AI ROI: Measure Results, Not Usage
21 August 2026
I read an interesting perspective from OpenAI on how companies should think about ROI in the AI age, and there are parts of it that I quite agree with.
The article argues that the real question is not:
How much AI are people using?
It is:
How much useful work is getting done, at what total cost, and with what level of reliability?
OpenAI describes this as “Useful Intelligence per Dollar.”
This framing is interesting because it changes what we should measure.
Token prices, model costs, number of prompts and even adoption tell us something about AI usage. They do not necessarily tell us whether AI is creating value.
So perhaps the ROI equation could be:
AI ROI = Valuable work successfully completed / Total cost of getting that work right
And that total cost matters.
It includes the AI cost, but also employee time, review, retries, corrections and escalation.
Eventually, the organisations that get AI ROI right may not be the ones using the most AI, or even the cheapest AI.
They may simply be the ones getting the most dependable, valuable work done for every dollar they spend.
Reference
