The AI-Native Finance Playbook: 5 Lessons From OpenAI's CFO

OpenAI’s finance team is aiming for a zero-day close. Not eventually. Now. And Sarah Friar, the company’s CFO, just published the playbook for how they’re getting there.

Most companies talk about “AI transformation” in the abstract. OpenAI shows the receipts: the workflows they rebuilt, the custom GPTs they built internally, and the hard lessons from a team that has access to their own frontier models and still found the technology was the easy part.

Here’s what’s worth stealing.

The hook: a zero-day close and continuous forecasting

The headline goal is a zero-day financial close with continuous, automated forecasting. That sounds impossible if you’ve ever sat through a month-end close. But Friar frames the real prize differently. A faster close means nothing if leaders can’t act on the information sooner. The value isn’t speed for its own sake. It’s collapsing the gap between what happened and what you can still influence.

This is the reframe that matters. Most finance teams chase faster closes because the close is painful. Friar is saying: chase it because it gives the business more time to act while the outcome can still change.

Lesson 1: Give everyone access, but give them a reason to use it

OpenAI ran an internal finance hackathon. Sales engineers paired with finance staff to build custom GPTs. IR-GPT for investor relations. Tools for procurement. Tools for tax. The hackathon wasn’t a gimmick. It was the mechanism that turned “access to AI” into “people actually using AI on real work.”

Hand every employee a ChatGPT license and walk away, and most of them will use it to draft emails. That’s fine, but it’s not transformation. The hackathon model works because it puts your closest-to-the-work people in a room with people who understand the tools, and gives them a deadline to build something specific.

Lesson 2: Redesign the whole workflow, not just the pieces

This is where most AI deployments stall. You automate one step, invoice processing say, and feel good about the efficiency gain. But the upstream and downstream steps are still manual, so the bottleneck just moves.

Friar’s point: you have to redesign the full decision workflow. Where does the data come from? Who reviews it? What decision does it trigger? Map the entire path from raw input to decision, then figure out where AI changes each step. Automating pieces of a broken workflow just makes a broken workflow run faster.

Lesson 3: Finance professionals become builders

The IR-GPT example is telling. Finance staff didn’t request a tool from IT and wait six months. They built it, with help from sales engineers during the hackathon. The ownership stayed with the people who understood the problem.

This is a real shift. Finance teams have traditionally been consumers of software that someone else built. The AI-native model turns them into builders. You don’t need them to become software engineers. But you do need them comfortable prototyping, iterating, and owning their own tools.

Lesson 4: Pair speed with accountability

Moving fast with financial data is risky. Friar is explicit that speed has to come with clear accountability and controls. This isn’t a throwaway caveat. It’s a design constraint. If your AI-assisted workflow produces a number that feeds a real decision, someone needs to own that number and be able to explain how it was derived.

The trap is treating AI-generated analysis as automatically trustworthy because it came from a powerful model. It didn’t. It came from a model that can hallucinate, and the controls need to assume that.

Lesson 5: Measure value per unit of intelligence

This is the most counterintuitive lesson, and the one most useful for anyone managing AI costs. The cheapest model isn’t always the cheapest overall. If a cheap model takes five attempts to get a reliable answer, or worse, produces confident wrong answers that propagate downstream, it’s more expensive than a pricier model that gets it right the first time.

Friar calls this “value per unit of intelligence.” Total cost isn’t just token price. It’s token price multiplied by the number of attempts you need to get a trustworthy result, plus the cost of catching errors you didn’t catch.

What to do now

If you’re in finance or operations, three things are worth pulling from this.

First, run a hackathon. Pick your smartest domain experts, pair them with someone who knows the tools, give them a day, and see what they build. The ROI of one good internal tool usually dwarfs the cost of the day.

Second, map one full decision workflow end to end. Pick one, forecasting or expense analysis or investor reporting, and trace it from raw data to the decision it informs. Find where AI changes each step. Don’t automate one piece and declare victory.

Third, start tracking cost per reliable answer, not cost per token. If your team is burning through cheap-model calls because half the outputs need redoing, your “cheap” model is the expensive one.

The OpenAI finance team has advantages most companies don’t: frontier models, internal expertise, a culture built around building. But the lessons transfer. The hard part was never the model. It was redesigning how work flows through an organization. That part is available to anyone willing to do it.

Based on OpenAI CFO Sarah Friar’s account of building an AI-native finance function.