Oracle Cuts Tasks From Days to Minutes With ChatGPT Work and Codex

Oracle has told OpenAI how it turns its employees' expert knowledge into fast, repeatable workflows. The company uses ChatGPT Work and Codex in recruiting, engineering and operations, and, in OpenAI's words, what used to take days now fits into minutes. This is a customer case study on OpenAI's own blog, so "days to minutes" should be read as a claim by the parties involved, not as an independent measurement. The available announcement contains no detailed metrics by department.

Three functions, one principle

The choice of departments is telling. Recruiting, development and operations run by different rules. The first involves a lot of text and judgment calls, the second involves code, tests and review, and the third involves procedures and approvals. They share one thing: each function has specialists whose knowledge lives in their heads and in scattered documents.

From the description, Oracle takes that knowledge, formalizes it as a reproducible scenario and hands it to an AI tool. The result no longer depends on who is on shift today or how much time they are willing to spend. The word "repeatable" is the key one here. Any company can get a one-off win from a chatbot, but a stable process that can scale across a team is rarer.

Codex, OpenAI's coding agent, covers the engineering side. ChatGPT Work, judging by its name, is aimed at work scenarios outside development. The announcement does not say how tasks are split between the two.

Why the Oracle case reads differently

Oracle is not an ordinary customer. According to press reports from 2025, the company became one of OpenAI's main compute suppliers under the Stargate project, with the agreements estimated at hundreds of billions of dollars. At the same time, Oracle sells its own OCI cloud and enterprise applications with built-in AI.

So the story has two sides. For OpenAI, it is a showcase: a major enterprise player demonstrating ChatGPT inside its own processes. For the market, it is a signal that Oracle's internal use cases are not limited to its own products. Still, the parties' interests are intertwined, and a case like this is best treated as marketing material with useful reference points, not as an audit report.

What is known from the announcement

• Company: Oracle.
• Tools: ChatGPT Work and Codex.
• Functions: recruiting, engineering, operations.
• Claimed effect: tasks that took days are cut to minutes.
• Source of the data: OpenAI's blog, with no independent verification.
• Not disclosed: the share of automated tasks, the number of employees, cost, error rate and the amount of manual review.

How to apply this yourself

If you want to replicate the approach, start with a task where a specialist makes the decision but 70–80% of the time goes to gathering and formatting materials. In recruiting, that is the initial résumé screening and preparation of structured notes. In engineering, it is routine edits, tests and migrations. In operations, it is preparing standard reports and approvals.

The sequence from there:

1. Measure the baseline completion time in hours before rollout. Without that number, there is nothing to check the claim "it took days, now it takes minutes" against. 2. Ask the expert to write down their algorithm as a step-by-step instruction, with examples of good and bad results. 3. Keep a human on the final decision and track the share of edits they make. 4. For Codex and similar agents, set up an isolated environment and mandatory tests so that a fast result does not turn into fast debt. 5. After a month, compare time and quality against the baseline, and only then scale the scenario to neighboring teams.

The main question about the Oracle case remains open: how many of those minutes are down to the model, and how many to the process built around it in advance. The answer will determine whether the result can be transferred to another company.