Uber gave its engineering team access to Claude Code, a popular AI tool that helps write software, in December 2025. Adoption was fast. By March 2026, 84% of Uber's roughly 5,000 engineers were active users. The company's internal leaderboard, which ranked teams by total AI tool usage volume, pushed adoption even faster. By April, the entire year's AI tools budget was gone.
The cost structure is what made this so hard to control. Claude Code charges a base fee per seat plus a usage charge for every request made. The more engineers use it, the higher the bill, with no fixed ceiling. At Uber, monthly costs per engineer ran between $500 and $2,000. Multiply that across 5,000 engineers for several months and the numbers become very large, very quickly.
The more damaging problem is what came next. Uber president Andrew Macdonald went on the record saying the company cannot draw a clear line between what it spent and anything better reaching its customers. About 25% of Uber's code commits last quarter came through Claude Code, but Macdonald said that metric tells him nothing about whether the app improved. He has now signalled that Uber will weigh AI tool costs directly against the cost of simply hiring more engineers.
Uber is not alone in hitting this wall. Microsoft cancelled most of its Claude Code licences in May, giving engineers a deadline of June 30, which is the last day of Microsoft's financial year. The stated reason was toolchain consolidation, pushing teams toward GitHub Copilot CLI, Microsoft's own product. The practical reason was cost. Claude Code's open-ended usage pricing was driving bills up in ways that flat-rate alternatives do not. Microsoft's CEO had said earlier this year that AI now writes around 30% of the company's code, which gives you a sense of the scale involved.
The broader picture across the enterprise is consistent. A recent survey found that fewer than 1% of companies report significant returns from AI, while most report only 1 to 5% productivity gains. Separately, around 42% of enterprises now say optimising AI spend is their top priority for 2026, having overtaken simply expanding AI use. The free-spending exploration phase is closing.
For anyone outside the tech industry watching this: the pattern here is familiar from other technology waves. Companies rushed to adopt AI tools under pressure to keep up, often without clear metrics for what success would look like. Usage became the proxy for value. Now the bills are landing and the question of actual business impact is overdue.
The practical takeaway is straightforward. If your organisation is spending on AI tools, usage volume is not a measure of value. Lines of code written, documents drafted, or prompts answered are activity metrics. The only question worth asking is whether the output changed something that matters to your customers or your bottom line. If you cannot answer that, you have the same problem Uber has, just at a smaller scale.