Uber's AI strategy isn't chasing moonshots. It's grinding.
CEO Dara Khosrowshahi spelled out the approach on the company's second-quarter earnings call: thousands of incremental improvements rather than a handful of bold bets. The numbers suggest it is working—AI spending stayed roughly flat even as adoption accelerated.
At a glance
- Uber contained AI costs by optimizing defaults per use case and using less expensive models
- The company lets employees track their own tool usage
- CTO previously said Uber exhausted its Claude Code allocation for 2026 within months
- Destination suggestions are correct roughly 75% of the time
- Self-driving trips remain under 0.5% of 300 million weekly rides
- Customer service headcount dropped about 10% last month
The cost discipline piece
Balaji Krishnamurthy, who leads finance at Uber, told analysts after the earnings report that the company had contained its AI expenses through three moves. Teams adjusted default settings for different tasks, switched to cheaper models where accuracy was sufficient, and gave engineers clearer visibility into their own usage patterns. The result: cost per token fell even as more teams used AI tools. Krishnamurthy called the outcome "broadly stable" overall spending.
The urgency behind this discipline became public earlier this year. Praveen Neppalli Naga, Uber's chief technology officer, posted that the company had burned through its entire Claude Code budget for the next several years within a few months. The remark circulated widely in tech circles and, according to people familiar with internal discussions, prompted the company to tighten controls.
Small hits compound into something real
Khosrowshahi offered a blunt assessment of how he views AI adoption. "People expect revolutionary technologies to produce some giant hit," he said. In his telling, that expectation misses how value actually accumulates. "Thousands of small improvements across our platform will drive growth for years," he added.
One concrete data point: when users open the app, AI suggests a destination before they type anything. The model gets it right about three-quarters of the time. The company also highlighted Cart Assistant, a feature that builds a grocery order from a recipe or list. Shoppers using it load carts roughly double the size of those who skip the tool.
The numbers on autonomous vehicles tell a different story. Robotaxis handle hundreds of thousands of rides weekly—fewer than half a percent of the 300 million trips booked through the app in the same window. The physics of the problem, plus regulatory approvals, constrain how fast that fraction can grow.
Why this matters for how teams adopt AI
The Uber case offers a useful reality check. The company isn't running a grand transformation program. It identified specific jobs—predicting destinations, populating shopping carts, routing customer requests—and picked the right tool for each. The CTO's viral budget overrun didn't trigger an AI pause. It triggered better defaults and employee visibility.
Consider a parallel: you manage a product team evaluating AI for sprint planning or backlog refinement. A vendor demo shows impressive results on a curated dataset. Before signing, you ask what the tool costs per token, who on the team controls access, and whether a smaller open-source model could handle the task with acceptable accuracy. That is the Uber question applied to your context.
The company also showed willingness to cut roles when automation delivers. Customer support headcount fell roughly 10% after AI handles more interactions. Krishnamurthy flagged support quality as a place where automation should lift results and trim effort simultaneously. That trade-off—headcount versus automation quality—will land on more managers' desks as these tools mature.
Bottom line
Uber is not betting its future on a single AI breakthrough. It is running an iterative portfolio of small wins and measuring whether each one earns its keep. That discipline kept spending in check and, if Khosrowshahi is right, will compound into real competitive advantage.
For project managers and team leads, the takeaway is practical: find the narrow problem, pick the right model, track the cost, and expand only what works. Moonshots are for press releases. Small hits pay the bills.
Key takeaways
- Uber controlled AI costs by optimizing defaults, choosing cheaper models, and giving engineers visibility into usage
- The CTO's viral post about overspending on Claude Code prompted a spending reset
- Destination prediction and grocery cart features show how targeted AI deployment delivers measurable results
- Self-driving rides are growing but remain a tiny slice of total volume
- Customer service headcount dropped after AI absorbed more repetitive work
Frequently asked questions
How did Uber manage to keep AI spending flat?
The company adjusted default settings for different tasks, deployed less expensive models where appropriate, and gave employees clearer sightlines into their own tool consumption.
What happened with Uber's Claude Code budget?
The CTO announced publicly that the company had consumed its full Claude Code allocation for the next several years within just a few months of the year starting.
How accurate is Uber's destination prediction feature?
The AI guesses correctly about 75% of the time when customers first open the app to book a ride.
What share of Uber rides involve self-driving cars?
Robotaxis handle under 0.5% of the roughly 300 million trips booked through the platform each week.
Why did Uber cut customer service staff?
The company reduced customer support headcount by about 10%, citing efficiency gains from AI handling a larger share of routine interactions.
Inspired by reporting from: businessinsider.com