OpenAI's finance team cut a five-day task to five hours because the person doing it mapped it first

Last Tuesday, CFO Dive published a short story about Kyle Kober, director of product finance at OpenAI. He and his colleagues used Codex, OpenAI’s coding agent, to automate the analysis of the company’s computing costs. By his account, the work went from about five days to roughly five hours.

The tool gets the headline. What interests me is who did the automating. Kober wasn’t an engineer handed a ticket. He was the finance person who knew the job, and he described himself as central to rebuilding the team’s manual processes. The hard part, he said, was reconciling product usage data with the accounting records before anyone could start the analysis.

That detail is the whole story for operators. The slow step was invisible from the outside. You’d only know about it if you had done the job.

By the numbers

5 days to 5 hours

Kober's estimate for OpenAI's monthly compute cost analysis, before and after

Source: cfodive.com

What happened at OpenAI

Kober had led product and corporate finance at Nextdoor before OpenAI. He is not a software developer. According to the CFO Dive piece, OpenAI’s engineers had their “Codex moment” in late 2025, and finance followed a couple of months later. The team now uses Codex alongside ChatGPT Work, a product OpenAI launched in July for longer tasks that span apps and files.

This sits inside a larger push. OpenAI’s CFO, Sarah Friar, wrote last month about building an AI-native finance function, including a move toward a zero-day close and forecasts that update continuously.

Kober’s own summary of the shift was that there’s now “the opportunity to just do it yourself.” He means finance people building their own tools without waiting in the engineering queue.

Two cautions are worth keeping in mind. The five-days-to-five-hours number is Kober’s estimate in an interview, not an audited figure. And the article doesn’t describe how the new workflow is checked. Boston Consulting Group, quoted in the same piece, warns that finance teams building with AI coding tools need guardrails such as auditability. OpenAI also has every reason to show its own tools working. None of that changes the useful part of the story.

The slow step is never the one on the org chart

If you asked most CFOs to describe the monthly close, they’d describe the outputs. The reports, the review meeting, the sign-off. They would not describe the two days someone spends lining up a usage export against the general ledger, chasing the mismatches and rebuilding the same pivot table.

That’s where time goes in most companies: in the gaps between steps. Pulling data out of one system to put it into another. Waiting for someone to confirm a number. Fixing the same formatting problem every month. None of it shows up in a process document, because nobody writes down the workarounds.

This is why so many automation projects disappoint. Someone automates the step they can see, like generating the report, and leaves the real bottleneck, like the reconciliation, exactly where it was. The report comes out faster. The close takes as long as it always did.

Before and after

The process on paper

Pull the data
Run the analysis
Write the report
Review and sign off

The process as it's done

Export from two systems
Match usage to the ledger
Chase every mismatch
Rebuild the same table
Then the analysis

The steps nobody writes down are usually where the days go.

How to map work before you hand it to an agent

You don’t need process mining software or a consultant to do this for one workflow. You need the person who does the work, a stopwatch and an honest hour. This is the method I’d use.

Sit with the person who does it, while they do it. Don’t ask them to describe it in a meeting. Watch them do it the next time it comes up, and write down every step, including the ones that feel too small to mention. Opening a second spreadsheet is a step. Messaging someone to ask which version is current is a step.

Time each step. Rough minutes are fine. You’re looking for the two or three steps that eat most of the time, and they are often not the ones people would guess.

Mark what kind of step each one is. Some steps are moving data from one place to another. Some are checking or matching. Some are judgment calls, where the person decides something based on experience. Agents are good at the first two when the rules are clear. The third stays with a person, at least at first.

Find the waits. Note every point where the work stops because it is waiting for someone else. Those waits often matter more than any step. An agent can’t speed up a person who hasn’t answered an email, but it can make sure the question gets asked on day one instead of day four.

Let the person who does the work own the fix. This is the part OpenAI got right. The person who knows the workarounds is the one who will notice when the agent’s output is wrong. Give them the tool and the time to rebuild the process, and make them the reviewer when it goes live.

FlowMapping one workflow in an afternoon

01Watch the person do it
02Time every step
03Label each step
04Find the waits
05Owner picks what to automate

The person who does the job maps it, owns the fix and checks the output.

The map becomes the instructions

There’s a second payoff to doing this properly. The page you end up with, the steps, the timings, the rules for what counts as a match and what gets escalated, is most of what an agent needs to do the job. Agents fail on vague instructions the same way a new hire does. A map written from watching the real work is the opposite of vague.

Say you run a 50-person distributor, and your controller spends three days each month calculating sales commissions. On paper it’s one step: apply the commission plan to closed deals. Watch her do it and you’ll likely find she exports deals from the CRM, cross-checks them against invoices actually paid, removes returns, applies two exceptions the sales VP agreed to last year, and emails three reps to confirm split deals. Each of those is a line in the agent’s instructions. The exceptions and the split deals stay with her. The export, the matching and the returns are the part to hand off.

Without the map, you’d have automated “apply the plan” and wondered why the controller still needed three days.

What to do this week

Pick one recurring task that takes more than a day each month, like a close, a client report or a commission calculation.
Book an hour to sit with the person who does it the next time it runs. Write down every step and time it.
Circle the two steps that take the longest and note whether each is moving data, matching data or making a judgment.
Ask the person who does the work what they would automate first if they could. Write down their answer before you suggest anything.
Give them a budget of hours over the next month to try it, with a simple check: a second person compares the automated output to the manual one for the first two cycles.