Recently, companies' attitudes toward AI have changed significantly. In the past year, many companies encouraged employees to use large models as much as possible. Now, the more common practice is to tighten permissions, set limits, and calculate costs and outputs together.

Microsoft revoked some of the Claude licenses

Microsoft has announced that it will revoke most Claude Code licenses from its Experiences + Devices division by June 30. This division is responsible for products such as Windows, Microsoft 365, Outlook, Teams, and Surface. Engineers working with this division will be migrated to the GitHub Copilot CLI. Microsoft stated that Claude Code has completed its learning and exploration phase, and the company should now focus more on using its own products.

Amazon and Uber are also doing their math.

The Financial Times revealed that Amazon has canceled its internal AI usage leaderboard because employees were excessively farming tokens to climb the rankings, even performing tasks with no real value. Uber's CTO also stated that the company's engineers used up the entire year's Claude Code budget in just four months. Uber's COO subsequently pointed out that there was no clear linear relationship between token consumption and the ultimately valuable products produced.

Businesses are starting to re-evaluate ROI

These changes aren't limited to major US companies. Duolingo had planned to incorporate AI usage into performance evaluations but later withdrew the arrangement. Meta has also been reported to have tightened internal incentives, shifting from encouraging "more grinding" to placing greater emphasis on actual output. Domestically, Zheng Yinhe of miHoYo mentioned that one Agent project burned through 2 million yuan in token fees in a single night.

Additional information:Axios previously reported that a company once burned through $500 million of Claude's bills in a single month, but its identity has not yet been disclosed.