AI costs are starting to worry tech companies as heavy token use cuts into productivity gains

Tech companies have spent the past year pushing employees to use AI tools more often, but the cost of that usage is now becoming harder to ignore. A new report says companies including Microsoft, Meta, and Amazon are seeing problems as internal AI use grows, especially when employees rely on expensive agentic AI tools for everyday work.

The issue is not only that AI costs money. The bigger problem is that people are using far more tokens than expected. Token prices may be falling as AI systems become more efficient, but total usage is rising so quickly that the savings can disappear. Agentic AI can be especially costly because it may use many more steps to complete a task than a normal chatbot response.

Cheaper AI tokens do not help much when employees use far more of them

One example mentioned in the report is Microsoft, where employees were reportedly encouraged to switch from Claude Code to Microsoft’s own Copilot CLI. The public reason was that Microsoft preferred an internal tool, but the cost of increasing Claude Code usage was also reportedly a major factor.

The same pattern is appearing elsewhere. As companies ask workers to use AI more, some employees start using it for nearly everything. This behavior is being described as “tokenmaxxing,” where people push AI usage higher to meet internal goals or show that they are adopting the technology.

Agentic AI is a major reason costs can climb quickly. Unlike a simple prompt and answer, agentic tools may plan, search, read files, write code, test results, retry, and revise. Each step consumes tokens. In some cases, that can make a task far more expensive than expected. The report mentions OpenClaw creator Peter Steinberger claiming that his team spent more than $1.3 million in token costs in just one month.

This creates a difficult question for companies. AI can help with coding, research, writing, support, and automation, but the return depends on whether the tool saves enough time or money. If a task uses thousands of extra tokens and only saves a few minutes, the business case becomes weaker.

The situation also resembles the Jevons Paradox. When a technology becomes cheaper or more efficient, people often use much more of it. That can make total consumption rise instead of fall. AI may now be facing the same pattern. Lower token prices encourage wider use, but wider use creates larger bills.

There is also a cultural problem. If companies reward employees for using AI, some workers may use it just to increase their internal usage numbers. The report says this has happened at Amazon, where some employees admitted using AI for unnecessary tasks to boost usage scores. Similar behavior has reportedly appeared at Microsoft and Meta.

That does not mean AI is useless. It means companies may need more careful rules. AI is valuable when it solves real problems, speeds up complex work, or handles repetitive tasks. It becomes wasteful when employees use it only because management wants higher adoption numbers.

The broader lesson is simple. AI adoption cannot be measured only by how many tokens employees consume or how often they open a tool. Companies need to measure whether AI actually improves output, saves time, reduces errors, or lowers costs.

For now, the AI cost crisis shows that the technology is entering a more practical phase. The early excitement around adoption is giving way to questions about efficiency, budgets, and real productivity. Tech giants may still keep pushing AI, but they are also learning that unlimited token use can become expensive very quickly.