enterprise AI AI costs Claude AI spending model selection

Half a Billion Dollars in One Month: What Happens When Nobody Watches the AI Tab

2 June 2026

An unnamed company reportedly spent $500 million on Anthropic's Claude in a single month after failing to set usage limits on its AI licenses, highlighting how quickly enterprise AI costs can spiral out of control. Broader industry examples, such as employees using AI to check the weather or misusing large models for simple tasks, point to widespread inefficiency in how companies deploy AI tools. Experts argue that businesses need greater internal AI expertise, better model selection, and smarter usage controls to manage costs and ensure quality outcomes.

Companies are finally starting to squint at their AI invoices, and some of what they're finding is genuinely alarming.

Microsoft recently pulled back on internal Claude Code licences, citing both strategic reasons and runaway costs. Uber's COO has publicly questioned whether AI spending is justifiable when the return on investment remains murky. Neither of those is particularly shocking. What Axios has now reported, though, is on another level entirely.

An unnamed company apparently burned through $500 million on Claude in a single month. The reason? Nobody thought to set usage limits on the licences. Enterprise AI pricing can look attractively flat-rate on the surface, but those contracts almost always cap request volumes per model. Blow past the cap, or fail to read the fine print, and the bill becomes something your CFO will be discussing at length.

That story sits alongside a broader pattern of questionable AI deployment. One CTO noted that staff were routing weather queries through Claude. It technically works. It also costs orders of magnitude more than typing into a search bar. Sophia Velastegui, formerly an AI lead at Microsoft, made the observation to Axios that organisations tend to aim AI at the tasks nobody wants to do rather than the work that actually matters commercially. That tracks.

The real cost culprits here are misuse and poor model selection. Misuse often comes down to context engineering failures: conversations that sprawl endlessly with bloated context windows because nobody structured the prompts or workflows properly. Poor model selection is even simpler to diagnose. Routing every task through an expensive frontier model when a cheaper, lighter one would do the job fine is just waste dressed up as innovation.

Not every problem needs a large language model. Plenty of things still run better in conventional software, and knowing when to reach for AI versus when not to is fast becoming a core operational competency. That skill gap has consequences beyond the finance department too. Quality degrades when people don't know what they're doing. There are documented cases of Copilot running in auto mode, confidently producing badly biased analysis that a thinking model would have handled correctly. Bad outputs have their own downstream costs, often harder to quantify than the API bill but no less real.

The companies that will get this right are the ones building genuine internal expertise in how these systems behave, how to structure tasks, and which tools fit which problems. Roles focused on AI agent orchestration are going to matter. Chucking licences at a workforce and hoping for the best, apparently, can cost you half a billion dollars before the month is out.