Uber’s Claude Code bill exposes the harder question for workplace AI

Uber has reached the less glamorous part of the AI cycle: the bills are visible, and the productivity gains still have to be proven. According to The Verge, the company reportedly exhausted its annual AI budget in the first four months of 2026. That would be manageable if the return were obvious. The problem, as Uber president and chief operating officer Andrew Macdonald told Rapid Response, is that the company is not yet seeing a clear enough connection between spending and output.

The most concrete signal concerns Claude Code, Anthropic's coding assistant. Macdonald's point is not that AI has no value. It is that higher token consumption does not automatically translate into more useful work. Engineers may use more AI assistance, systems may generate more responses, and inference bills may rise, but that does not prove that more reliable software shipped, that problems were solved faster, or that expensive operational bottlenecks disappeared.

That is a meaningful shift in tone. In the first phase of corporate AI adoption, adoption itself often became the metric: how many teams used the tool, how many prompts were sent, how many internal prototypes were launched. Uber's case exposes the weakness of that logic. Tokens are a cost, not an outcome. If they are not tied to a measurable business result, they become another cloud expense wrapped in futuristic language.

For Uber, the question is especially sensitive because the company runs a vast operational network: rides, delivery, dynamic pricing, support, safety processes and internal engineering work. AI should, in theory, have many places to add value inside a system like that. That is exactly why Macdonald's skepticism matters. If an organization with large data flows, large engineering teams and constant pressure to improve efficiency cannot easily demonstrate return, smaller AI buyers will have an even harder time making the case.

This is not the end of enterprise AI. It is the end of the period in which nearly any AI spend could be defended as a strategic experiment. Large companies will increasingly ask sharper questions: which team actually saves time, which workflow produces fewer errors, how much manual work disappeared, how much faster the delivery cycle became, and who pays when model usage outruns the plan. In that sense, Uber's comment is not just a budget note. It is a market signal to AI tool vendors.

For Anthropic and the wider AI infrastructure layer, the message is direct. Tools such as Claude Code must prove themselves not only through answer quality, but through the economics of real work inside real teams. If usage grows faster than measurable impact, customers will start demanding limits, internal policies, more selective deployment and clearer value reporting. AI is entering a more normal and stricter phase: fewer transformation slides, more questions about the invoice.