OpenAI just published the cheat sheet every CFO will copy, and every AI vendor will pretend they invented first.
The Summary
- OpenAI's CFO Sarah Friar released a four-metric framework for measuring AI ROI: useful work completed, cost per successful task, system dependability, and return on compute
- This is OpenAI turning enterprise sales into a science, giving buyers a language to justify AI spend beyond vibes and demos
- The framework standardizes how companies evaluate AI systems, potentially reshaping procurement and vendor accountability across the industry
The Signal
OpenAI just did something more strategic than shipping another model upgrade. Sarah Friar published a measurement framework that gives enterprises a common vocabulary for AI procurement. Four metrics: useful work completed, cost per successful task, dependability (success rate over time), and return on compute (business value per dollar of infrastructure).
This matters because AI buying decisions still happen in a fog. CIOs greenlight six-figure contracts based on impressive demos and vendor promises, then struggle to measure whether the system actually delivers. Friar's scorecard creates accountability where there was mostly hope.
"The shift from measuring inputs to measuring outcomes changes who wins AI deals."
The useful work metric is the sharpest edge here. Not tasks attempted. Not queries processed. Useful work completed means the AI actually finished something a human would have done, correctly, without requiring a do-over. That's a much higher bar than most vendors want to clear. It forces the conversation from "how fast is your model" to "how often does it actually work."
Cost per successful task flips the pricing model. Right now, most AI tools charge per seat or per API call, regardless of whether the output was garbage. This metric makes vendors eat their own failure rate. If your system halts 40% of the time, that 40% shouldn't be billable.
Key accountability shifts this creates:
- Vendors can't hide behind raw speed or parameter counts anymore
- Finance teams get a framework that maps to existing ROI models
- Procurement can compare AI tools the same way they compare SaaS, with clear unit economics
Dependability and return on compute are the enterprise-grade concerns. Dependability tracks whether the system degrades over time, catches drift, maintains accuracy as data changes. Return on compute matters because AI infrastructure is expensive, and CFOs want to know if throwing more GPUs at the problem actually moves the needle on business outcomes.
This scorecard also signals where OpenAI sees the market heading. They're not positioning as the scrappy research lab anymore. They're positioning as the enterprise AI platform, the one that speaks CFO language and understands procurement cycles. Friar isn't just measuring OpenAI's tools. She's teaching the market how to measure all AI tools, and setting the terms on which those comparisons will happen.
The Implication
If you're buying AI, download this framework and use it in your next vendor call. Ask for each metric explicitly. Watch which vendors squirm. If you're building AI tools, start tracking these four numbers now, because your buyers will ask for them within six months. And if you're a CFO, this is your Rosetta Stone for translating AI hype into line items that make sense on a P&L.
The companies that can report clean numbers on this scorecard will win enterprise deals. The ones that can't will get relegated to point solutions and pilot purgatory.