What cost to completion means, how it differs from cost per token, and why it is the only AI unit economics measure a CFO can defend to the board.
AI spend has become a board-level line item faster than any category before it. Gartner put worldwide generative AI spend at $644 billion in 2025, up 76% in a year, and most enterprises report running over their AI budgets. Yet few finance teams can say what a unit of AI work costs. They can say what was spent. They cannot say what it bought.
Cost to completion is the total cost of getting one piece of work finished and correct. It has five parts:
Together they are the whole cost of finishing the work. Each one counts money the others do not.
Cost per token is a supplier price, not a unit cost. It is like measuring a construction project by the price of a bag of cement. It looks precise, it is easy to compare across vendors, and it leaves out the parts that drive the total: how much material was wasted and how many hours of labor went into fixing mistakes.
Once cost to completion exists, AI spend becomes manageable like any other operating cost. Finance can see cost per completed task by team and task type, compare it month over month, and attribute savings to specific decisions. Rework that used to hide in payroll lands on the same page as the invoice.
The figures also need to be defensible. Every number should say whether it was observed, inferred or estimated, and carry the label of its weakest input, so a reviewer can follow any saving to its source and stop wherever they stop believing it.
CompletionPrism™ measures cost to completion on your own traffic. A two-week shadow baseline, with nothing rerouted, reports cost per completed task, rework, completion rate and a projected annual saving with its assumptions shown. After that, it lowers the number by choosing the right model for every step before the money is spent.