OpenAI CFO Sarah Friar has published a four-part scorecard for measuring AI's return on investment, shifting the yardstick from how clever a model sounds to how much useful work it actually delivers per dollar of compute [S1]. The framework, posted on OpenAI's site on 17 July 2026, names four metrics every AI buyer may soon be asked to track. Whether it becomes the industry default or just one company's opinion depends on details the announcement leaves out.

A finance chief's framing, not a researcher's

The four metrics Friar names are useful work, cost per successful task, dependability, and return on compute [S1]. Each one reframes AI value as a finance department would measure it, not a research lab.

"Useful work" asks whether the model completed something a human would have had to do, not whether it produced a fluent paragraph. "Cost per successful task" divides spend by tasks that actually worked, not total attempts. "Dependability" measures whether the output is reliable enough to ship without a human checking every response. "Return on compute" ties the whole thing back to the hardware bill, the line item that has come to dominate AI budgets.

This is a CFO talking. Friar runs the numbers at the company whose models are among the most expensive to run in the industry. Her framing puts the cost of inference, the price of actually running a model, at the centre of the value question.

Why the gap between demos and reality

For the last two years, companies buying AI have been flying blind on ROI. They can cite benchmark scores and demo wins, but most cannot answer a simple question: for every dollar spent on AI compute, how much labour did we actually replace or accelerate?

Friar's scorecard is an attempt to give that question a standard shape. The four metrics are not new as concepts. Packaging them as a named framework from OpenAI's CFO gives them weight. A mid-sized company trying to justify its AI spend to a board can now point to a framework from the biggest AI vendor's own finance chief and say: this is how we should measure it.

The shift from benchmarks to business metrics matters because the gap between demo performance and real-world results has been the single biggest complaint from AI buyers. A model that scores well on tests can still fail at a specific task a company needs done, and until now there has been no shared vocabulary for capturing that gap.

What it means

The scorecard's real contribution is conceptual. It tells AI buyers to stop asking "how smart is the model?" and start asking "how much value does it deliver per dollar of compute?" That is a question a CFO can put on a spreadsheet, and it is a question vendors will have to answer with numbers, not narratives.

For a technology that has been sold on vibes and benchmark curves, a four-metric framework from the vendor's own finance chief is a quiet but real shift. It signals that the industry is moving from the phase where AI is bought on promise to the phase where it is measured on output.

What it means for business

A two-person consulting firm paying $2,000 a month for AI tools across several platforms now has a framework to audit that spend. They can ask: how many tasks did the AI complete that we would have done ourselves? What did each successful task cost? How often did we have to redo the work?

A suburban real estate agency using AI to draft property listings can track dependability: what percentage of generated listings needed no human edits before going live? If that number is 60 per cent, the cost per successful task is higher than the raw API bill suggests, because 40 per cent of outputs required rework.

The return on compute metric is the one most operators have not been tracking. It asks whether the GPU hours spent running a model produced output worth more than the compute cost. For small businesses using API-based tools, this is buried in the monthly invoice. For larger firms running their own infrastructure, it is the line that determines whether an AI project survives the next budget review.

What we don't know yet

The announcement does not include specific numerical benchmarks, thresholds, or dollar figures [S1]. It names the four metrics but does not publish the methodology for calculating each one. There is no detail on how "useful work" is defined or measured, what counts as a "successful task," or how dependability is quantified.

OpenAI has not said whether it has deployed the scorecard internally or published results from its own use. The word "introduces" in the announcement confirms the framework exists as a published concept, not that it has been tested in production [S1].

No other organisation has publicly adopted the scorecard. Whether it becomes an industry standard or remains OpenAI's proposal depends on whether other AI vendors, independent auditors, and corporate finance teams pick it up and refine it.

The next signal to watch: whether OpenAI publishes a detailed methodology guide, or whether a third party releases an independent implementation of the four metrics that companies can actually run against their own AI spend.

Subscribe for the follow-up when OpenAI publishes the methodology, or when someone tries to build an independent version.

Sources

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.