A finance leader does not need to understand token sampling to get to the bottom of a high AI bill. The diagnosis is a sequence of questions, and the answers point to the fix. This is the checklist to work through with your engineering team when the OpenAI bill lands higher than plan.
1. Can we attribute this bill to teams and use cases?
If the answer is no, stop here, because this is the root problem. A bill you cannot break down by owner is a bill nobody manages. The first ask is one API key per team or use case so that next month's bill arrives already decomposed. Attribution is the foundation everything else stands on.
2. Did call volume grow, or did cost per call grow?
These are two different problems. Volume growth points to more usage, retries, or new automations. Cost-per-call growth points to heavier context or more expensive model choices. Ask engineering to split the change into these two, because the fix depends entirely on which one moved.
3. Is our spend growing faster than our usage?
If AI spend is outpacing the growth in whatever the AI produces, users served, tasks completed, revenue supported, that gap is where efficiency has slipped. Spend growing in line with value is healthy. Spend growing faster than value is the definition of a problem, and this ratio is the one to watch every month.
4. Are we running the right model for each task?
Ask what share of spend is on the most expensive model and how much of that work genuinely requires it. Routine tasks on a frontier model are the most common large waste on an AI bill, and reallocating them to a cheaper model is often the fastest big saving available.
5. Do we know about every workload that is running?
Ask for an inventory of every AI workload against the bill. Anything on the bill that is not on the inventory is unmanaged spend, and it frequently hides forgotten prototypes and abandoned jobs. The gap between the inventory and the bill is a number worth asking for directly.
6. What is our cost per outcome, not per call?
Push for a unit-economics number: what does one completed task, resolved ticket, or served customer cost in AI terms. This is the metric that connects AI spend to the business and lets you judge whether the spend is worth it. A bill without a cost-per-outcome number is a bill you can only shrink, not optimize.
7. Do we find out about spikes in days or at month end?
If the first time anyone sees a number is on the invoice, every problem is a surprise. Ask for weekly spend visibility and anomaly alerts so that a spike announces itself while it is still small and cheap to fix.
Work through these seven and the bill stops being a mystery. Every high AI bill resolves into some combination of missing attribution, volume or rate growth, model misallocation, unmanaged workloads, and no cost-per-outcome measure. The checklist does not require code. It requires asking the right questions in the right order.