Uber just became the first major case study proving that AI agent pricing is fundamentally broken for enterprise finance. The company didn't fail. Its engineers loved Claude Code so much they burned the entire 2026 budget in 120 days.

According to CTO Praveen Neppalli Naga in confirmation to The Information, Uber exhausted its full-year AI budget by April. Not because the tool didn't work. Because it worked exactly as designed, and no finance team had a model for what "working as designed" would cost.

The numbers tell the story of runaway adoption. Claude Code spread from 32% of engineers in February to 84% "agentic coding users" by March, hitting 95% monthly AI users by spring. That's viral enterprise adoption that would make any SaaS founder weep with joy. Until the bills came in.

Monthly costs per engineer ranged from $150-$250 on average, with power users hitting $500-$2,000. The CTO himself burned $1,200 in a two-hour session during a personal demo. Uber made it worse by creating internal leaderboards ranking engineers by Claude Code usage, essentially gamifying token consumption.

By spring, 70% of committed code originated from AI tools. Even more striking: 11% of live backend updates were written by agents with no human in the loop. The productivity gains were real, but they couldn't be netted against AI costs in the same quarterly line item.

Here's the timing problem: Anthropic announced on May 13 that paid subscribers will face separate monthly credit meters for agent tools, billed at full API rates starting June 15. GitHub is moving Copilot to credit-based pricing June 1. Meanwhile, Microsoft 365 Copilot Enterprise charges a flat $30/user/month - the exact opposite model.

The industry is split between subscription models (predictable costs, limited usage) and consumption models (unlimited potential, unpredictable bills). Enterprise finance teams are built for the first model. The second model just torched Uber's entire year.

Only 43% of organizations have formal AI governance policies, according to the data. That means most companies are walking into the same trap. CFOs are about to discover what every Fortune 500 finance team will learn the hard way: when AI tools actually work, usage scales exponentially, not linearly.

This isn't a failure story. It's a preview. Uber's engineers found a tool that made them dramatically more productive. They used it. A lot. The productivity gains are measurable and real. The company just had no framework for budgeting what "dramatically more productive" actually costs when you pay per token.

Every enterprise evaluating AI tools needs to ask: what happens when this actually works? Because if your engineers love it as much as Uber's did, your finance team is about to get a very expensive education.