Across thousands of companies, an employee at the 99th percentile, meaning only one in a hundred spends more, spends $699.74 a week on AI. The typical employee spends $2.49. The surprise is which of the two is growing: over six weeks the typical employee's weekly spending more than doubled from $1.10, while the heavy users' spending stayed roughly flat.
The heavy users are not where the growth is
A coding agent that works all day costs far more than an assistant that answers a few questions, so heavy users are the first target of almost every cap. Picture a company of 1,000 people. Its ten heaviest users spend at least $7,000 a week between them. The other 990, each spending the typical $2.49, come to about $2,500. Today those ten cost nearly three times as much as everyone else together.
Now keep the ten flat and let the typical spending keep its recent pace for twelve more weeks, which is two more doublings. Each of the 990 would spend about $10 a week, so together they would reach about $9,900 and pass the ten. The middle climbs because assistants have moved from answering a question to doing multistep work, and harder tasks use more tokens. Every ordinary employee drifts toward what the heavy users already do.
So a company that spends its energy finding power users is managing last quarter's bill. The next one is made of a thousand small habits.
Cloud went through this, slowly
The cloud version took most of a decade. Finance lost sight of what engineers were spending in the early 2010s, and the FinOps Foundation was only set up in February 2019. The AI version is moving faster. Microsoft capped AI spending in every division in July, and in November it began showing each employee and vendor their own usage and spend. Within four months, a company of that size added visibility to limits.
Three lessons from cloud carry over. Count usage by person and team before you set any limit, because a limit set blind will hit the wrong people. Plan the budget for the worst week, not the average one, because an agent that loops can cost more in a day than a month of questions. And judge spending by cost per finished piece of work, not per token. Some companies have started giving every employee an AI budget, which only works if each person can see how fast they are using it.
A big bill cannot tell value from waste
A $700 week, about $36,000 a year, can be the best money a company spends or the worst. The number alone does not say which. In one coding test, two models finished 89.1% of tasks. One cost $0.05 a task and the other $1.62, so a heavy user on the pricier model pays about 32 times more for the same result. Choosing the model for each task is how that gap closes, and it is a habit that employees learn only when they can see their own spending.
The first question about a large AI bill is therefore what it produced, and only a manager who knows the work can answer it.