Artificial intelligence is transforming software development at an unprecedented pace, helping engineers write code faster, automate repetitive tasks, and boost productivity. But for Uber, the rapid adoption of AI-powered coding assistants has come with an unexpected challenge: soaring costs.

According to a Bloomberg report, the ride-hailing giant has introduced spending limits on popular AI coding tools such as Anthropic’s Claude Code and Cursor after exhausting its annual AI budget within the first four months of the year.

The move highlights a growing dilemma facing technology companies worldwide. While AI tools promise significant productivity gains, their usage costs can quickly spiral out of control as adoption scales across thousands of employees.

Credits: Investopedia

New Spending Caps for Employees

To keep expenses under control, Uber has implemented a monthly spending limit of $1,500 per employee for each AI coding tool.

The cap applies separately to individual tools, meaning employees who exhaust their allocation on one platform can continue using other approved AI coding assistants. Uber has also reportedly introduced an internal dashboard that allows workers to monitor their AI token consumption in real time.

Employees who require additional usage for their projects can seek approval to exceed the standard budget allocation.

An Uber spokesperson described the decision as a balanced approach to managing costs while continuing to encourage experimentation with emerging AI technologies.

“We think this is all a pretty straightforward way to responsibly encourage agentic AI adoption and experimentation at scale across the company,” the spokesperson told Bloomberg.

A Budget Crisis Triggered by Claude Code

The spending limits come after Uber’s leadership acknowledged that AI adoption had far exceeded expectations.

Back in April, Uber Chief Technology Officer Praveen Neppali Naga revealed that the company had effectively gone “back to the drawing board” regarding AI spending after unexpectedly blowing through its entire annual budget. The surge was largely driven by growing employee reliance on AI-powered coding assistants, particularly Claude Code.

The incident demonstrates just how quickly AI tools have become embedded in software engineering workflows. What began as productivity experiments have rapidly evolved into essential daily tools for many developers.

AI Is Already Writing a Significant Portion of Uber’s Code

Despite the budget concerns, Uber’s AI transformation appears to be progressing at remarkable speed.

In a recent post on X, Naga revealed that the company’s internal background coding agent is now responsible for approximately 1,800 code changes every week. He also noted that around 95% of Uber engineers use AI tools at least once every month.

Even more striking, Uber CEO Dara Khosrowshahi recently disclosed that nearly 10% of the company’s code is now submitted and built directly by AI agents.

These figures suggest that AI is no longer simply assisting engineers—it is increasingly becoming an active participant in software development.

Microsoft Is Also Reassessing AI Tool Usage

Uber is not alone in reevaluating its AI spending strategy.

Just a day before reports emerged about Uber’s new restrictions, Microsoft reportedly informed employees that support for Claude Code would be phased out. According to The Verge, Microsoft plans to discontinue employee access to the tool by June 30 and is encouraging workers to transition to GitHub Copilot CLI instead.

The decision reflects a broader trend among technology companies that are seeking to consolidate AI investments around preferred platforms while keeping operational costs under control.

The Big Question: Does More AI Mean Better Products?

While AI-generated code continues to increase, some executives remain unconvinced that higher usage automatically translates into better outcomes for customers.

Uber Chief Operating Officer Andrew Macdonald recently raised this concern during an episode of the Rapid Podcast. Although approximately 25% of Uber’s code commits reportedly came through Claude Code last quarter, he questioned whether those productivity gains are directly leading to more valuable features being delivered.

According to Macdonald, measuring AI’s true impact remains difficult. While teams may be shipping more code than ever before, establishing a clear connection between AI-generated output and meaningful customer improvements is still a challenge.

As companies race to embrace AI-powered software development, Uber’s experience offers an important lesson: adoption is easy, but proving business value—and managing costs—may be the harder task. The next phase of the AI revolution could be less about how much code machines can write and more about whether that code creates better products for users.