Why the AI invoice is the smallest part of what AI work costs, and how the labor spent fixing wrong answers became the largest line nobody reports.
Every AI provider bills the same way: per request, per token. That number is precise, it arrives every month, and finance teams have learned to manage it. It is also the wrong number to manage, because nobody in the business buys a request. They buy a finished piece of work: a contract reviewed, a bug fixed, a report drafted, a customer question answered.
A finished piece of work takes as many attempts as it takes. When the first answer is wrong, someone notices, works out why, and either asks again or fixes it by hand. That time is paid through payroll, not through the AI invoice, so it never appears next to the token spend it was meant to save.
Consider one hard task done on a model chosen to keep the invoice low. It takes four attempts at a total of $0.12 in tokens, and then an hour of a person's time to fix what still came back wrong. At $100 an hour, the real cost is $100.12.
The same task on the right model costs $0.50 in tokens and six minutes of someone checking the answer, a real cost of $10.50. The cheaper model saved 38 cents on the invoice and cost the business $90.
Rework hides for three reasons. It is spread across hundreds of people in small amounts, so no single instance looks expensive. It is paid from a different budget than AI spend, so nobody adds the two together. And most AI cost tools only see what passes through the provider, so they report the invoice in detail and the rework not at all.
The result is that companies optimize the number they can see. They push employees toward cheaper models, cap usage, and celebrate a falling AI bill while the labor cost of the work quietly rises.
The fix is to measure the cost to finish the work, not the cost of each request. That means counting every attempt a task takes, pricing the time people spend on rework at their actual labor rates, and adding the charges a price list leaves out, such as reasoning a model bills but never shows and conversation history that is charged again on every turn.
CompletionPrism™ does this before each request is sent. It prices the cost to finish that step on every model a company has approved, rework included, and runs it on the path with the lowest total cost that still holds quality. Sometimes that is a cheaper model. Sometimes it is a more expensive one, because one good answer costs less than four bad ones.
A falling AI invoice is not proof that AI is getting cheaper. The question a CFO should ask is what it cost to get the work finished and correct. Until rework is on the same page as the invoice, nobody can answer it.