Uber is not rejecting AI, but it is asking the harder question every company will soon face: what exactly is all that usage buying?
Uber has built one of the most AI-dependent consumer businesses in the world, so Andrew Macdonald's latest comments land with more force than a routine executive complaint about software bills. The company's president and chief operating officer said Uber is finding it harder to justify rising AI costs when it cannot draw a clean line from token usage to more useful features for riders, drivers, couriers, restaurants, or merchants.
That is the part worth watching. This is not a legacy company discovering ChatGPT late and wondering why the pilot project did not change the business. Uber already uses machine learning across pricing, routing, matching, fraud detection, demand forecasting, logistics, and support. If a company with real operational uses for AI is asking where the return shows up, the question is moving from skepticism to management discipline.
In a Rapid Response interview published by Masters of Scale, Macdonald said the headline metrics around AI can make executives' heads explode: more AI-generated code, more token usage, more employee adoption. But after talking with senior engineering leaders, he said the missing piece is still attribution. It is hard to say that because a certain share of code commits came through Claude Code, Uber is now producing a proportional increase in useful consumer functionality.
The timing matters because Uber's own AI tooling bill has already become a story inside the industry. CTO Praveen Neppalli Naga previously told The Information that Uber had blown through its 2026 AI budget early in the year, a disclosure that quickly circulated among engineers and enterprise software buyers. DesignRush later reported that Claude Code adoption had spread rapidly across thousands of Uber engineers, with monthly spending varying sharply between average users and heavy users.
That is the problem with token-based economics. A traditional software license is annoying, but at least finance teams know roughly what it costs. Agentic coding tools are different. One engineer using a model for basic suggestions might consume modest amounts. Another asking agents to inspect large codebases, write tests, refactor services, and run repeated loops can generate a bill that behaves less like a subscription and more like cloud infrastructure under stress.
For a while, companies could treat that as the price of staying current. AI adoption looked like the metric that mattered. Leaders wanted employees to use the tools, vendors wanted usage to rise, and teams often treated more AI activity as evidence of transformation. But usage is not the same thing as value. A restaurant does not become more profitable because the kitchen burns more gas. It becomes more profitable when the food is better, faster, or cheaper to produce.
Macdonald's comment is important because it shifts the conversation away from whether AI can produce work. It clearly can. The harder issue is whether the work improves the product enough to justify the expense. More code can also mean more review, more integration work, more bugs, more duplicated features, and more maintenance. In software, shipping faster is useful only when the right things are being shipped.
The wider AI market should pay attention
Uber is not alone in this tension. Business Insider noted that Macdonald's remarks come as companies such as Duolingo have also had to rethink how aggressively they push employees to use AI. Duolingo CEO Luis von Ahn has spoken publicly about walking back AI usage as a performance-review input after employees questioned whether they were being pushed to use AI for its own sake.
That is a practical lesson for every company buying AI tools in 2026. The easy dashboard shows tokens consumed, prompts sent, seats activated, and code generated. The useful dashboard shows cycle time, defect rates, customer conversion, support resolution, revenue per employee, infrastructure savings, and products that were previously impossible to build. One set of numbers flatters adoption. The other tests whether the adoption is worth paying for.
This is also where vendors should expect tougher procurement conversations. Anthropic, OpenAI, Cursor, GitHub, Google, Microsoft, and the cloud platforms all benefit when AI usage becomes habitual inside large companies. But large customers eventually push back when costs become unpredictable. The novelty premium fades quickly once a CFO asks why the bill is growing faster than the measurable output.
None of this means Uber is pulling away from AI. Macdonald also described AI as a major long-term frontier for Uber, especially because the company sits between digital systems and the physical world. Autonomous vehicles, delivery logistics, marketplace balancing, and local commerce remain deeply tied to advances in intelligent software. The point is narrower and more serious: AI has to compete for capital like everything else.
The next phase of enterprise AI will probably be less about access and more about proof. Companies will still buy the tools, but they will demand guardrails, model choice, usage caps, internal chargebacks, and better measurement. Uber has given other executives permission to say the quiet part plainly. If the bill is real, the return has to be real too.
Also read: ByteDance raises the price of keeping its AI talent • China's AI travel curbs show the real doomsday risk is fragmentation • India and the U.S. are turning rare earths into an AI supply chain bet
