Most companies deploying AI right now are flying blind. They know they’re spending money on it. They’re pretty sure it’s doing something useful. But ask them to prove it — to quantify what that investment is actually returning — and the conversation gets uncomfortable fast. OpenAI CFO Sarah Friar thinks that’s a solvable problem, and she’s published a practical framework to prove it. The AI scorecard she’s proposing is four metrics wide, surprisingly concrete, and aimed squarely at the boardroom argument that AI budgets still need to win.
Why This Is Happening Now
The timing isn’t accidental. We’re roughly three years into the mass enterprise adoption of generative AI, and the honeymoon phase is ending. Early pilots ran on enthusiasm and novelty. Now CFOs across industries are asking the same question Friar herself sits in the hot seat to answer: what are we actually getting for this?
The pressure is real. Enterprise software spend on AI has ballooned — Gartner estimates that AI-related IT spending will exceed $600 billion globally by 2025. But the measurement frameworks haven’t kept pace. Companies are still reaching for legacy productivity metrics — time saved, headcount ratios — that don’t map cleanly onto what AI actually does.
This is especially acute now that agentic AI is entering the picture. When an AI model answers a question, it’s relatively easy to assess quality. When an AI agent completes a multi-step business process autonomously, the evaluation problem gets significantly harder. Traditional ROI tools weren’t built for that world.
Friar’s scorecard is an attempt to give enterprise buyers — and frankly, OpenAI’s own sales conversations — a shared vocabulary for this new environment. It’s also, let’s be honest, a smart piece of market positioning. If your CFO is the one defining how AI gets measured, you have some influence over what “good” looks like.
The Four Metrics, Unpacked
The framework Friar proposes centers on four core measures. Here’s what each one actually means in practice:
- Useful Work: This is the foundational metric — the volume of tasks the AI completes that genuinely move something forward. Not tokens generated, not prompts processed, but outcomes that matter. Think reports filed, tickets resolved, contracts reviewed. The emphasis on “useful” is doing a lot of heavy lifting here; it explicitly rejects vanity metrics that inflate activity without driving results.
- Cost Per Successful Task: This is where the real financial accountability lives. It takes total AI spend — compute, licensing, integration, maintenance — and divides it by the number of tasks completed to a satisfactory standard. It forces a conversation about what “successful” means, which is harder than it sounds but absolutely necessary.
- Dependability: Essentially an uptime and reliability score for AI outputs. Does the system produce consistent quality? Does it fail gracefully? For enterprise workflows where AI is embedded into critical processes, a model that works brilliantly 80% of the time and hallucinates the other 20% isn’t usable, regardless of average output quality. This metric captures that.
- Return on Compute: The most technically specific of the four. This measures the value generated per unit of compute consumed — essentially asking whether you’re getting efficient use of the underlying infrastructure. As model costs evolve and companies make decisions between running models via API versus on-premise or private cloud deployments, this becomes a genuine strategic variable.
Taken together, these four metrics create a 2×2 of sorts: you want high useful work and high return on compute on one axis, low cost per task and high dependability on the other. A system that scores well across all four is genuinely delivering. A system that scores brilliantly on one while tanking on another reveals a specific kind of problem you can actually fix.
What’s Missing From the Scorecard
To be fair to the framework, it’s deliberately practical rather than exhaustive. But a few things are conspicuously absent. There’s no metric for risk — the cost of errors, compliance exposure, or reputational damage from AI mistakes. For industries like healthcare, finance, or legal, that’s not a footnote; it’s potentially the most important number on the page.
There’s also no explicit measure of user adoption or satisfaction. A technically efficient AI deployment that employees actively route around is still a failed investment, and cost-per-task won’t surface that problem if usage is artificially low.
How This Compares to What Else Is Out There
Friar’s isn’t the first attempt at an AI ROI framework. McKinsey, Deloitte, and a dozen boutique AI consultancies have all published their own versions. The difference here is the source: this is coming from the CFO of the company whose products most enterprises are actually buying. That gives it a different kind of gravity in a procurement conversation.
It’s also notably more operational than most. Competing frameworks tend to be high-level — focus on strategic alignment, change management maturity, governance readiness. Friar’s scorecard skips most of that and goes straight to the numbers a finance team can pull into a spreadsheet. That’s either a strength or a weakness depending on where your AI deployment actually is.
Our earlier analysis of how to measure and manage AI investment in the agentic era pointed to exactly this gap — the industry has been long on aspiration and short on operational metrics. Friar’s framework is a real step toward closing it.
The Agentic AI Dimension
This framework clearly anticipates a world where agentic AI — systems that execute multi-step tasks autonomously — is the dominant deployment model. Cost per successful task and dependability are almost tailor-made for evaluating agents rather than simple chat interfaces.
That’s not surprising given where OpenAI is pushing its product roadmap. ChatGPT Work and the broader Operator and Agent frameworks OpenAI has been building out are all bets on agentic workflows becoming standard enterprise infrastructure. A scorecard that helps customers justify those deployments is good for OpenAI’s revenue, full stop.
It also implicitly raises the stakes for competitors. If enterprises start adopting these metrics as standard, then Anthropic’s Claude, Google’s Gemini, and Microsoft’s Copilot suite all get evaluated on the same axes. That might seem fair — but OpenAI helped define the axes. That’s not nothing.
What This Means for Different Audiences
For Enterprise Buyers
If you’re a CTO or CIO trying to justify AI spend to your board, this framework gives you a starting structure. The four metrics are defensible, relatively easy to communicate, and grounded in actual business outcomes rather than technical proxies. I’d treat it as a starting template, not a finished answer — you’ll almost certainly need to add sector-specific metrics and weight them differently depending on your use cases.
For AI Vendors Beyond OpenAI
This is a competitive pressure play whether Friar intended it that way or not. If your sales team is walking into enterprise deals where the buyer has already internalized OpenAI’s scorecard, you need to be able to answer those four questions credibly — and ideally add a metric or two where you have a genuine edge. Anthropic’s focus on safety and predictable outputs, for instance, maps well onto the dependability metric. That’s a real talking point.
For Smaller Businesses
The framework scales down reasonably well, but cost per successful task requires some upfront definition work that smaller teams may not have bandwidth for. The most immediately useful metric for a 50-person company is probably useful work — just start tracking what AI actually completes versus what gets kicked back or reworked. Even rough data there is more useful than none.
FAQ
What is OpenAI’s AI scorecard?
It’s a four-metric framework proposed by OpenAI CFO Sarah Friar for measuring the return on AI investment. The metrics are useful work, cost per successful task, dependability, and return on compute. It’s designed to give enterprises a concrete way to evaluate whether their AI deployments are delivering real business value.
Who is this scorecard designed for?
Primarily enterprise buyers and finance leaders who need to justify AI spend in business terms. It’s most directly applicable to organizations running OpenAI’s products at scale, but the metrics are general enough to apply across vendors and deployment types.
How does this relate to agentic AI deployments?
The framework is clearly built with agentic workflows in mind — systems that complete multi-step tasks autonomously rather than just answering individual queries. Metrics like cost per successful task and dependability are particularly relevant for evaluating agents, where the stakes of failure are higher and the value of consistency is more significant. For more context, see how ChatGPT Work fits into this agentic shift.
Does this scorecard account for AI risk or safety?
Not directly. The four metrics focus on performance and efficiency rather than risk exposure or compliance. Organizations in regulated industries will need to supplement this framework with additional measures covering error costs, audit trails, and governance requirements. OpenAI has separate work underway on AI safety evaluation systems that could eventually complement this kind of operational scorecard.
The most interesting test of this framework will come in 12 to 18 months, once enterprises have had time to actually implement it and report back on whether the numbers told them anything useful. If cost per successful task starts appearing in quarterly earnings calls alongside traditional SaaS metrics, that’s a sign the industry is finally growing up around how it thinks about AI value. And if it doesn’t land that way — if companies find the metrics too hard to operationalize or too easy to game — then the next version of this conversation will be more interesting still.
