Nobody asks the cost-per-task question until month four

An agent demo costs nothing. An agent in production costs a number nobody calculated, and the first time most teams see it is on a monthly invoice in month four.

The unit to model is cost per completed task, not cost per token or cost per call. A task that takes eleven calls and succeeds is cheaper than a task that takes three and gets escalated to a person.

Three places the spend runs away. First, retrieval that returns too much: every irrelevant chunk is paid for twice, once to embed and once to reason over. Second, retries without a budget — a loop with no ceiling will find one. Third, the escalation path, which is usually costed at zero and is actually the most expensive route in the system.

Put the figure on the same dashboard as the quality metric. A model that is ten percent better and four times more expensive is a decision, not an upgrade, and it should be made by someone who can see both numbers at once.

If your risk committee cannot read the cost model, it isn't finished.