Enterprise AI spending surges as companies struggle to prove ROI

In December 2025, Uber provided its engineers with Claude Code and established internal leaderboards to track token consumption. [1]

By April, Uber's entire AI coding budget for 2026 was completely depleted. [1]

Uber President and COO Andrew Macdonald noted that there was no link yet between this high usage and shipping better products for drivers and riders. [1]

This episode highlighted a broader problem known as "tokenmaxxing," which involves surging token consumption without the return on investment to justify it. [1]

To control costs, Uber now caps spending at $1,500 per employee each month per agentic coding tool. [1]

Gartner projects that spending on AI agent software will approach $207 billion this year, representing a 139% increase from $86.4 billion in 2025. [1]

Unlike traditional software costs, token pricing can produce vastly different invoices for the same engineer depending on whether they are auto-completing suggestions or running parallel agents. [1]

Microsoft canceled Claude Code licenses across its Experiences and Devices division after questioning their cost. [1]

Additionally, Duolingo reversed a plan to factor AI usage into performance reviews following employee pushback. [1]

The shift from unchecked AI experimentation to strict budget caps marks a critical maturation phase in enterprise AI adoption. Uber's aggressive initial push—tracking token usage on leaderboards—demonstrates how companies initially conflated high utilization with productivity. The subsequent pivots by Uber, Microsoft, and Duolingo suggest that the industry is abandoning AI adoption for its own sake. With Gartner projecting agentic software spending to hit $207 billion, CFOs are being forced to model highly variable token costs that defy traditional SaaS billing structures. What the sources don't address: how organizations plan to definitively measure the ROI of AI tools now that usage alone is no longer considered a reliable metric of success.