OpenAI unveils an AI ROI scorecard: four metrics instead of counting tokens

Sarah Friar, OpenAI's chief financial officer, published an "AI-era scorecard" on the company's blog on 17 July 2026. It is a system of four metrics for evaluating the return on enterprise AI: useful work, cost per successful task, reliability and return on compute.

What metrics does the scorecard include?

OpenAI's scorecard has four metrics that together answer what Friar calls the central economic question for CFOs: is the value of the work AI performs growing faster than the cost of producing it? She calls the resulting formula "useful intelligence per dollar."

• Useful work — AI should complete meaningful tasks: resolved customer requests, deployed code, reviewed contracts, rather than merely showing "active usage"
• Cost per successful task — the full price of a result: AI spend, retries and human time spent on review, not just the price per token
• Reliability — how consistently the system delivers results that can be trusted without constant oversight
• Return on compute — whether each dollar invested in AI yields more and more value as usage grows
• Publication — 17 July 2026, OpenAI blog, a signed column by CFO Sarah Friar

Why shouldn't AI be measured in tokens?

Sarah Friar argues that familiar software metrics — seats purchased, active users, subscription renewals — do not capture the real value of AI.

"For years, software was measured by adoption: seats, active users, renewals.

AI is different: it needs to be measured by work completed," — Sarah Friar, OpenAI CFO.

Friar adds that as capabilities grow, models take on ever longer and more complex tasks: they hold context, reason across multiple steps and work with several tools at once. The stakes are high: according to figures cited by CFO Dive, global AI spending will reach $2.59 trillion in 2026, up 47% from a year earlier.

"Tokens create value when they turn into work that people can use,"

Friar writes on the OpenAI blog.

How many companies are earning back their AI investments?

Only 12% of executives said AI had both cut costs and increased revenue at their companies, according to a January PwC survey cited by CFO Dive. Another 33% or so of respondents saw an effect in at least one area, either costs or revenue, while 56% of organizations have not yet recorded a significant financial result.

Friar's scorecard is a direct response to this gap: OpenAI is urging CFOs to count not subscriptions and request volume but the share of tasks carried through to a useful result, and their full cost. This approach makes it possible to compare AI projects with one another and with the "human" alternative.

What this means

The largest AI vendor is itself setting the framework for evaluating AI payback — a signal that the market is moving from experimentation to strict financial accounting. For companies, Friar's scorecard works as a ready-made checklist: if the cost per successful task is not falling and "useful work" is not rising, it is time to revisit the AI budget.

Frequently asked questions

What is OpenAI's AI-era scorecard?

It is a system for evaluating AI payback built on four metrics — useful work, cost per successful task, reliability and return on compute — which OpenAI CFO Sarah Friar presented on the company's blog on 17 July 2026.

How do you calculate cost per successful task?

According to Friar's method, it is the full cost of getting a result: spending on AI usage, retries and employee time spent on review, not just the price of tokens.

Are companies' AI investments paying off right now?

According to the January PwC survey, only 12% of CEOs see both revenue growth and cost reduction from AI, while 56% of organizations have not yet seen a significant financial effect.