The workplace story around AI used to be simple: give employees access, push usage, and expect productivity to rise. By May 2026 the bills were arriving, and some of the companies pushing AI hardest had started to slow usage down.
Microsoft reportedly started cancelling most licences for Claude Code, Anthropic’s AI coding tool, in parts of its Experiences and Devices division. Many employees were being moved to Microsoft’s own GitHub Copilot tool instead. Claude Code had become popular inside Microsoft after access was opened to thousands of employees. Developers, product managers and designers were using it because it helped them work. As it became part of the working day, the bill became harder to manage.
AI can be useful and expensive at the same time. Microsoft remained deeply committed to AI, yet it still had to control how much its own teams spent on AI tools. Heavy use does not remove the need to prove the return.
Uber shows the same pressure from another angle. The company reportedly burnt through its 2026 budget for AI coding tools within the first four months of the year. It had pushed usage, employees had adopted the tools, and the cost arrived much earlier than expected.
The hard question is whether the extra spend is creating enough extra output. AI systems process text in small chunks called tokens, and the bills rise with them. Andrew Macdonald, Uber’s president and chief operating officer, said it was difficult to draw a clear line between rising token use and more useful features for consumers. Usage can rise quickly without showing that productivity rose with it.
Heavy usage can still show that the tools are becoming part of daily work. If the productivity gains catch up, these cost problems may look like growing pains. But if the bill rises before the business value is obvious, teams will face usage limits and tool switches.
Nvidia adds a warning from inside the AI industry, where the labour-saving story does not always hold. Bryan Catanzaro, its vice president of applied deep learning, said the cost of computing power for his team was far beyond the cost of employees.
