OpenAI finance: a five-day month-end analysis now takes five hours

OpenAI's finance team cut the monthly analysis of its computing costs from about five days to roughly five hours, its director of product finance, Kyle Kober, told [CFO Dive in an interview published on 1 September 2026](https://www.cfodive.com/news/inside-openai-experiment-ai-coding-finance/829250/). Finance staff did it themselves with Codex, OpenAI's AI coding agent: a tool that writes and runs software from instructions typed in plain English. The report gives no cost and names no reviewer.

The job

Kober described the work as one of the most cumbersome parts of OpenAI's month-end reporting: reconciling product-usage data with the accounting records, then turning the result into analysis and reports on what the company spends on computing capacity. He and colleagues used Codex to automate "much of the process", CFO Dive reports. Which steps are still done by hand is not stated.

No engineer is credited. "There's now the opportunity to just do it yourself," Kober said. His background is finance: Nextdoor's product and corporate finance team and, before that, an investment bank. He said finance picked up Codex "a couple of months" after OpenAI's engineers did in November or December 2025, which puts the start in early 2026; CFO Dive gives no exact date. The team now also uses ChatGPT Work, a product OpenAI launched in July 2026 that works across apps and files.

What the number is, and is not

• "About five days to roughly five hours" is Kober's figure. CFO Dive does not say whether five days means one person's working week or elapsed time across several people, how many closes the five hours has held for, or how it was timed.
• OpenAI sells Codex and ChatGPT Work, so this is a seller describing its own use of its own product, inside the company whose engineers build the tool.
• No cost is given: no usage bill, no build hours, no time spent checking the output.
• The analysis is one part of the close. OpenAI has not said its books close faster. Its chief financial officer, Sarah Friar, set a "zero-day close" (transactions recorded and reconciled as they happen, so little is left to do at period end) as a goal in a blog post on 10 August 2026 and wrote: "We are still building toward both ambitions," [CFO Dive reported](https://www.cfodive.com/news/openai-aiming-zero-day-close-cfo-says-ipo/827521/).

The people side and the catch

On people, the sources are nearly silent. CFO Dive says Friar is building out the finance team as the company grows; no headcount is given, and nothing says what the staff who spent five days on the analysis do with the time.

The catch is control. A script written by an AI agent now produces numbers that feed month-end reporting, and Kober's account does not say who reviews the code, who checks its output against the ledger, or how a change is logged for an auditor. CFO Dive quotes Boston Consulting Group's caution that finance leaders need "clear guardrails" for AI coding, auditability among them. Friar's post, as [PYMNTS summarised it](https://www.pymnts.com/news/artificial-intelligence/2026/openai-cfo-sarah-friar-built-an-ai-native-department-mid-market-cfos-can-start-smaller/), says finance should decide which data the AI can reach, which actions need approval, and that every output should tie back to a reliable source. Neither report shows how OpenAI does that for this workflow.

What to do if you run finance operations

• Pick one recurring close task with a named owner and a clear output, such as a reconciliation between an operational system and the ledger. Time it for two closes. That is the baseline Kober's figure lacks.
• Before anyone automates it, have your controller write down who reviews the script, who ties its output to the ledger each month, and where changes are logged. Show that note to your external auditor.
• Count the review time and the usage bill against the hours saved, and expand only if the net is positive after two closes.