Uber's useful AI story isn't that one finance task now takes 30 minutes. It's that the company got there only after its first AI rollout blew through a full-year budget in four months.
Praveen Neppalli Naga runs engineering at Uber, and in May he became an unlikely cautionary tale. According to The Information and Forbes, Uber burned through its full-year AI coding budget in four months after Claude Code usage spread across its roughly 5,000-person engineering organization. Naga himself said he once spent $1,200 in a two-hour personal demo. That's real money for a demo. Bloomberg later reported that Uber put a $1,500 monthly token cap on individual use of agentic coding tools such as Claude Code and Cursor. The warning was plain. Token-metered AI tools can spend money faster than finance teams can model them.
Now Naga is telling a different story, and this one is more useful if you're trying to make AI work inside a real company. Good. The budget blowout was never the interesting part on its own.
Uber pulled 30 of its most AI-proficient engineers off their normal work and embedded them, in pairs and small teams, inside functions including finance, legal, HR, marketing, customer support and procurement. Each placement runs two weeks. The engineers don't show up with a product pitch. They sit with the team, watch how the work actually gets done, and then build a custom AI agent for that specific workflow. Uber calls the structure an Agentic Pod, and according to Business Insider, it has run 16 of them over the past two months.
That's the whole model. The results, which Naga described in a post on X and Business Insider reported on July 9, are concrete rather than aspirational. Financial pacing reports that used to take a two-person team two days now take 10 minutes. Capital allocation decisions across the roughly 150 cities where Uber operates, previously a 15-hour exercise, now run in 30 minutes. Those aren't projections. They're old workflows with new machinery underneath them.
Naga's broader claim, in the same post, is that 99% of Uber's engineers now use AI tools day to day. More than 70% of pull requests are attributed to local or cloud agents, and engineers across the company have built more than 2,500 agent skills spanning the software development lifecycle. That explains why Uber felt confident enough to move the same muscle into departments where the work doesn't look like software engineering at all.
The Work Was The Product #
Here's the part worth paying attention to if you want to copy this. Naga says the biggest wins didn't come from bolting AI onto a single task. They came from redesigning the whole workflow around it: cutting handoffs between teams, killing approval steps that had outlived their purpose, retiring old tools nobody wanted to defend anymore, and in some cases reducing vendor spend outright. That is not a software rollout. It's management work.
An agent that drafts a report faster is a nice trick. An agent that removes the approval emails, spreadsheet chasing and manual reconciliations sitting between the report and the decision is a different kind of win. That is where Uber says the time disappeared.
That distinction matters because it's the opposite lesson of the budget blowout. The coding-tool overrun happened because usage scaled faster than the company's workflow for controlling it. Engineers just kept prompting. The meter kept running. The Agentic Pods worked, by Uber's own account, because someone spent two weeks understanding the actual bottleneck before writing a line of agent logic.
Naga says Uber is now standing up a dedicated team to do this at scale. The goal is depth: going deeper into individual workflows rather than wider across more departments. Not wider. That's a sharper goal than adding another chatbot to every corner of the company and hoping people find a use for it.
For other CTOs watching Uber's AI spending with a mix of sympathy and dread, the useful data point isn't only the 15-hour-to-30-minute result. A company can spend its way into an AI budget problem and still walk out months later with a repeatable model for turning expensive experimentation into operational work. The two stories aren't a contradiction. They're the same company learning, expensively, that unmanaged AI adoption and deliberately designed AI adoption produce very different receipts. Also read: A Chinese Hacker Used DeepSeek to Attack 460 Systems With One Command • AI Firms Are Buying and Pulping Rare Books After Scanning Them for Training Data • Google Built ChatGPT's Ancestor in 2018 and Watched Its Creator Leave Twice