OpenAI built an internal AI data agent with a two-person engineering team in three months, and it now serves more than 4,000 of the company’s roughly 5,000 employees daily. The tool, which runs on GPT-4.5 and connects to Slack, web interfaces, IDEs, and OpenAI’s internal ChatGPT app, accepts plain-English questions and returns charts, dashboards, and analytical reports in minutes.
The scale of the underlying data problem makes the achievement notable. OpenAI’s data platform spans more than 600 petabytes across 70,000 datasets. Before the agent existed, a finance analyst looking to compare revenue across geographies and customer cohorts could spend hours just identifying the right table, writing SQL queries, and verifying schemas. The same task now takes a single Slack message.
Built Fast, Mostly by AI
Seventy percent of the agent’s code was written by AI. The two engineers who built it completed the project in roughly three months. Emma Tang, head of data infrastructure at OpenAI, said the tool saves two to four hours of work per query, though she acknowledged the larger benefit is harder to quantify: the agent gives employees access to analysis they simply could not have performed on their own, regardless of available time.
“Engineers, growth, product, as well as non-technical teams, who may not know all the ins and outs of the company data systems and table schemas” can now pull sophisticated insights independently, according to Tang’s team.
What It Actually Does
The use cases span the organization. OpenAI’s finance team queries it for revenue comparisons across geographies and customer cohorts. Product managers use it to track feature adoption. Engineers ask it whether a specific ChatGPT component is slower than the previous day and, if so, which latency factors explain the change.
Tang described one telling example. A user noticed discrepancies between two dashboards tracking Plus subscriber growth. The agent broke down the differences factor by factor. “There turned out to be five different factors,” Tang said. “For a human, that would take hours, if not days, but the agent can do it in a few minutes.”
What separates this deployment from most enterprise AI efforts is its horizontal reach. Most corporate AI agents operate within a single department. OpenAI’s cuts across all of them. A senior leader can combine sales figures, engineering metrics, and product analytics in one query. “That’s a really unique feature of ours,” Tang said.
The Hardest Problem: Finding the Right Data
With 70,000 datasets in play, locating the correct table remains the agent’s most persistent technical challenge. Tang was direct about it. “That’s the biggest problem with this agent,” she said. The company has pointed to its Codex coding agent as part of the solution, using it to help navigate the data catalog more reliably.
The team launched the tool department by department, building specific memory and context for each group before eventually connecting everything into a shared system. The gradual rollout helped surface edge cases and calibrate the agent’s reliability before broader exposure.
A Signal for Enterprise Data Strategy
OpenAI’s internal experience points to a finding that goes beyond any single tool: the limiting factor for AI-driven organizations is not model capability. It is data infrastructure. The bottleneck Tang’s team spent three months solving, navigating massive, poorly catalogued internal datasets, is one nearly every large organization faces.
Tang confirmed that OpenAI intends the project to serve as a replicable blueprint. Any enterprise with the right data foundations, she suggested, can build something similar.
Photo by Zac Wolff on Unsplash
This article is a curated summary based on third-party sources. Source: Read the original article
