OpenAI’s CFO Turns AI ROI Into a Four-Metric Scorecard

Sarah Friar, OpenAI’s chief financial officer, published a framework earlier this week for a question enterprise buyers keep asking without a clean answer: how do you know if AI spending is actually paying off? In a post titled “A scorecard for the AI age,” Friar argues that the metrics built for judging traditional software, mainly adoption and usage, do not hold up for agentic AI that takes autonomous action inside a business. Her proposed replacement is a single idea she calls “useful intelligence per dollar,” broken into four measurable parts, and it comes from the CFO of the company with the most riding on enterprises accepting that AI spending pays off.

The Four Questions Behind “Useful Intelligence Per Dollar”

Friar’s post reframes AI evaluation around four questions, shifting the conversation from whether people are logging into a tool to whether the tool is producing value that outpaces its cost.

Is the Work Actually Getting Done

The first question asks whether AI is completing tasks that matter to the business: resolving a customer issue, shipping a piece of code, reviewing a contract, rather than simply being opened or queried. In Friar’s framing, usage statistics are a starting point, not a verdict. A tool employees log into is not the same as a tool that finishes work correctly.

What a Successful Task Actually Costs

The second question is cost per successful task, and Friar is explicit that this is not the same number as price per token or per API call. The full cost of an AI task includes the compute to run it, plus the employee time spent on human review, retries, and rework whenever the first attempt falls short. Her post lays out the arithmetic directly: add up the full cost of completing the work, count only the tasks that met the required quality bar, then divide the cost by that number. A cheap model that needs three attempts to clear the bar can end up costing more per successful outcome than an expensive model that gets it right the first time. As a simple illustration: a $0.50 query that needs two retries at $0.50 each has a real cost per successful task of $1.50, not $0.50, a distinction per-token pricing hides by design.

Can People Depend on the Result

Dependability is the third measure: whether an AI system’s output is accurate, well sourced, and consistent, and whether it escalates to a human when the system is not confident, rather than delivering a wrong answer with false certainty. Friar ties this directly to cost. Results people can trust need less reviewing, correcting, and repeating, which is what allows a business to expand AI into more consequential workflows in the first place rather than keeping it confined to low-stakes drafts.

Does the Dollar Go Further as Usage Grows

The fourth question, return on compute, asks whether the cost per successful task falls, or the value generated per dollar rises, as an organization’s AI usage scales up. A pilot that looks efficient at low volume but does not get cheaper or more valuable per unit as it scales is, by this measure, not actually working yet; it is just small enough that the inefficiency has not shown up on the invoice.

Why Price Per Token Is the Wrong Number

The argument threading through all four measures comes down to one line from Friar’s post: “The basic economic question facing CFOs and other business leaders is whether the value of the work AI completes grows faster than the cost of producing it.” That framing is a direct challenge to how AI vendors typically compete: on benchmark scores and per-token pricing. A model that is cheaper per token but wrong more often can cost a business more once retries, human review, and downstream errors are counted. The lowest sticker price and the lowest real cost are not always attached to the same model, and a scorecard built around per-token pricing will pick the wrong winner every time the cheaper option needs a second try.

A CFO Metric for a Job That Has Changed

Friar’s post lands as CFOs are being pulled deeper into decisions that used to sit with engineering or product teams. According to McKinsey research cited in Fortune’s coverage of the scorecard, the share of CFOs with strategy responsibilities reporting to them has grown from about one-third five years ago to roughly two-thirds today, a shift McKinsey’s Andy West ties to CFOs moving from experimenting with AI to leading enterprise-wide transformation. A CFO-authored measurement framework fits that shift. Finance leaders are increasingly the ones being asked to approve AI budgets, which means they are also the ones who need a defensible way to say no to a vendor’s pitch.

The Company Proposing the Yardstick Has $500 Billion Riding on the Answer

None of that makes the four-question framework wrong. Cost per successful task, measured against quality and against retries, is a genuinely more honest number than adoption rates or per-token pricing, and it gives a buyer a way to compare a cheap, error-prone model against an expensive, reliable one on equal footing. But it is worth reading the source. OpenAI is not a neutral standards body publishing disinterested research into enterprise software economics. It closed a $122 billion funding round at an $852 billion post-money valuation in March, with Amazon committing $50 billion and Nvidia and SoftBank each committing $30 billion, and it is carrying out a four-year, $500 billion commitment under the Stargate project to build 10 gigawatts of AI infrastructure in the United States, with an initial $100 billion phase already underway. Every one of those numbers depends on enterprises concluding that AI spending clears the bar Friar just described.

That does not make the scorecard a sales document dressed up as a framework. Cost per successful task is a metric OpenAI’s own enterprise customers can and do apply to any vendor, not just OpenAI, and a company confident its models produce fewer retries and less rework has a real incentive to popularize a metric built around exactly that comparison. The framework is genuinely useful to a buyer trying to evaluate AI spending. It is also, unavoidably, the specific case for AI spending that its author’s company needs the market to accept.

The Takeaway

Friar’s scorecard gives finance teams a concrete alternative to counting logins: track the full cost of a successful task rather than the price of a token, and treat dependability as an economic variable instead of a technical footnote. That is a real improvement over how most organizations currently measure AI value, and it is specific enough that a finance team could start applying it this quarter. Whether OpenAI’s own models, or any vendor’s, actually clear that bar is a separate question, and it is one the scorecard cannot answer for you. Applying “useful intelligence per dollar” honestly means being willing to run OpenAI’s own numbers through it and finding they fall short too.