OpenAI's AI data agent now powers 4,000 staff, engineers say anyone can replicate it

OpenAI’s data‑agent, a GPT‑5.2‑powered assistant that translates plain‑English prompts into charts, dashboards and narrative reports, was built by just two engineers in a three‑month sprint, with roughly 70 % of its code generated by AI itself, according to Emma Tang, head of data infrastructure, in an exclusive interview with VentureBeat. The tool now runs on every major internal channel—Slack, a web UI, IDE extensions, the Codex CLI and the company’s internal ChatGPT app—so that analysts can ask a question like “show revenue by region for Q4 2025” and receive a polished visualization in minutes. Tang’s team estimates the agent saves two to four hours per query, but she stresses that the deeper impact is the democratization of analysis: “Engineers, growth, product, as well as non‑technical teams… can now pull sophisticated insights on their own,” she said.

The scale of OpenAI’s data problem underscores why the agent matters. The company’s data platform stores more than 600 petabytes across roughly 70 000 distinct datasets, a volume that would normally require hours of schema hunting and SQL writing for a single analyst, VentureBeat reports. With a workforce of about 5 000 employees, OpenAI’s internal data‑tooling team already supports over 4 000 daily users; the new agent extends that reach by allowing virtually any employee to query the entire data lake without needing to know table names or column definitions. Tang notes that the agent is “used for any kind of analysis” and that “almost every team in the company uses it,” highlighting its role as a universal front‑end to the otherwise unwieldy data stack.

Beyond routine reporting, the agent has proven its value in multi‑step, investigative workflows. Tang described a recent incident where a finance analyst noticed a mismatch between two dashboards tracking Plus‑subscriber growth. By feeding the discrepancy into the agent, the analyst received a series of ranked charts that isolated five distinct factors contributing to the variance—insights that would have taken a data scientist days to surface manually. Similar use cases span revenue breakdowns, latency debugging and product‑performance retrospectives, illustrating the tool’s versatility across both technical and business domains.

OpenAI’s claim that “anyone can replicate” the agent rests on the open‑source‑friendly methodology the team employed. The engineers leveraged the same GPT‑5.2 model that powers OpenAI’s external products, combined with internal APIs for data discovery and visualization, and documented the pipeline in a public blog post. While the company has not released the full codebase, the architecture—natural‑language parsing, automated schema inference, and chart generation—mirrors patterns emerging in the broader AI‑agent ecosystem, as noted by TechCrunch’s coverage of OpenAI’s enterprise‑agent platform. Tang emphasizes that the bottleneck for smarter organizations is not model quality but data accessibility, a point that aligns with industry analysts who see internal AI agents as the next frontier for productivity gains.

The rapid internal adoption—4 000 of 5 000 employees daily—makes OpenAI’s deployment one of the most aggressive AI‑agent rollouts in any enterprise, according to VentureBeat. The company’s broader AI strategy, which includes recent enhancements to its voice agent for developers (ZDNet) and plans to monetize enterprise AI tools (TechCrunch), suggests that the data‑agent is both a proof‑of‑concept and a template for future products. If other firms can emulate the three‑month, two‑engineer build process, the competitive landscape could shift quickly, turning data‑accessibility into a commodity service rather than a bespoke internal capability.