Uber wanted its engineers using AI. So it built a leaderboard, ranking teams by how much AI they used.
Four months into 2026, the entire year's AI coding budget was gone.
Andrew Macdonald, Uber's president and COO, went on a podcast and said the quiet part out loud. Asked whether the spending was paying off, he didn't reach for a confident number. "That link is not there yet," he said. "It's very hard to draw a line between one of those stats and 'Okay now we're actually producing like 25% more useful consumer features.'"
A leaderboard ranks whatever you tell it to rank. Uber told it to rank usage. Usage is exactly what went up.
The Leaderboard Did Its Job Perfectly
Here's what a leaderboard optimizing for volume actually produces, reliably, every time. Teams compete on the number the leaderboard shows them, not on some other number nobody's tracking. If the leaderboard shows total AI tool usage, teams increase total AI tool usage. That isn't a failure of the mechanism. It's the mechanism working exactly as designed.
This is worth stating plainly because it isn't a new lesson, just a newly expensive version of an old one. Sales organizations learned decades ago that ranking reps on calls made rather than deals closed produces more calls and no more revenue. Customer service organizations learned that ranking agents on tickets closed rather than problems actually solved produces faster closes and angrier customers who call back. Every one of these failures traces to the identical root: a visible, competitive number that measures activity standing in for an invisible, uncompetitive number that measures outcome. Uber's leaderboard is that same failure, running on infrastructure that turns the mistake into a line item instead of a slow leak.
The problem was never that Uber's engineers used the tool badly. The problem is that nobody built a second leaderboard, or any equivalent signal, measuring what the usage actually shipped. Macdonald's own CEO, Dara Khosrowshahi, told analysts on an earnings call that roughly ten percent of Uber's committed code is now built by autonomous agents, and that adoption has spread past engineering into legal and marketing teams. Real usage, confirmed by the company's own leadership. What Khosrowshahi didn't have, and what Macdonald admitted to a podcast host weeks
later, was a matching number for value delivered.
Two executives, two honest numbers, and neither one connects to the other. That gap is not a coincidence. It's what happens when an organization builds a scoreboard for one variable and assumes a second, unmeasured variable will rise alongside it out of goodwill.
Why Gamifying Adoption Backfires In A Way Training Never Would
A leaderboard is a stronger intervention than most people give it credit for, which is exactly why it's dangerous to point at the wrong target. Training nudges behavior. A leaderboard commands it, because nobody wants to be the team at the bottom in a public ranking their peers and their boss can see. That's precisely why it worked so well at driving usage up, and precisely why it was reckless to run without a companion measure for output.
Consumption-based AI pricing turns this into a financial problem on a much faster clock than a normal incentive misfire would. A leaderboard optimizing the wrong number in a fixed-cost tool is embarrassing. The same leaderboard optimizing the wrong number in a tool billed by the token is a budget line item burning down in real time, and Uber found out how fast that clock runs: four months, not four quarters.
Compare that to what a misdirected incentive costs in a business with fixed-cost tools. A sales team gamed to maximize call volume instead of closed deals wastes attention and morale, but it doesn't blow through an annual budget line by springtime. Bill by usage, and the exact same incentive design flaw becomes a cash problem with a visible countdown, discovered by finance rather than by the people who built the leaderboard in the first place. That's the part of this story that generalizes well beyond Uber: any organization pricing AI by consumption is running every internal incentive through a meter, whether it built the incentive on purpose or backed into it through a well-intentioned adoption push.
Microsoft hit a version of the same wall from a different angle. The company began canceling most of its direct Claude Code licenses, moving engineers instead to a different tool entirely. Different mechanism, same underlying lesson: usage that isn't tied to a measured outcome is a cost center wearing a productivity metric's clothes, and once finance notices, the response is often a blunt one, a canceled license or a capped budget, rather than the more precise fix of building the missing measurement in the first place.
What The Missing Leaderboard Should Have Measured
The fix isn't fewer leaderboards. It's a second one, sitting next to the first, tracking the thing Macdonald couldn't produce a number for: features shipped, defects resolved, cycle time on a specific workflow, whatever output the business actually sells. Rank teams on that number too, publicly, with the same visibility as the usage leaderboard, and the incentive stops rewarding volume for its own sake.
This is not a call to slow down or reverse the rollout. Uber's engineers did exactly what the incentive told them to do, and doing that well is not the failure here. The failure sits one level up, with whoever approved a leaderboard that could only ever measure effort, never result, and called that measurement enough.
Before your organization builds its own version of this leaderboard, or discovers it already has one buried inside a well-meaning adoption dashboard, ask the question Uber's own president had to answer live on a podcast: if someone asked you right now to draw a line from AI usage to a specific business outcome, could you actually draw it, or would you find yourself saying, on the record, that the link isn't there yet.
The honest version of that answer is worth more than another quarter of rising usage numbers. Uber's leaderboard did not lie to anyone. It reported exactly what it was built to report, faithfully, right up until the budget ran out and a reporter asked the question the dashboard was never designed to answer.
References
Angelo, Jake. "Uber's COO Says It's Getting Harder to Justify the Company's AI Spend: 'That Link Is Not There Yet.'" Fortune, fortune.com, May 26, 2026.
