The token bill is an identity problem

In late 2025, Uber implemented Claude Code for 5,000 engineers, quickly exhausting its annual AI budget by April 2026, reflecting a broader trend of uncontrolled AI spending due to lacking governance infrastructures. Traditional enterprise software costs are predictable, but token consumption in AI is notably different, marked by non-linearity, invisibility, and non-attribution, which complicates financial oversight. The emerging FinOps response emphasizes real-time monitoring, business-unit chargeback, and ROI thresholds, contingent on establishing a robust authorization architecture. This involves distinct agent identities, tool-level scoping, and session boundaries to facilitate accurate cost attribution. The Model Context Protocol has yet to resolve these issues at the protocol level, but the FinOps Foundation's initiatives and tools like WorkOS are developing frameworks to manage these challenges effectively, underscoring the necessity for early architectural decisions to prevent budget crises.