Codex Cuts Trade Validation From 30 Minutes to 4

Chatham Financial, a financial services company, built OpenAI's Codex into its trade validation process. A trade used to take about 30 minutes to check. Now it takes less than 4 minutes. This is the company's "first pass" at these results. They are still checking the agent's work against scores from their expert reviewers.

The task here is clearly defined. You compare the trade terms against the client agreement. The agent genuinely helps the human reviewers here, and I am not just saying that to be nice. But on fuzzier tasks, I would not expect this kind of productivity gain. An agent will always produce something that looks like a correct answer. You just cannot always tell right away that it is wrong. This example comes from OpenAI's own blog, so the results deserve a degree of skepticism.