Cost per successful task.
Also known as: Cost per correct task, CPST
The full cost of getting one piece of AI work done properly. That means the model and licence costs, but also people’s time, the checking, the fixing, and all the attempts that failed. You divide it by the tasks that were actually finished and usable, not by the number of prompts or drafts. It is one way people are trying to answer the AI ROI question, because it measures work done rather than activity.
In the wild
“Shift AI metrics from tracking inputs (licenses, tokens, usage) to the cost per successful task”
Seen here: Harvard Business School AI Institute, September 2026
For AI users
For anything you do again and again, notice how many attempts and how much checking a good result really takes. A cheap first answer can turn into an expensive finished job by the time the model has taken several turns and you’ve made the final manual edits.
For business leaders
As with any task or project, define what success looks like with the person who owns the output before you start. Then put tools and processes in place to track what it costs to get there. Include the model and licence costs, but also the time spent reviewing, correcting, and redoing the work. Compare it with how the job was done before, and only compare tasks that are alike.