When an OpenAI finance analyst wanted to check income throughout geographies and buyer cohorts final yr, it took hours of labor — looking by 70,000 datasets, writing SQL queries, verifying desk schemas. At the moment, the identical analyst varieties a plain-English query into Slack and will get a completed chart in minutes.
The software behind that transformation was constructed by two engineers in three months. Seventy % of its code was written by AI. And it’s now utilized by greater than 4,000 of OpenAI's roughly 5,000 workers each day — making it one of the aggressive deployments of an AI knowledge agent inside any firm, anyplace.
In an unique interview with VentureBeat, Emma Tang, the top of knowledge infrastructure at OpenAI whose workforce constructed the agent, provided a uncommon look contained in the system — the way it works, the way it fails, and what it indicators about the way forward for enterprise knowledge. The dialog, paired with the corporate's weblog submit saying the software, paints an image of an organization that turned its personal AI on itself and found one thing that each enterprise will quickly confront: the bottleneck to smarter organizations isn't higher fashions. It's higher knowledge.
"The agent is used for any form of evaluation," Tang stated. "Virtually each workforce within the firm makes use of it."
A plain-English interface to 600 petabytes of company knowledge
To grasp why OpenAI constructed this method, think about the size of the issue. The corporate's knowledge platform spans greater than 600 petabytes throughout 70,000 datasets. Even finding the proper desk can devour hours of an information scientist's time. Tang's Information Platform workforce — which sits below infrastructure and oversees massive knowledge programs, streaming, and the info tooling layer — serves a staggering inner person base. "There are 5,000 workers at OpenAI proper now," Tang stated. "Over 4,000 use knowledge instruments that our workforce gives."
The agent, constructed on GPT-5.2 and accessible wherever workers already work — Slack, an online interface, IDEs, the Codex CLI, and OpenAI's inner ChatGPT app — accepts plain-English questions and returns charts, dashboards, and long-form analytical reviews. In follow-up responses shared with VentureBeat on background, the workforce estimated it saves two to 4 hours of labor per question. However Tang emphasised that the bigger win is tougher to measure: the agent provides folks entry to evaluation they merely couldn't have achieved earlier than, no matter how a lot time they’d.
"Engineers, development, product, in addition to non-technical groups, who might not know all of the ins and outs of the corporate knowledge programs and desk schemas" can now pull refined insights on their very own, her workforce famous.
From income breakdowns to latency debugging, one agent does all of it
Tang walked by a number of concrete use circumstances that illustrate the agent's vary. OpenAI's finance workforce queries it for income comparisons throughout geographies and buyer cohorts. "It could, simply actually in plain textual content, ship the agent a question, and it is going to be in a position to reply and provide you with charts and provide you with dashboards, all of this stuff," she stated.
However the actual energy lies in strategic, multi-step evaluation. Tang described a latest case the place a person noticed discrepancies between two dashboards monitoring Plus subscriber development. "The info agent can provide you a chart and present you, stack rank by stack rank, precisely what the variations are," she stated. "There turned out to be 5 various factors. For a human, that may take hours, if not days, however the agent can do it in a couple of minutes."
Product managers use it to know function adoption. Engineers use it to diagnose efficiency regressions — asking, as an illustration, whether or not a selected ChatGPT element actually is slower than yesterday, and in that case, which latency parts clarify the change. The agent can break all of it down and examine prior intervals from a single immediate.
What makes this particularly uncommon is that the agent operates throughout organizational boundaries. Most enterprise AI brokers right this moment are siloed inside departments — a finance bot right here, an HR bot there. OpenAI's cuts horizontally throughout the corporate. Tang stated they launched division by division, curating particular reminiscence and context for every group, however "sooner or later it's all in the identical database." A senior chief can mix gross sales knowledge with engineering metrics and product analytics in a single question. "That's a very distinctive function of ours," Tang stated.
How Codex solved the toughest drawback in enterprise knowledge
Discovering the fitting desk amongst 70,000 datasets is, by Tang's personal admission, the only hardest technical problem her workforce faces. "That's the largest drawback with this agent," she stated. And it's the place Codex — OpenAI's AI coding agent — performs its most creative function.
Codex serves triple responsibility within the system. Customers entry the info agent by Codex by way of MCP. The workforce used Codex to generate greater than 70% of the agent's personal code, enabling two engineers to ship in three months. However the third function is probably the most technically fascinating: a day by day asynchronous course of the place Codex examines essential knowledge tables, analyzes the underlying pipeline code, and determines every desk's upstream and downstream dependencies, possession, granularity, be part of keys, and comparable tables.
"We give it a immediate, have Codex take a look at the code and reply with what we want, after which persist that to the database," Tang defined. When a person later asks about income, the agent searches a vector database to search out which tables Codex has already mapped to that idea.
This "Codex Enrichment" is certainly one of six context layers the agent makes use of. The layers vary from primary schema metadata and curated knowledgeable descriptions to institutional information pulled from Slack, Google Docs, and Notion, plus a studying reminiscence that shops corrections from earlier conversations. When no prior info exists, the agent falls again to dwell queries in opposition to the info warehouse.
The workforce additionally tiers historic question patterns. "All question historical past is all people's 'choose star, restrict 10.' It's probably not useful," Tang stated. Canonical dashboards and government reviews — the place analysts invested important effort figuring out the proper illustration — get flagged as "supply of fact." The whole lot else will get deprioritized.
The immediate that forces the AI to decelerate and assume
Even with six context layers, Tang was remarkably candid concerning the agent's greatest behavioral flaw: overconfidence. It's an issue anybody who has labored with giant language fashions will acknowledge.
"It's a very massive drawback, as a result of what the mannequin usually does is really feel overconfident," Tang stated. "It'll say, 'That is the fitting desk,' and simply go forth and begin doing evaluation. That's truly the improper strategy."
The repair got here by immediate engineering that forces the agent to linger in a discovery section. "We discovered that the extra time it spends gathering potential eventualities and evaluating which desk to make use of — simply spending extra time within the discovery section — the higher the outcomes," she stated. The immediate reads virtually like teaching a junior analyst: "Earlier than you run forward with this, I really need you to do extra validation on whether or not that is the fitting desk. So please verify extra sources earlier than you go and create precise knowledge."
The workforce additionally discovered, by rigorous analysis, that much less context can produce higher outcomes. "It's very straightforward to dump all the things in and simply anticipate it to do higher," Tang stated. "From our evals, we truly discovered the other. The less stuff you give it, and the extra curated and correct the context is, the higher the outcomes."
To construct belief, the agent streams its intermediate reasoning to customers in actual time, exposes which tables it chosen and why, and hyperlinks on to underlying question outcomes. Customers can interrupt the agent mid-analysis to redirect it. The system additionally checkpoints its progress, enabling it to renew after failures. And on the finish of each process, the mannequin evaluates its personal efficiency. "We ask the mannequin, 'how did you assume that went? Was that good or unhealthy?'" Tang stated. "And it's truly pretty good at evaluating how nicely it's doing."
Guardrails which are intentionally easy — and surprisingly efficient
In relation to security, Tang took a practical strategy that will shock enterprises anticipating refined AI alignment strategies.
"I feel you simply should have much more dumb guardrails," she stated. "We’ve actually robust entry management. It's at all times utilizing your private token, so no matter you have got entry to is just what you have got entry to."
The agent operates purely as an interface layer, inheriting the identical permissions that govern OpenAI's knowledge. It by no means seems in public channels — solely in non-public channels or a person's personal interface. Write entry is restricted to a short lived check schema that will get wiped periodically and may't be shared. "We don't let it randomly write to programs both," Tang stated.
Consumer suggestions closes the loop. Workers flag incorrect outcomes instantly, and the workforce investigates. The mannequin's self-evaluation provides one other verify. Long term, Tang stated, the plan is to maneuver towards a multi-agent structure the place specialised brokers monitor and help one another. "We're transferring in direction of that ultimately," she stated, "however proper now, whilst it’s, we've gotten fairly far."
Why OpenAI gained't promote this software — however needs you to construct your individual
Regardless of the plain industrial potential, OpenAI advised VentureBeat that the corporate has no plans to productize its inner knowledge agent. The technique is to offer constructing blocks and let enterprises assemble their very own. And Tang made clear that all the things her workforce used to construct the system is already out there externally.
"We use all the identical APIs which are out there externally," she stated. "The Responses API, the Evals API. We don't have a fine-tuned mannequin. We simply use 5.2. So you possibly can undoubtedly construct this."
That message aligns with OpenAI's broader enterprise push. The corporate launched OpenAI Frontier in early February, an end-to-end platform for enterprises to construct and handle AI brokers. It has since enlisted McKinsey, Boston Consulting Group, Accenture, and Capgemini to assist promote and implement the platform. AWS and OpenAI are collectively creating a Stateful Runtime Setting for Amazon Bedrock that mirrors among the persistent context capabilities OpenAI constructed into its knowledge agent. And Apple lately built-in Codex instantly into Xcode.
In accordance with info shared with VentureBeat by OpenAI, Codex is now utilized by 95% of engineers at OpenAI and evaluations all pull requests earlier than they're merged. Its international weekly lively person base has tripled because the begin of the yr, surpassing a million. Total utilization has grown greater than fivefold.
Tang described a shift in how workers use Codex that transcends coding fully. "Codex isn’t even a coding software anymore. It's rather more than that," she stated. "I see non-technical groups use it to prepare ideas and create slides and to create day by day summaries." One in all her engineering managers has Codex assessment her notes every morning, establish an important duties, pull in Slack messages and DMs, and draft responses. "It's actually working on her behalf in numerous methods," Tang stated.
The unsexy prerequisite that can decide who wins the AI agent race
When requested what different enterprises ought to take away from OpenAI's expertise, Tang didn't level to mannequin capabilities or intelligent immediate engineering. She pointed to one thing way more mundane.
"This isn’t attractive, however knowledge governance is basically essential for knowledge brokers to work nicely," she stated. "Your knowledge must be clear sufficient and annotated sufficient, and there must be a supply of fact someplace for the agent to crawl."
The underlying infrastructure — storage, compute, orchestration, and enterprise intelligence layers — hasn't been changed by the agent. It nonetheless wants all of these instruments to do its job. However it serves as a basically new entry level for knowledge intelligence, one that’s extra autonomous and accessible than something that got here earlier than it.
Tang closed the interview with a warning for corporations that hesitate. "Firms that undertake this are going to see the advantages very quickly," she stated. "And corporations that don't are going to fall behind. It's going to drag aside. The businesses who use it are going to advance very, in a short time."
Requested whether or not that acceleration anxious her personal colleagues — particularly after a wave of latest layoffs at corporations like Block — Tang paused. "How a lot we're in a position to do as an organization has accelerated," she stated, "however it nonetheless doesn't match our ambitions, not even one bit."
