Consumption pricing is efficient when usage is predictable and valuable. It is painful when usage jumps, pricing changes, or a workflow quietly multiplies calls to a model.
Why AI consumption spikes
- A new workflow or agent goes live and calls the model many times per task.
- Users switch to a more expensive model tier.
- Longer documents or conversations mean more tokens per request.
- A vendor changes its pricing unit or what is included.
- A loop or error repeats calls until someone notices.
A real pricing surprise
In June 2025 Cursor, an AI coding tool, moved its Pro plan from a request-based limit to usage-based pricing. On July 4, 2025 it apologized, saying the changes "were not communicated clearly", and offered to refund unexpected charges from the previous three weeks. Its plan now includes a spend-limit option to control overage. Finance tools are not exempt from the same dynamics.
Controls that work
- Caps and alerts in every vendor console, set below the budget.
- Spike alerts internally: flag month-over-month growth above a threshold (for example 25%).
- Run-rate forecasts: project the year end from the last few months, not from the annual budget.
- Unit economics: cost per invoice, per reconciliation or per report, compared with the value per unit.
- Model routing: use cheaper models for simple tasks; reserve expensive ones for hard tasks.
- Contract terms: price-change notice, caps, and the right to exit.
Using the spend tracker
Enter monthly actuals per vendor and team. The tracker forecasts the year end from the last N months, raises overspend and spike alerts at your thresholds, and shows unused seats.