Why do AI bills surprise people?
Unit prices per token keep falling while bills keep rising, because agentic work consumes far more tokens per task than a single answer does, runs continuously rather than on request, and pulls in more internal data each quarter. A per-seat mental model cannot predict a per-token bill, which is why most organizations report overshooting their original AI cost projections.
Which AI FinOps levers pay off, and in what order?
Start with attribution: cost per team, per project, per model, because nothing can be optimized while the bill is one line. Then match the model to the work through model routing, cut repeated input with prompt caching, and retire agents that no longer earn their keep as part of a real agent lifecycle. Cost control is only half the job; the number that matters is cost against outcome, tracked per unit of work as cost per task, which is where AI ROI measurement and the original AI business case come back into it.
What does AI FinOps look like inside our own company?
Two habits carry most of it. First, attribution at the source: every model call our systems make is tagged with the job that caused it, so the bill decomposes into tasks instead of arriving as one line, the setup in our guide to token cost attribution. Second, budgets where the spending happens: each of our unattended agent runs carries its own cap and checks it before every paid call, so the worst case for a misbehaving run is its own budget, never the account balance. The attribution is also what finds the savings: seeing exactly where one coding agent’s tokens went is how we halved its usage without changing models.