Forget vanity metrics — OpenAI's CFO just handed every business a four-point scorecard that finally answers whether your AI investment is actually working.
The AI ROI Scorecard Breakthrough
Sarah Friar, OpenAI's Chief Financial Officer, has published a practical framework for measuring AI return on investment — and it's a genuine shift from how most organisations currently think about AI spend. Instead of tracking logins, prompts fired, or time saved on paper, Friar's scorecard centres on useful work: did the AI actually complete something that mattered?
The four pillars are: useful work completed, cost per successful task, dependability (how often the AI finishes without human rescue), and return on compute. Each one is measurable, comparable quarter-on-quarter, and ruthlessly honest about whether an AI deployment is earning its keep.
Why "Useful Work" Changes Everything
The concept of useful work is the real breakthrough here. Most AI ROI conversations get stuck on inputs — how many hours of prompting, how many API calls — rather than outputs that move a business forward. Friar's framing flips that: a task that takes ten seconds but fails isn't cheap, it's expensive.
Cost per successful task is equally sharp. It forces teams to account for retries, human corrections, and failed runs — the hidden tax that makes many AI deployments look cheaper on a spreadsheet than they are in practice. If you want to go deeper on why orchestration and task completion matter so much, our course on Multi Agent Architecture That Actually Works breaks down exactly how to design systems that actually finish jobs reliably.
Dependability — the percentage of tasks completed without human intervention — is perhaps the most honest metric of all. It's the number that separates a genuinely autonomous AI workflow from an expensive autocomplete tool.
What This Means for Learners
If you're building AI skills right now, this scorecard is your new north star. Understanding how to design, evaluate, and improve AI systems against these four metrics is quickly becoming a core professional competency — not just for CFOs, but for anyone deploying AI at work.
Return on compute, in particular, is a concept worth getting comfortable with fast. As AI inference costs evolve and organisations run more workloads, knowing how to squeeze genuine output from compute spend will separate the AI-literate from the AI-dependent. Our Future of AI Inference course covers exactly how inference economics work and why they matter for real-world deployments.
The broader lesson: AI literacy in 2026 isn't just about prompting well — it's about measuring well. Teams that can audit their AI against frameworks like this one will make smarter decisions, cut waste, and build genuine competitive advantage.
