OpenAI finance team builds toward a zero-day close with AI
An unnamed author who says they joined OpenAI two years ago to build its finance function from the ground up describes redesigning that function around AI. The team set two goals: a zero-day close, meaning a real-time, reconciled, traceable view of the company's financial position, and continuous, automated forecasting that updates as the business changes. Both are described as ambitions still in progress, not finished systems, but the author says the work has already pushed the team away from static spreadsheets, manual record searches and slide decks toward live tools built on the company's full data and context.
The author lists five lessons. First, broad AI access paired with structured experimentation: the team ran a finance hackathon with sales engineers, out of which came IR-GPT, a custom GPT grounded in approved investor-relations materials that answers diligence questions, plus other custom GPTs under development for procurement and tax. Second, AI changes the unit of work, so finance should redesign the whole path from source data to decision rather than automate individual steps; for the close, that means connecting approved spending plans, general-ledger actuals, purchase orders, accruals and transaction details into one continuously reconciled view, with AI drafting an initial variance explanation and flagging exceptions while finance validates the numbers and keeps final sign-off, and forecasting building on that same reconciled base by combining statistical models, sales conversations, account-level evidence and finance judgment in one live view.
Third, finance staff are now building their own tools. The author cites OpenAI research finding that 40% of finance professionals' specialized AI use involves work outside traditional finance, and 22% involves engineering-related tasks. As an example, a teammate on the advertising side who had never coded used Codex to build a tool that turns the monthly advertising forecast into weekly and daily plans, accounting for weekdays and holidays and keeping every number tied to the approved model. Fourth, control has to scale with speed: IR-GPT cut diligence answers that used to take hours, sometimes overnight, down to a first draft in seconds, but the investor relations team still reads every draft and checks consistency; the author's rule is that every AI output should trace to a reliable source, every forecast should carry a clear explanation, and every change to an approved baseline should require finance authorization, with CFOs working alongside IT and governance to set data access, action limits, approval thresholds and budget controls (the author calls the earlier phase of unrestrained AI spending 'tokenmaxxing' and says it has passed).
Fifth, measurement should track operating performance rather than seats or tokens purchased. The author proposes asking four questions of every AI-assisted workflow: did it complete work that mattered, what did it cost including employee time and rework, was the result good enough to use, and did it help the team move faster or decide better. Suggested close metrics include cycle time, the share of transactions reconciled automatically, the number of exceptions needing review and the time to explain a variance; suggested forecasting metrics include accuracy, refresh frequency, time to produce a new scenario and the quality of decisions the forecast supports. The author adds that the cheapest model is not always the most economical one if a costlier model reaches a reliable answer with fewer attempts and less review. The piece closes by framing an AI-native finance function as faster cycles, stronger controls, better decisions and more time for judgment, and argues finance's central view of strategy, capital, data, risk and performance makes it a template other parts of the company can follow.
Key facts
- OpenAI's finance team is pursuing two still-unfinished goals: a zero-day close giving a real-time, reconciled, traceable view of the company's finances, and continuously updated automated forecasting.
- A finance hackathon that brought in sales engineers produced IR-GPT, a custom GPT grounded in approved investor-relations materials that turns diligence answers that used to take hours, sometimes overnight, into a first draft in seconds.
- Cited OpenAI research finds 40% of finance professionals' specialized AI use is work outside traditional finance, and 22% is engineering-related tasks.
- A teammate on the advertising side who had never coded used Codex to build a tool that converts the monthly advertising forecast into weekday-aware weekly and daily plans.
- The author proposes scoring each AI-assisted workflow on four questions: did it complete work that mattered, what did it cost, was the result good enough to use, and did it help the team move faster or decide better.
Why it matters
OpenAI is publishing, under its own name, a blueprint for how it wants finance departments to run: less manual reconciliation, more real-time visibility, decisions redesigned end to end rather than automated step by step. The underlying argument goes beyond finance: AI changes the 'unit of work' available to a team, letting non-specialists like the advertising teammate who built a forecasting tool with Codex take on work that used to require a dedicated engineer.
Who it affects
CFOs and finance leaders weighing whether and how to rebuild reporting, closing and forecasting workflows around AI. More broadly, anyone assessing how AI reshapes back-office functions, since the author argues finance's central view of strategy, capital, data, risk and performance makes it a template the rest of a company can follow.
How to use it
This is a first-person essay, not a product launch, and names no pricing. It does name the tools used internally: ChatGPT Work and Codex for building custom dashboards, plus custom GPTs such as IR-GPT for investor relations and others in development for procurement and tax. The practical steps described are pairing broad AI access with structured hackathon-style experimentation, redesigning full decision workflows rather than single steps, letting non-technical staff build their own tools, setting explicit governance (data access, action limits, approval thresholds, budget controls), and scoring workflows with the four-question test rather than counting seats or tokens.
How solid is it
A single first-person opinion piece on OpenAI's own blog, written by an author whose name, title and gender are never stated in the text. It describes internal practices rather than externally audited results. The two headline figures, 40% and 22%, are attributed only to 'recent OpenAI research' with no citation, methodology or link. No dollar figures, headcount or cost-and-savings numbers back the zero-day close or forecasting claims, and both are explicitly described as ambitions still being built toward, not completed systems.
Risks and caveats
OpenAI is describing its own staff's use of its own products, ChatGPT Work, Codex and custom GPTs, on its own blog, which makes the piece implicit marketing as well as advice. No adoption figures, error rates or usage volumes are given for IR-GPT or the advertising forecasting tool, so both examples are qualitative anecdotes rather than measured outcomes, and none of the claims can be checked against a source beyond the text itself.
“The cheapest model isn't always the most economical.”
— the post's author