Somewhere, a company’s finance team is still processing what happened. According to a report by Axios, an unnamed enterprise gave its entire workforce unrestricted access to Anthropic’s Claude — no spending caps, no per-user limits, no real-time dashboards. One month later, the bill was $500 million.
Not per year. Not per quarter. One month.
This isn’t a fluke. Claude, like all enterprise AI tools, charges per token — every word in, every word out. A simple chat query is affordable. But an AI agent looping through dozens of research tasks, generating reports, calling external tools — that’s a hundred times more expensive per session. Multiply that across thousands of employees running agentic workflows with no kill switch, and the meter doesn’t just run — it sprints.
The real problem isn’t the bill
It’s that the controls were there all along. Claude’s admin console has per-user limits, monthly spending caps, and usage dashboards built in. They just aren’t on by default. Someone had to go configure them — and nobody did.
Uber burned through its entire 2026 AI budget by April from Claude Code alone. Microsoft pulled most of its Claude Code licenses after per-engineer costs hit $2,000 a month.
What builders should do right now
If you’re deploying Claude or any token-based AI tool inside a company — even a small one — set hard monthly caps per user before you roll it out. The cost of agentic workflows is not intuitive, and it scales faster than any other software expense you’ve seen. Treat it like cloud infrastructure: always cap first, monitor always, and never assume “it won’t get that bad.”