OpenAI's chief financial officer, Sarah Friar, has rolled out a new internal metric called 'useful intelligence per dollar' — a scorecard designed to measure the return on investment from the company's AI products. The move signals a shift away from tracking raw adoption numbers toward evaluating the actual economic value those tools generate for customers.
The new metric
Friar introduced the scorecard as a way to answer a question that has dogged the AI industry since the ChatGPT boom: is all that spending actually paying off? Instead of counting how many users sign up or how many queries they run, the 'useful intelligence per dollar' framework tries to capture the tangible business outcomes tied to each dollar spent on OpenAI's models. The company hasn't disclosed the specific formula or what data feeds into the calculation, but the concept is straightforward — measure the useful output, not just the activity.
Why the shift matters
For months, investors and corporate buyers have pressed AI vendors for clearer proof that the technology delivers real returns. Adoption metrics like monthly active users or API calls can look impressive but don't show whether a company is actually saving money, boosting revenue, or solving problems. By focusing on economic value, OpenAI is trying to speak the language of CFOs and procurement officers who need to justify AI budgets. The scorecard also puts pressure on competitors like Google and Anthropic to come up with their own ROI frameworks.
What's inside the scorecard
Details are sparse. Friar didn't release a public version of the scorecard or give examples of how it works in practice. The announcement came during an internal meeting, according to people familiar with the matter. What's clear is that the metric is meant to guide product development and pricing decisions inside OpenAI. If a model delivers high 'useful intelligence per dollar,' it could get more resources. If it doesn't, the company might rethink its approach.
OpenAI hasn't said whether it will publish the scorecard externally or make it available to customers. For now, the framework is an internal tool. But if it proves useful, it could become a benchmark for the entire industry. The question is whether 'useful intelligence' can be defined in a way that's both honest and consistent across different use cases — from coding assistants to customer service bots. Friar's team is expected to refine the metric over the coming quarters.
