Uber's CTO has said that 99% of Uber engineers use AI tools. Over 70% of pull requests are attributed to agents. More than 2,500 agent skills built in-house across the software development cycle. Impressive numbers, but Naga says the real problem wasn't there: it was in Finance, Legal, Operations, Marketing, Customer Support, HR and Procurement. Functions that run on manual workflows, full of exceptions and judgement, spread across dozens of systems. And that can't be automated by reading process diagrams, because the diagrams don't tell you how the work actually gets done.
So Uber created Agentic Pods: 30 of the company's most AI-experienced engineers, each paired with a domain expert from a business function. Two weeks per pod, with a rigid structure. Days one and two: zero code, observation only (the engineer sits next to the person doing the work, documents every step, understands where the friction is). Day three: opportunities are prioritised by scale, repeatability, impact and data availability. Days four and five: a working agent is built together with the person who actually does that job. Days six to nine: validation with others doing the same job, to see whether it generalises. Day ten: it goes into production.
In two months, 16 pods across 16 functions. Capital allocation across 150 cities: from 15 hours to 30 minutes. Financial pacing reports: from 2 days to 10 minutes. Marketing web QA: from 2 weeks to 50 minutes. 9,000 manual customer support workflows replaced by self-service automation.
But what does Naga say about the biggest lesson? That the real gains don't come from automating a task, they come from rethinking the entire workflow. And rethinking a workflow, in practice, means eliminating handoffs, removing approvals that no longer earn their place, retiring legacy tools, cutting vendor spend.
Stop and look at that list. Removing an approval isn't a technical act: it's someone losing a power of veto. Retiring a legacy tool means a budget changes owner. Eliminating a handoff means a boundary between two teams ceases to exist. Who decides that that approval dies? Who answers for it if the agent gets it wrong? Who loses a piece of their role when an activity goes from 15 hours to 30 minutes? The pod works because someone, above the pod, has already answered these questions. They gave the engineer + expert pairing the authority to touch the process, not just to observe it. And the ten-day timebox isn't an agile-methodology affectation: it's a constraint that makes politics impossible. In ten days you either ship or you've failed, and there's no room for the committee that defends the status quo.
This is why I say the pods aren't a method you can replicate like a recipe. They are organisational design: an explicit redistribution of authority, with a clear mandate and a deadline. Copying the recipe without copying the mandate produces yet another pilot project that dies in demo.
There's also a crack that Naga's post doesn't mention, and that makes the case even more instructive. Uber exhausted its annual Claude Code budget months before the end of the year. And in May, Andrew MacDonald, president and COO, publicly admitted that the scale of AI spending is becoming hard to justify: internal productivity is growing, but it hasn't yet translated into proportional improvements for riders and drivers. Even those doing pods better than anyone else haven't yet closed the economic loop. Adoption is running faster than the economic design, and sooner or later the bill arrives. Uber is still writing this part, and I don't have the answer ready any more than they do.
In Italy the picture is the one Dattoli describes in a recent LinkedIn post, and it matches what we see: companies buying tools without ever redesigning the workflow. "We've integrated AI" often means: we've laid a copilot on top of a process that wasn't working in the first place. The ILLUMINA research has been telling us for months that the majority of Italian companies are stuck at POC stage. And they stay there for a precise reason: a POC doesn't require redistributing anything. No approval dies, no boundary moves, nobody loses a veto. That's why it's comfortable. And that's why it doesn't produce ROI.
This is the piece we work on every day at FAIRFLAI: before building agents, we map how the work actually gets done by sitting next to the people doing it, exactly like Uber's pods - and then we explicitly negotiate who has the authority to change what. The uncomfortable part isn't the prompt. It's the conversation in which you decide which approval gets eliminated and who takes the consequences if something goes wrong. The cost has to be stated, because it's real. Doing proper pods costs more than buying licences: it means taking a function's best person away from it for two weeks, and it takes a sponsor willing to make a few enemies when the redesigned workflow steps on the toes of whoever had built a piece of power on that workflow. Anyone not ready to pay this price will do a POC. Which is an elegant way of not deciding.
And a question for those reading: in your company, who has the authority today to eliminate an approval? If the answer is "nobody, we'd need a meeting", you already have your diagnosis. Let's talk.
