents is a growing trend, with some engineers at OpenAI now using hundreds of billions of tokens weekly across multiple agents to build software. This signals a shift in software development from manual coding to guiding teams of AI agents. While this specific data agent is an internal tool, its creation and success demonstrate a blueprint for how other companies might leverage AI for their own data analysis needs.
