What CFOs Are Finally Asking About AI Spend

Cursor's CFO council reveals the gap between AI pilot budgets and production reality. The conversation has shifted from 'what can it do' to 'what does it cost to keep running.'

The Pilot Trap Is Over

Cursor gathered CFOs from companies actually running AI in production. The consensus: pilot budgets are fiction. A $50k proof-of-concept becomes a $500k annual line item once you add infrastructure, observability, eval pipelines, and the engineers who maintain it. Nobody accounted for the cost of keeping the model working.

Token Economics Don’t Scale Linearly

Everyone models API costs at prototype volume. At production volume, three things break the model: context window bloat from RAG retrieval, retry loops when outputs fail validation, and the hidden cost of human review queues. One CFO reported their per-transaction cost 4x’d between pilot and month six.

The Headcount You Didn’t Budget

You’re not buying a tool. You’re hiring a team. Prompt engineers, eval specialists, infrastructure ops, compliance reviewers. The CFOs agreed: the first AI hire is expensive. The fifth is a new org chart. If your ROI model assumes existing staff absorbs the work, it’s wrong.

What Changes Now

Stop asking vendors for ROI calculators. Build your own: infrastructure + model calls + human review + engineering maintenance + compliance + opportunity cost. Run it for three months. Then decide if the use case survives contact with reality.