Two goals: zero-day close and continuous forecasting—both a work in progress

She set two goals, and they're sequential.

Why the first goal first? Because forecasts are built on book figures. With a constantly updated ledger, forecasts can be recalculated anytime. If the books only close at month-end, then even frequent forecasting uses stale data—the more often you calculate, the faster you compound the error.

Why spend two years chasing these? The payoff isn't speed itself. Faster closes and frequent forecasts are means to an end: seeing a problem while you can still do something about it. Discovering a line item collapsed after the quarter ends leaves no options; spotting it mid-quarter means you can reallocate budget, pull back on ads, or change pricing.

Both goals are still under construction. The article is about the direction and how work processes have already changed, not about finished results: it doesn't give a single number for how long a close takes now or how often forecasts refresh.

A one-day finance hackathon produced IR-GPT

The first step was giving everyone access and letting people experiment with their own tasks. But access alone isn't enough—most people try it twice and then drop it, because nobody has time to figure out "how could this task be done differently?"

So they ran a hackathon: one day, the entire finance team set aside regular work. Each person picked their most tedious repetitive task and built a tool for it on the spot. They brought in sales engineers to help—the tech-savvy folks who usually build solutions for clients. Anyone stuck had someone to ask. By the end of the day, someone could go the whole distance: pick a task they do monthly, build something usable, have a colleague try it, and iterate immediately.

One output was IR-GPT. When investors consider funding or follow-on rounds, they send a barrage of questions: How was this revenue calculated last year? Why did this cost rise? Why doesn't this number match last quarter? Previously, each round meant someone digging through old materials and drafting answers by hand. IR-GPT feeds the "approved for external communication" documents to a custom GPT, which produces a first draft. Procurement and tax teams each built their own later.

The lesson here is a two-sided effort: the people doing the work know which steps are painful and what good looks like, but they often build small utilities; leaders have to identify what matters most to the company and concentrate resources there. This follows: when the points below align with what leadership has identified as most important, it makes sense to scale it up.

Work backwards from the decision, redesign the whole process

The hackathon was just the starting point. The real shift is something else: AI changes the size of the chunk you can rework at once.

In the past, efficiency efforts targeted individual steps: make a spreadsheet faster, find data quicker, write documents more efficiently. Speeding up each step by 20% leaves the overall chain unchanged, because the waiting and handoffs between steps remain. Now you can operate at a different scale: redesign the entire path from raw data to a decision.

The method is directly transferable: pick a real decision that needs to be made—say, "should we increase next quarter's ad budget?"—then trace backwards, listing what data you need, whose approval is required, and how many handoffs occur. Then go step by step and ask: Can AI compute this step? Can it help chase things down? Can it do it entirely on its own?

Five data sources auto-reconcile; AI drafts, finance approves

Applying the backwards approach to the close produces a workspace. It pulls everything previously scattered across systems into a single page, and it's constantly auto-reconciling: approved budgets, actual spend from the finance system, signed-but-unpaid POs, estimated accruals, and individual transactions.

Once connected, spotting "we're over budget this month" allows you to click through to the specific transaction or PO. AI drafts an explanation of the variance, flags anomalies, and prepares a review; a human verifies the numbers, judges materiality, and signs off.

Reconciliation workspace: left column shows three data sources (budget, finance system, POs), each marked with a green "matched" indicator; the middle column walks through how the monthly expense was calculated; the right column shows Codex's review note, marked "Draft." Notice the period is "April · Demo year" and the PO number is DEMO-1042, with all amounts as $x.x—this is demo data. Image / OpenAI

The bottom-right of this screenshot has a line that captures the division of labor perfectly:

Codex handles the mechanics. Finance owns the judgment.

Caption from screenshot: Codex automates the mechanics. Finance owns the judgment.

Another line worth remembering: the close doesn't disappear—what disappears is the frantic post-period scramble to reconstruct last month.

Forecast suggestions show four evidence trails; human acknowledgment required

The reconciliation workspace tells you what happened last month. The forecasting workspace handles something harder: what next quarter might bring, and what justifies changing that number.

The second screenshot is a 38-second screen recording of a tool called StratFin DS. It opens on a table with three numbers for the same quarter side by side: what the model calculated on its own, the previous approved version, and the current working version; the next two columns show you the exact differences. This layout makes it immediately obvious who bumped a number up, and by how much.

Scrolling down, you can adjust individual accounts: select a client, drag a point on the curve to where you think it should be, and the surrounding numbers update instantly.

The most revealing part is the "Recommendations" page. AI suggests bumping a client's number up, and alongside it lays out "here's why," each point annotated as connected to its source:

Here's the key point: it breaks the reasoning into things you can verify one by one. You can listen to the Gong call, or judge for yourself whether the Slack pricing change holds up. You can see for yourself whether the suggestion still stands if one piece is missing.

A 38-second recording: from the three-number quarterly table, to selecting a client and dragging a point, to the evidence-backed AI recommendation. The client name APEX Inc. is anonymized. Image / OpenAI

One level up in the same interface is budget allocation: seeing where marketing spend is still producing returns, where it's flattening, and where the next dollar might be better spent.

Marketing spend efficiency curve: the x-axis is daily spend, and the y-axis is return. As the curve flattens to the right, additional spend yields less. R²=0.94 indicates a tightly fitted curve. The table below lists channel reallocations. Image / OpenAI

Finance staff without a coding background are building tools with Codex

With tools like this, the profile of who builds them has changed too.

In late July, OpenAI's Economic Research team published a report with two relevant figures: for finance professionals, 40% of occupation-specific AI usage involves tasks outside the traditional scope of finance, with 22% related to engineering.

To understand the 40%, context helps. Comparing across eight occupations, finance is actually among the least cross-disciplinary:

So read the 40% this way: even in one of the least cross-disciplinary roles, four out of ten tasks push beyond the traditional job description.

Her team now uses ChatGPT Work and Codex daily to build their own tools. ChatGPT Work is OpenAI's office agent released in July—it can work across software, run for hours, and deliver finished spreadsheets, slides, and web pages. Codex is OpenAI's coding agent. Deliverables have changed accordingly: instead of a static Excel file plus a PPT that dies on delivery and needs a rebuild when data changes, the output is a live page you can query further; when the underlying data shifts, it updates itself.

The most concrete example: a team member supporting ad operations, who previously had zero coding experience, used Codex to build a tool that splits the monthly ad forecast into weekly and daily plans. It accounts for business days and holidays, compares forecast versions, and locks every number to the approved model. The marketing lead opens it to see what changed, where to invest, and what the extra spend buys.

Finance professionals haven't become less specialized. Their expertise now extends further: instead of submitting a request and waiting for IT, they can build what they need themselves. And the act of building forces you to rethink why the work has always been done this way.

AI can draft, but it can't change an approved number

When everyone can build tools and AI can generate output in seconds, the question shifts: who owns the error?

Back to IR-GPT. Historically, when an investor came with questions, an analyst would dig through past materials, draft answers, double-check consistency with what previous investors were told, and run a review across several people. It took hours, sometimes overnight. Now, the "approved for external communication" documents are fed into it, and it produces a decent first draft in seconds.

But it's a first draft. The article doesn't state how long a human needs to read and refine it, so "hours to seconds" shouldn't be treated as the full speedup. The investor relations team's work hasn't decreased: they still read the draft, add judgment and context, and check that different investors get consistent answers. AI saves the time spent hunting for materials and typing. The responsibility stays with people.

So they set hard boundaries. Together with IT and risk management, they put four things in writing:

Additionally, three non-negotiables:

The third one is the hardest and deserves its own explanation. "The approved version" refers to the forecast the company has signed off on, by which departments allocate budgets and the external narrative is set. If that number changes, everything downstream moves: do budgets need re-splitting? Does the board presentation change? So AI can't change it, no single person can change it without review, and finance must approve. The "needs human review" tag and the "Accept suggestion" button in the forecasting tool are exactly this rule rendered in UI.

The article also touches on an industry buzzword: the brief "tokenmaxxing" craze is over.

Why it's over: all the necessary controls are now standard. Usage caps per person, budgets, role-based access, routing simple tasks to cheaper models, and approval gates for big-ticket items. AI spend can be managed like utility bills or travel expenses, while still leaving room for teams to experiment.

With boundaries set, teams feel safer letting AI move faster, because they know which step to check and who to call when something goes wrong. And when reconciliation, aggregation, and material prep speed up, people free up time for what can't be delegated: challenging assumptions, advising the business, making the call.

Four Questions, Two Scorecards—Not About Seat Counts

Once controls are in place, the remaining question is simple: is it worth it? The answer: number of seats or token counts say nothing. What matters is whether the work got done well and what it truly cost.

Ask four questions for every workflow:

For a finance team, "useful work" can be specified: a forecast update, an explained variance, a closed audit query, an answered board question. Drilling down further, close and forecasting each have their own quantifiable scorecards:

The final note is a corrective to "cost-saving" logic: the cheapest model isn't necessarily the cheapest overall. A stronger model might give you a usable answer in one pass, saving more in human iteration and rework than the model price difference.

To Get Teams to Use It, First Show Them What It Looks Like

The piece closes on a favorite line of hers: you can't be what you can't see. To help finance teams believe what AI can do, they first need to see it—and that has to start with the CFO herself.

The direction has two parts: see the books anytime, and let forecasts follow. Forecasting sits on top of the close—if the books aren't reconciled, frequent forecasts are just using stale data.

The starting point was a hackathon: one day, each finance team member turned their most tedious task into a tool. The method that followed is the bigger shift: AI changes the size of the process chunk you can redesign at once—the whole path from data to decision.

Steal the method: pick a decision, list its data, tool, approval, and handoff steps backwards, then ask whether AI can analyze, coordinate, or finish each step.

When the forecasting tool suggests an adjustment, it lays out its reasoning alongside: four evidence sources tagged "Connected," plus a confidence score. The change only takes effect after a user clicks "Accept."

Approved numbers are off-limits to AI. It can draft and suggest, but changing the baseline requires finance sign-off. Colleagues with no coding background now use Codex to break down forecasts.

Seat counts and token burn tell you nothing.

• Does it do meaningful work?
• What's the real cost? Including staff time, review, rework.
• Is the quality usable as-is?
• Does it make us faster or help us decide better?

For the close, track cycle time, auto-reconciliation rate, anomaly count, and time to explain a variance. For forecasting, track accuracy, refresh speed, scenario turnaround, and decision quality. The full metrics and hard rules are in the main article.