Annual budgets set in January don't survive contact with token-based AI billing. Microsoft reportedly canceled internal licenses for Anthropic's Claude after usage-based costs blew through full-year allocations within months - a pattern that's becoming an uncomfortable reality for enterprises moving from flat-rate seat licenses to pay-per-token models.
Token-based billing charges per unit of text processed (a "token" is roughly four characters, so a 1,000-word document is about 750 tokens in and whatever the model generates out). It's more efficient than per-seat licensing when usage is light, but it scales directly with how much employees actually use the tools. When a company deploys AI writing, coding, or summarization tools widely, token costs compound fast in ways that a traditional software budget - built around fixed annual license fees - simply doesn't accommodate.
The Budget Math Nobody Did
The structural problem is that enterprise procurement cycles are annual. Finance signs off on a line item in Q4, the tool deploys in Q1, and by Q2 a heavily-used AI tool has consumed twelve months of budget. That's not misuse - that's the product working as intended.
Microsoft's position here is awkward. The company has its own Copilot suite built on OpenAI models and has invested heavily in that stack. Maintaining internal Anthropic licenses was always somewhat redundant strategically. The budget overrun may have given procurement the justification to cut licenses it was already looking for a reason to cut.
For teams evaluating enterprise AI contracts, this is a real operational risk worth pricing in. Token-based costs need usage forecasts, not just budget approvals. The alternative - negotiating a token commitment or hybrid flat-rate deal upfront - is increasingly what large buyers are asking for.