Uber CTO Praveen Neppalli Naga pulled 30 of the company's most AI-proficient engineers out of their normal jobs and embedded them across business teams.
Naga calls them "Agentic Pods." For two weeks at a time, engineers sat with teams in finance, legal, and HR, watched how people actually got work done, and built AI agents to automate the most painful parts. Uber ran 16 of these pods over two months.
The results were hard to argue with. Financial pacing reports that used to take two days now take ten minutes. Allocating capital across the 150 cities Uber operates in went from 15 hours to 30 minutes. "You can't automate them effectively by looking at process diagrams or documentation," Naga said. "You have to understand how the work actually gets done."
The approach was almost comically simple for a company that spent $3.4 billion on R&D last year. Instead of building a self-serve AI platform or handing out a prompt library, Uber physically moved engineers into departments they'd never set foot in and had them watch. They spotted inefficiencies that insiders had stopped noticing. Uber's finance team had actually been out in front on AI adoption, with tools like Finch, a conversational agent that lets employees query financial data in plain English. The pods just scaled what was already working.
Then the costs showed up.
Uber rolled out Claude Code to roughly 5,000 engineers in December 2025. By April, the company had burned through its entire 2026 AI budget. Adoption tripled in a single month, from 32% of engineers in February to 84% by March. Heavy users racked up $2,000 per month, and during one hands-on demo, Naga himself burned $1,200 in tokens in two hours. "I'm back to the drawing board," he said, "because the budget I thought I would need is blown away already."
The blowout makes sense when you understand how agentic AI actually eats resources. A regular chatbot query is one question, one answer. An agentic session generates five to 30 separate model calls per task, and industry estimates suggest that agentic coding tasks can consume roughly 1,000 times more tokens than a standard query. When thousands of engineers start running multi-step agents instead of asking simple questions, costs compound fast.
Even with those efficiency gains, Uber's own COO isn't convinced. Andrew Macdonald said on a podcast in May that while productivity metrics are "trending in a really astronomical direction," he can't connect them to useful features actually reaching customers. "If you're not actually able to draw a direct line to how much useful features and functionality you're producing," Macdonald said, "that's going to be a real challenge."
CEO Dara Khosrowshahi has been more bullish, calling AI tools "superpowers" for employees during Uber's Q1 2026 earnings call. As of last quarter, 95% of Uber's engineers use AI tools monthly, 70% of committed code is AI-generated, and roughly 10% is written by fully autonomous agents. Uber is now forming a dedicated team to scale the pods further.
You can't embed engineers in every department of every company, and that's what makes Uber's success story so tricky to replicate. The thing that worked required pulling expensive talent out of their jobs and physically sitting them in rooms with accountants for two weeks. Someone has to eventually build the tool that gets those results without the engineering rotation. Until that exists, the fastest-moving companies on AI will also be the ones spending the most, and even Uber's leadership can't tell you if it's paying off.
