The AI cost exposure shift

The first wave of AI coding tools trained organisations to think in seats: a predictable monthly subscription, a familiar procurement motion, and a productivity story that mostly lived outside the finance model. That framing is changing quickly.

GitHub Copilot is moving to usage-based billing. Anthropic Enterprise separates access from token consumption. Frontier models are becoming more expensive even when they are more capable and, in some workflows, more token-efficient. The commercial signal is clear: the subsidy is becoming visible.

This creates a new management question. It is not enough to ask whether a team has AI tools. Leaders need to know whether people are using expensive intelligence well: when to escalate to a frontier model, when to use a lighter one, how to keep context efficient, and when an agentic loop is still useful rather than simply consuming spend.