The rush to make artificial intelligence part of everyday software development is facing a new reality: AI can be expensive when employees use it at scale.
Companies including Microsoft, Uber and Amazon have reported or taken measures to control rising costs associated with AI coding tools and autonomous AI agents. The developments are not a sign that companies are abandoning AI altogether, but they show that businesses are becoming more careful about how much they spend on usage-based AI services.
Uber Burned Through Its Annual AI Budget in Four Months
One of the clearest examples has come from Uber. The company’s Chief Technology Officer, Praveen Neppalli Naga, said the company had exhausted its entire 2026 AI coding budget by April, just four months into the year.
Uber had expanded access to AI coding tools, particularly Anthropic’s Claude Code, among its engineering teams. As adoption increased, employees began using the tools for tasks including writing code, testing, refactoring and working on multiple AI agents.
The rapid increase in usage created a much larger bill than the company had originally expected. Reports said Uber subsequently introduced a $1,500 monthly spending limit per employee for individual agentic coding tools, including Claude Code and Cursor. Employees can monitor their usage through an internal dashboard, with exceptions possible in certain cases. :contentReference[oaicite:0]{index=0}
Uber’s experience highlights an important difference between traditional software subscriptions and modern AI services. While a conventional software licence may have a predictable cost per employee, AI agents can generate significantly different bills depending on how heavily they are used and how many tokens they consume.
Microsoft Pulls Back Claude Code Access
Microsoft has also moved to reduce its internal use of Claude Code. After giving employees access to Anthropic’s coding tool, the company reportedly began cancelling most direct Claude Code licences in parts of its organisation and directing engineers toward GitHub Copilot CLI.
The decision was not simply about abandoning AI. Microsoft continues to invest heavily in artificial intelligence and has also expanded the availability of Claude through Microsoft 365 Copilot. Instead, the company’s internal changes reflect a push to control costs and consolidate developer tools around Microsoft’s own ecosystem. :contentReference[oaicite:1]{index=1}
Microsoft has also introduced tighter controls around AI token usage. Recent internal guidance reportedly warned employees against excessive AI consumption, with the company focusing on getting more business value from each token rather than simply encouraging employees to use AI as much as possible. :contentReference[oaicite:2]{index=2}
Amazon Reports $1.8 Million AI Cost Overrun
Amazon provides another example of how quickly AI expenses can grow when autonomous systems are allowed to operate at scale.
One internal Amazon project using Claude reportedly accumulated approximately $1.8 million in costs, significantly exceeding its original budget. The project was designed for a relatively routine task involving product data, but its token consumption continued to increase without being detected quickly enough.
Reports said the project represented an 860% budget overrun. Amazon has described such incidents as part of its experimental use of AI rather than normal operating expenditure, but the episode has nevertheless highlighted the difficulty of predicting costs when AI agents can perform large numbers of tasks automatically. :contentReference[oaicite:3]{index=3}
Other Companies Are Also Watching AI Bills
The issue is not limited to Claude. Technology companies are increasingly finding that the cost of AI depends heavily on how employees use the tools.
TechCrunch reported that an employee at Priceline said a routine renewal for its Cursor AI contract came back at roughly four to five times the previous cost. The publication also reported that companies across the industry were beginning to reconsider unlimited or loosely controlled AI usage as token consumption increased. :contentReference[oaicite:4]{index=4}
The problem becomes more pronounced with AI agents. Unlike simple chatbots that answer a question and stop, coding agents can repeatedly read files, search documentation, generate code, run tests, identify errors and make further changes. Each step can consume additional tokens and increase the final bill.
Are Companies Moving Back to Human Workers?
Despite headlines suggesting that companies are “moving back to humans”, the reality is more complicated.
Microsoft is not replacing Claude with human programmers; it is moving engineers toward GitHub Copilot and tightening internal AI spending. Uber has not abandoned AI coding either, but has introduced spending controls. Amazon continues to experiment with AI despite individual projects exceeding their budgets.
At the same time, some companies are becoming less interested in using AI simply for the sake of saying they are an “AI-first” organisation.
Duolingo, for example, backed away from using employees’ AI usage as a performance metric after workers questioned whether they were being encouraged to use AI even when it did not improve their work. CEO Luis von Ahn said the company ultimately decided that employees should be judged on how well they perform their jobs, rather than how much AI they use. :contentReference[oaicite:5]{index=5}
The AI Industry’s New Question: Is It Worth the Cost?
The latest developments suggest that the next stage of the AI race may not simply be about which company uses the most AI. It could increasingly be about which company gets the highest return from every AI dollar spent.
AI coding assistants can deliver substantial productivity gains. A recent study of Microsoft’s early 2026 rollout of Claude Code and GitHub Copilot CLI found that engineers who adopted the tools merged approximately 24% more pull requests during the study period, although the researchers noted that merged pull requests are only a proxy for output and do not necessarily measure the value of the work. :contentReference[oaicite:6]{index=6}
That creates a difficult balancing act for technology companies. AI can allow engineers to complete more work, but if usage costs rise faster than productivity gains, the financial benefit can quickly disappear.
The lesson from Microsoft, Uber and Amazon is therefore not that companies are giving up on AI. Instead, the industry appears to be entering a more cautious phase — one where AI usage must increasingly justify its cost.
For businesses that once encouraged employees to use AI without strict limits, the era of the unlimited AI budget may be coming to an end.
