Uber gave a new artificial intelligence coding tool to about 5,000 engineers last December. The engineers loved it. By April the company had spent its entire 2026 AI budget, according to the Financial Times.
That is not a story about technology failing. It is a story about technology working, and a bill nobody knew how to estimate.
The culprit has a name: agents. A chatbot answers a question and waits for the next one. An agent takes an assignment and goes off to do it — reading files, writing code, checking its own work, starting over when it fails. A single instruction can touch off hundreds of steps.
AI is sold by the token, roughly three-quarters of a word, counted going in and coming out. Each step an agent takes drags along everything that came before it. Researchers at the Stanford Digital Economy Lab found that agent tasks can burn a thousand times the tokens of ordinary chat, with most of the cost in what the machine reads rather than what it writes. The same task run twice can differ thirtyfold, and the models are poor judges of their own appetite.
So companies started metering. Uber now caps employees at $1,500 a month per tool. Walmart limits tokens on its in-house coding assistant. Microsoft moved thousands of engineers off a tool this summer after individual bills reached $2,000 a month.
A year ago some firms kept leaderboards celebrating heavy AI use. The word for it was tokenmaxxing. This year’s word is tokenminimizing.
Not everyone is pulling back. Databricks still gives its engineers an unlimited budget. And most businesses never felt any of it. The median American company spends $11.38 per employee per month on AI, according to the Ramp AI Index. The top 1 percent spend about $7,500.
Rick Good spent ten years in the newspaper business, including with The News & Observer (Raleigh, NC) and The Greenville News (Greenville, SC). He co-wrote, with Melinda Coleman, Rutledge to Riley: Governors of South Carolina, 1776–Present (Published by The Greenville News), and is the author of Mind Games Trivia #1 (Amazon). He has recently spent 15 years creating original trivia content through Sounds Good Mobile Entertainment and continues that with his new company, SGME Software. He edits The Smart Reader’s money, real estate, and business coverage.
