As AI Use Grows, Companies Begin Putting Limits on the Cost

Companies that once encouraged employees to use artificial intelligence as much as possible are beginning to place tighter limits on that use as the bills become harder to predict.

Uber learned that lesson quickly.

The company gave a new AI coding tool to about 5,000 engineers last December. Employees embraced it, but by April Uber had already spent its entire 2026 AI budget, according to the Financial Times.

The problem was not that the technology failed. It was that the cost of using it grew much faster than expected.

A big reason is the rise of AI agents.

Traditional chatbots generally answer a question and wait for another prompt. AI agents are designed to take on a larger task and keep working through it. They can read files, write code, check their own work, and start over when something goes wrong.

That can turn a single user request into hundreds of separate steps behind the scenes.

AI services are often priced per token, which is a small unit of text. A token is roughly equal to three-quarters of a word.

The cost is based on the amount of text the system reads and produces. With agents, that amount can grow very large because each new step may incorporate information from everything that came before.

Researchers at the Stanford Digital Economy Lab found that agent-based tasks can use up to 1,000 times as many tokens as ordinary chatbot conversations.

Most of that cost can come from what the AI system reads rather than what it writes.

The cost can also vary widely.

According to the research, the same task run twice can differ in token use by as much as thirtyfold. AI models are also not especially good at estimating in advance how many resources they will need.

That unpredictability has led some companies to start limiting employee use.

Uber now caps employee compensation at $1,500 per month per AI tool.

Walmart has placed token limits on its in-house coding assistant.

Microsoft also moved thousands of engineers away from one AI tool this summer after some individual bills reached about $2,000 a month.

The shift marks a departure from how some companies approached AI just a year ago.

At that time, some businesses encouraged employees to use the tools extensively and even maintained internal leaderboards showing who used the most AI.

That practice became known as “tokenmaxxing.”

Now, with companies paying closer attention to costs, a new idea has taken hold: “tokenminimizing.”

The goal is not necessarily to use less AI overall, but to use it more efficiently and avoid incurring large bills for tasks that do not require expensive models or long chains of automated work.

Some companies, however, are still willing to spend freely.

Databricks continues to give its engineers an unlimited AI budget.

That approach is not typical for most businesses.

According to the Ramp AI Index, the median American company spends about $11.38 per employee each month on AI.

At the top end, the difference is dramatic. The top 1 percent of companies spend about $7,500 per employee per month.

That gap shows how uneven AI adoption remains.

For many businesses, AI is still a relatively small expense. For companies that rely heavily on coding agents and other advanced tools, though, the cost can become substantial very quickly.

The challenge is that AI can be most expensive when it is also most useful.

An agent that can work through a complicated coding problem, review its own output and make corrections may save an engineer a great deal of time.

But each of those steps adds more computing and more tokens.

That means companies are increasingly trying to balance productivity gains against unpredictable usage costs.

The new spending limits also suggest that businesses are moving into a more mature phase of AI adoption.

The first stage was often focused on access and experimentation: give employees the tools and see what they can do.

The next stage is looking more like traditional budgeting.

Companies are now asking which tools are worth the cost, how much each employee should be allowed to spend, and whether the most powerful AI models are necessary for every task.

The answer will likely vary by company and by job.

But one thing is becoming clear: AI may feel almost effortless to the person using it, while the cost behind the screen can add up very quickly.