OpenAI CFO pitches a four-part scorecard to measure AI's returns

The News

Sarah Friar, chief financial officer of OpenAI, has laid out a way for companies to judge whether their spending on artificial intelligence is actually paying off. In a piece published on 17 July 2026, she argues that the industry has plenty of enthusiasm and very little agreement on how to keep score.

Her answer is a four-part scorecard. Friar asks organisations to track the useful work an AI system delivers, the cost per successful task, the dependability of that system, and the return on compute. The framing is deliberately plain: instead of counting pilots, licences or dazzling demos, boards should count outcomes and the unit economics behind them.

The thrust of her case is that AI budgets are climbing fast while the vocabulary for justifying them has lagged behind. A shared set of measures, she suggests, would let a finance team compare one deployment against another and against the money it took to run.

Why It Matters

The timing is telling. After two years of headlong experimentation, the conversation around enterprise AI is shifting from what the technology can do to what it is worth. When the person holding a leading model maker's purse strings starts publishing frameworks about return on investment, it signals that the buyers have grown sceptical enough to demand proof.

There is precedent for this pivot. During the dot-com years, firms first chased raw traffic and "eyeballs", then, after the 2000 crash, retreated to hard measures like cost per acquisition and revenue per user. The cloud wave repeated the pattern, maturing only once chief financial officers forced a discipline of unit economics onto every workload. Friar's scorecard reads like an attempt to fast-forward AI to that same stage of accountability.

There is also self-interest at work, and it is worth naming. A vendor whose revenue depends on continued adoption has every reason to hand buyers a language for proving value, because a customer who can measure a win is a customer who renews. That does not make the metrics wrong, but it does mean Indian buyers should treat the scorecard as a starting template rather than gospel.

Indian Angle

For India's technology services giants, this is close to home. TCS, Infosys, Wipro and HCLTech have spent the past two years selling AI transformation to global clients, and those clients increasingly want evidence rather than slideware. A metric such as cost per successful task maps directly onto the outcome-linked pricing these firms are edging towards, replacing the old habit of billing for seats and hours.

The economics bite harder in rupee terms. Indian developers and startups pay for model access in dollars, so return on compute is not an abstraction but a live constraint on gross margins. Homegrown model builders such as Sarvam and Krutrim have staked their pitch on delivering acceptable quality at a fraction of the compute cost, and a scorecard that elevates efficiency plays straight to that strength.

There is a governance dimension too. As Indian boards and audit committees start asking where the AI money went, a crisp four-line scorecard is exactly the kind of instrument they can adopt. For firms weighing a domestic model against a foreign one, dependability and cost per successful task offer a neutral basis for the choice, rather than brand loyalty or hype.

FAQ

What problem is the scorecard trying to solve?

It targets the gap between soaring AI budgets and the absence of a common way to judge them. By fixing on outcomes and unit economics, it aims to replace vanity counts of pilots and licences with numbers a finance team can defend.

Does this apply only to large enterprises?

No. The logic scales down neatly. A small Indian startup weighing token costs against the value a feature ships can use the same four questions, and arguably needs the discipline more given tighter runways.

How is this different from ordinary ROI maths?

Standard ROI looks at cost against benefit in aggregate. This scorecard breaks the benefit into task-level success, reliability and compute efficiency, which suits systems that succeed probabilistically rather than every single time.

Should Indian buyers accept it wholesale?

Treat it as a template, not a verdict. Because a model vendor authored it, buyers should adapt the measures to their own definitions of a successful task before letting the framework drive procurement.

This story was reported by OpenAI. Read the full original coverage at OpenAI.