Budgets are the new estimates
There is currently no cost estimation tool for using AI at scale / agents. However, departments are already burning tokens, validating if this spend is justified is nearly impossible.
The more the path of agents is adopted, and proprietary models are used, the harder it becomes to predict and control costs. Systems will become tightly integrated with a specific vendor.
The constraint use to be availability of experts and the hours spent. Now the constraint is tokens, and having a clear goal, getting there takes retries, context windows, and the moment you are halfway through a task with no tokens left. That is not a billing annoyance. That is an unfinished system.
What actually helps:
- a budget per repo, per agent, per feature — with a hard stop, not a surprise invoice
- cheaper models for search and boilerplate, expensive ones only when the task needs them
- alerts before the cliff, not after
- a written fallback: who finishes the work if the agent cannot
If you cannot answer “what did this feature cost in tokens, and did we get what we paid for?”, you are not managing cost. You are hoping.
Estimation is dead. Budgeting is alive. Put a number on the run before you start it.
One of the few calculators available An AI Cost calculator like this would be nice