Claude: $500 Million Spent in One Month

A company recently spent the staggering sum of $500 million in just one month on the AI model Claude. This astronomical expenditure occurred after the company failed to impose usage limits on Claude's licenses. Typically, AI models for businesses are offered with flat-rate pricing, but these plans often include restrictions on the number of allowed queries.

Uber's Chief Operating Officer expressed concerns about the growing difficulty of justifying AI expenses, especially when the return on investment remains unclear. In a similar context, Microsoft reportedly recently reduced its internal licenses for Claude Code, citing strategic reasons and rising costs.

Another tech executive noted that some employees are using AI systems for tasks as simple as checking the weather, which turns out to be much more expensive than a traditional search. Sophia Velastegui, former head of AI at Microsoft, observed that companies tend to use AI for undesirable tasks rather than for revenue-generating activities.

These examples illustrate a recurring problem: when AI becomes essential to a company's revenue generation, it is crucial to have skilled personnel to effectively use and guide these systems. Emerging roles, such as AI agent orchestrators, are set to become critical.

High costs are often due to misuse and poor model selection. Inefficient usage results in a lack of contextual engineering, leading to unnecessary and costly interactions. Choosing a powerful model for simple tasks that cheaper models could handle is also a source of unnecessary expenses.

It is important to note that not all tasks require the use of generative AI, particularly for language or reasoning models. Many tasks are better managed by traditional software. Developing this skill is essential for any company looking to optimize its use of AI.

Finally, the impact of AI is not limited to costs. The quality of outcomes can also suffer if these technologies are not mastered. A recent example shows that Copilot, in automatic mode, failed in a data analysis task, producing biased results. Switching to a reflective model corrected the error, highlighting the importance of thoughtful AI usage.