OpenAI Finance Runs at 20% of Peer Headcount With AI-Native Workflows

Stacie Faggioli describes OpenAI’s finance organization as trying to build “the finance team of the future” while still managing the financial performance of the business. The operating model she lays out is not simply that finance employees should use AI tools more often. Her claim is stronger: AI should change the workflow design, the staffing model, and the location of technical talent inside finance.

The first principle is to be “AI-native by design.” Faggioli says that when the team deploys a tool or agent, it does not treat it as a bolt-on to an existing process. The finance team tries to reimagine how the workflow should be done, propagate best practices through the team and cross-functional stakeholders, and increasingly rethink hiring and org design around agents.

The second principle is measurable headcount leverage. Faggioli says a recent PwC assessment found OpenAI’s finance team was about 20% the size of comparable technology peers. She presents that as evidence that, with the right tooling, a finance organization can produce more output without adding headcount in the same proportion.

The third principle is to deploy early and iterate quickly. Faggioli says the finance team’s bias is not to wait for a more stable or final version of AI, because the technology is changing too quickly. Instead, the team deploys tools and continues iterating so that finance employees can grow with the technology in real time.

That principle shapes the structure of the organization. Faggioli describes three finance pillars. Strategic finance allocates capital, raises capital, and plans growth. Finance operations runs the finance engine: collecting revenue, paying bills, filing taxes, and closing the books. Enterprise FinTech owns finance systems and data platforms, including ERP systems, planning tools, integrations, datamarts, software engineering, agent development, MCPs, and ChatGPT connectors.

The important organizational choice is where engineers sit. Faggioli says OpenAI embeds engineers directly in finance, inside the Enterprise FinTech pillar, rather than treating them as a separate IT function receiving requirements from finance at a distance. Her argument is that engineers sitting beside finance subject-matter experts speed deployment and iteration, allowing tools to evolve alongside the work itself.

The investor relations use case began with fundraising diligence. Stacie Faggioli says she has been part of two historic equity raises at OpenAI: a $40 billion private-market raise last year and, a few weeks before the presentation, the close of $122 billion in financing. Behind those transactions, Faggioli says, the team was overwhelmed by diligence requests from investors.

The team built an investor relations agent with ChatGPT. Faggioli calls it a personal favorite because it was simple enough for a non-technical person to implement and, in her view, produced very high ROI.

Faggioli says the agent was trained on OpenAI’s internal data and on the tone of a “public company grade” investor relations professional. The configuration shown on screen described the agent as an “investor relations helper for OpenAI” whose job was to answer questions from current and prospective investors “in a clear, accurate, and professional way.” The visible instructions emphasized factual accuracy, sound judgment, and caution over completeness when those goals conflict. They also stated that the agent should answer investor relations questions about OpenAI using only information that OpenAI had publicly shared.

The operational change, according to Faggioli, was that diligence responses no longer had to be manually assembled each time. The agent could produce data-grounded, factual, consistent, strategically framed answers in minutes. Faggioli frames the scarce resource as the CFO’s ability to provide investor-grade answers.

Faggioli says this contributed to OpenAI running both raises entirely in-house, which she describes as saving hundreds of millions of dollars in advisory fees.

The same agent also changed how finance supported recruiting and internal equity-storytelling. Faggioli says finance’s influence scaled because the agent helped the company tell its equity story in contexts beyond fundraising. When OpenAI tries to recruit a senior executive in research, engineering, or product, and the candidate receives a large equity compensation package, that person wants to understand what the equity is worth. Finance has shared the investor relations agent with recruiting teams and executives so they can tell the same story the CFO would tell.

Stacie Faggioli treats ChatGPT for Excel as particularly important for finance because it works inside the medium finance teams already use. Her emphasis is not just on answer generation, but on how the tool structures analysis and writes formulas directly into the workbook.

When given a task, she says, ChatGPT for Excel first thinks through how to structure the analysis, then writes formulas into Excel. The result, in her view, is traceable and auditable: a user can inspect the formulas, toggle assumptions, and run scenarios rather than receiving an opaque written answer.

To demonstrate the point, Faggioli uses a private-equity-style exercise built around a company called OpenDesk. She says that as a former private equity associate, she remembers the all-nighters spent building leveraged buyout models. In the demonstration, she imagines being asked to analyze the take-private of a publicly traded company called OpenDesk. The on-screen source material is an equity research report for “OpenDesk, Inc. (ODK)” with financials and valuation information.

The prompt shown in the ChatGPT pane tells the tool to act as a senior private equity investor and build a comprehensive one-sheet LBO model. The requirements are demanding: clear assumptions, executive-readable outputs, full LBO mechanics, a five-year hold period, sources and uses, entry valuation, capital structure, debt schedule, cash sweep, exit valuation, sponsor returns, interest-rate sensitivities, revenue-growth sensitivity, and headline outputs such as MOIC, IRR, cash-on-cash, enterprise value, deleveraging path, and implied equity check. The prompt also tells the tool to use conservative SaaS assumptions where data is missing and label them as assumed.

The important finance point is not only that the system can produce a polished spreadsheet. It is that the system produces a workbook a finance user can review. ChatGPT for Excel plans the model, extracts information from the report, populates the spreadsheet, formats it, adds assumptions and outputs, and checks tie-outs. The visible workbook includes an executive snapshot, core inputs and assumptions, sources and uses, capital structure, and operating forecast summary.

Faggioli highlights that the tool does not merely summarize the PDF. It builds a model with projections, capital-structure assumptions, cost-of-capital assumptions across equity and debt tranches, sources and uses, sensitivities, and outputs a human investment associate would expect. It also produces an investment recommendation. In the demonstration, the final output says: “Recommendation = Do not proceed.” Faggioli says the tool concluded the investment did not meet return thresholds.

The time matters to her argument. She notes that the demonstration was a speed run and that, based on the timestamps shown, ChatGPT took about 10 minutes to produce the model. Faggioli is not claiming that OpenAI finance spends its days building LBOs; she says the same underlying capabilities — structuring analyses, running scenarios, and creating outputs — are part of the team’s daily workflows.

Stacie Faggioli says Codex has put coding and software engineering capabilities into the hands of non-technical users like herself, creating use cases across data analysis and automation.

One use case is a marketing spend dashboard her team actually uses, with some numbers blurred. Faggioli says OpenAI receives extensive marketing data from agencies, broken down by geography, channel, and keyword. The constraint is not data availability; it is the time required to analyze the data and extract useful insights.

The finance team uploaded the data into Codex and asked it to build an ROI dashboard. The resulting dashboard lets the team toggle among channels and understand spend and ROI by channel, including when ROI begins to diminish because a channel is becoming saturated. Faggioli says they further asked Codex for the top five recommendations for reallocating spend based on the data.

The dashboard shown includes a scatter plot of daily spend against blended daily CAC, a saturation model, and a table of proposed shifts from less efficient donor market-channels to more efficient receiver market-channels. The visible text describes the dashboard as testing shifts from higher-CPA or higher-CAC donor channels into receiver channels where incremental return remains positive after diminishing returns and penalty assumptions.

The operational change is cadence. Faggioli says the team now rebalances marketing spend weekly, moving dollars from less efficient channels to more efficient ones. She calls this a game changer, but the substantive mechanism is clear: Codex converts a large, granular dataset into a dashboard with reallocation recommendations that finance can revisit regularly.

For sales insight, OpenAI used Codex to avoid adding another CRM reporting burden. Faggioli says she and OpenAI’s CRO wanted leading indicators on whether sales reps were actually discussing a newly launched product with customers or defaulting to legacy products.

The traditional approach, as she describes it, would be to add a field to the CRM, require reps to enter whether they discussed the product, download that data, and have a human analyze it. OpenAI instead used Codex to pull from places where the relevant signals already existed: Gong transcripts and emails between customers and reps. The resulting dashboard shows, by segment, geography, and account, which reps are talking to customers about the new product.

The dashboard shown includes filters for time frame, owner segment, top accounts, account owner manager, and account owner. It shows executive-summary metrics such as total calls, calls where the product “Starling” was discussed, accounts with calls, and accounts with calls where Starling was discussed. When filtered to Enterprise (AMER), the dashboard shows updated call counts, Starling discussion counts, segment-level call coverage, call-stage mix, and account-week review rows.

For Faggioli, the value is that the company can see before the quarter ends whether field motion is developing as expected. If not, it can redeploy resources or bring in a team that specializes in the product to help existing reps sell it.

Because Codex can write code, Stacie Faggioli says it can build front-end interfaces in ways similar to a front-end engineer. That means it can help with presentation workflows that previously required humans to gather data, reconcile it, apply accounting logic, and turn it into polished charts.

The use case is a compute margin slide for a monthly CFO review. Faggioli says the slide is real but the numbers shown are made up. The visible frame is explicitly labeled “Synthetic sample data for illustration only.” It shows a “Straffin Slide Creator” prompt asking for a new slide for paid compute margin and a generated slide with a line chart titled “API & Web Compute Margin | 68%, +3pts M/M and +5pts vs. Plan,” plus a product-line margin table.

Before Codex, Faggioli says, creating this type of slide was highly manual. The team had to reconcile raw infrastructure telemetry data across products and GPU types. Accounting then had to overlay rules for cost allocation. Finally, a human had to turn the result into executive-ready charts.

With Codex and reusable skills, she says, that workflow has been compressed into a few hours of Codex work. Human review remains built into the process. Finance staff still inspect the data, create QA checks or evals, and stress-test the numbers to make sure they pass the team’s “sniff test” and look accurate.

The tool compresses preparation time, but human finance judgment remains in the review loop. Faggioli says the process saves probably days of time every month.

Stacie Faggioli distinguishes between individual productivity tools and agents embedded at the organizational workflow level. The latter, she says, automate tedious and routine work by orchestrating across multiple systems.

She gives four examples: procurement questions, customer credit checks, contract reviews, and vendor risk.

The procurement agent handles procurement and travel questions that humans previously answered. Faggioli gives a simple example: an employee traveling to London asking how much can be spent on a hotel each night. She says the agent can deflect about 60% of these questions automatically and is improving over time.

The credit check agent is used for customers spending above a certain threshold. Previously, Faggioli says, credit risk analysts researched each customer, pulled relevant information, and created a composite credit risk score. Now, a credit risk agent can do that in minutes. The score is embedded directly into the CRM so sales reps can quickly see whether to continue with a customer.

The contract review agent ingests agreements in bulk, structures the data, and flags non-standard terms. Faggioli frames this as important for both accounting and deal desk work. Accounting previously had to read agreements to identify terms that might affect revenue recognition under GAAP ASC 606. The deal desk also had to manually review agreements for terms that were too non-standard. The agent now flags these issues in bulk.

The leverage point is scale. As agreement volume grows, Faggioli says OpenAI does not have to scale the accounting team linearly in order to keep closing the books on time. Customers also receive faster deal desk responses.

The vendor risk agent automates research and compilation that previously went into vendor risk reports. The resulting risk score is embedded in the procurement software approval flow, so approvers can either proceed or escalate based on the score.

Stacie Faggioli closes on what she calls an AI mindset. Before taking on a difficult task, she says the team is now conditioned to ask what ChatGPT can do or how ChatGPT can make the work easier. The point is not a single tool choice, but a default angle of attack: assume the workflow may be redesigned with AI before doing it manually.

She also emphasizes democratized access. A photo from a team hackathon shows employees seated together and raising their arms to form the letters “A” and “I.” Faggioli says OpenAI put ChatGPT and Codex into the hands of everyone, and that none of the examples she presented were her ideas.

Her argument is that finance innovation comes from the people closest to the underlying work: the people in the data, in the systems, and in the daily workflows. In the model she describes, the finance team of the future is not only smaller relative to output. It is organized so that finance experts, embedded engineers, and broadly available AI tools can continuously redesign how finance work gets done.