Inside OpenAI’s in-house data agent
OpenAI
OpenAI built its own internal AI agent for crunching company data, powered by GPT-5 and Codex. Instead of engineers waiting days for analytics, they now get answers in minutes.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI has a data problem, or rather, a data solution it's decided to talk about. The company built an in-house agent that lets employees ask plain-language questions about massive internal datasets and get back reasoned answers in minutes rather than days. It's powered by GPT-5 for reasoning, Codex for writing and running code, and a memory layer that lets it hold context across a session instead of starting from scratch every time.
The pitch here isn't just speed. Anyone can wire a chatbot up to a database and call it a day. The harder problem is reliability: getting an agent to pull the right tables, write correct queries, sanity-check its own output, and flag when it's unsure rather than confidently making something up. OpenAI says its agent handles multi-step analysis on its own, chaining together queries, running Python for statistical checks, and iterating when the first pass doesn't hold up. That kind of self-correction loop is the actual engineering achievement, not the natural-language interface, which at this point is table stakes.
Worth noting is who this was built for. This isn't a product OpenAI is shipping to customers today, it's a tool for its own analysts and researchers, the people who'd otherwise be stuck writing SQL by hand or waiting on a data science team's backlog. Using GPT-5 and Codex on your own infrastructure is also a convenient way to stress-test your own models on real, messy, high-stakes internal work before you ever put them in front of paying customers.
There's an obvious subtext: OpenAI is showing what its models can do on genuinely gnarly enterprise problems, the kind involving petabyte-scale logs and years of inconsistent schemas, not toy benchmarks. If the agent can be trusted to summarize usage trends or flag anomalies without a human double-checking every number, that's a meaningfully different claim than just being good at writing an email. Expect this kind of internal-agent framing to show up again soon, dressed up as an actual product for the enterprise data-analytics market.
My take — AI-written commentary, not fact-checked reporting
I'll believe the reliability claims when someone outside OpenAI gets to poke at this thing, because self-graded homework is still self-graded homework. That said, using your own frontier models to solve unglamorous internal plumbing is a smarter signal of real capability than another cherry-picked benchmark screenshot, and I'd like more labs to be this transparent about their own dogfooding instead of just the shiny demo reel.
Read more about this at: OpenAI