What happened
Uber CTO Praveen Neppalli Naga revealed in early August that the company had burned through its entire 2026 AI budget by April after incentivizing unlimited Claude Code usage with engineer leaderboards. The solution: Agentic Pods—two-person teams pairing one AI-proficient engineer with one domain expert from Finance, Marketing, Ops, Legal, or HR. Each pod runs a 10-day sprint: shadow the expert, build alongside them, ship an agent. Uber ran 16 pods across 16 business functions in two months.
The numbers
Financial pacing reports dropped from two full days to 10 minutes. Capital allocation analysis across 150 cities fell from 15 hours to 30 minutes. Marketing quality assurance went from two weeks to under an hour. Despite quadrupling the number of employees using frontier AI tools, Uber's AI costs are now decreasing, not rising. Naga called it "the end of the tokenmaxxing era"—treating efficiency as an engineering problem, not a budget problem. The company improved prompt caching, adjusted default models, and gave engineers real-time visibility into their AI usage costs.
Why it matters
This is the counter-narrative to "just use the API." Uber didn't solve the cost crisis by switching vendors or negotiating discounts—they built agents that automate workflows using models as tools, not oracles. Short, embedded sprints with domain experts let them ship pragmatic automation wins that earn the right to swing bigger. Enterprises that hand frontier models to everyone and hope productivity appears are renting intelligence. Uber is buying it back, one 10-day sprint at a time, by owning the agent layer.