OpenAI's CFO Proposes a 'Scorecard for the AI Age' with Four ROI Metrics

OpenAI CFO Sarah Friar has published an article discussing how businesses should measure the returns on their AI investments. In response to the question executives keep asking — how much value are we actually extracting from our AI spend — she argues companies shouldn’t rely on traditional software metrics like seat counts or license counts, but should instead evaluate “Useful Intelligence per Dollar.”

Details

• The underlying economic question: Friar points out that the fundamental issue is whether the value of the work AI completes is growing faster than the cost of producing it
• Metric 1, useful work completed: Emphasizes that tokens only create value once they translate into real outcomes — customer issues resolved, code shipped, contracts reviewed, or time saved
• Metric 2, cost per task: Calculated as total cost of completion divided by the number of tasks that met quality standards. This needs to account not just for model usage fees and compute, but also for the time spent on human review and redoing work. A more expensive model can actually deliver a better ROI if it reduces the need for redos and reviews. As an example, the new model “GPT-5.6 Sol” achieves comparable results on coding benchmarks at an estimated API cost 36.2% lower than competitors
• Metric 3, dependability: Tracks outputs across three categories — “usable as-is,” “needs correction,” and “requires human escalation.” Higher dependability reduces review overhead and allows AI to be embedded more deeply into operations. This also depends on having clear governance around data access and permissions; ChatGPT Work includes built-in enterprise-grade security and compliance features
• Metric 4, economies of scale: Continuously monitors whether the volume of work completed is growing faster than cost as scale increases, and whether quality is being maintained or improved. Friar argues that as the underlying platform keeps improving, better models lead to better products, which in turn drives greater adoption and further research investment — a virtuous cycle

How to try it

• This scorecard complements OpenAI’s earlier “five principles” for managing AI investment in the agentic era, published July 14, 2026, and can serve as a framework for CFOs and IT administrators evaluating the cost-effectiveness of their company’s AI use
• Companies already using ChatGPT Enterprise and the redesigned Admin Console to gain visibility into usage and cost can apply these four metrics to reassess their own ROI evaluation