Total AI spend is the number everyone tracks and the number that explains the least. It tells you what happened without telling you why, and it cannot tell you what happens next. The metrics that actually matter sit underneath the total, and tracking them changes AI cost from something you react to into something you can manage.
Cost per call, trended over time
Divide total spend by total calls to get average cost per call, then watch it move week over week. A rising cost per call means your calls are getting heavier, more input tokens, larger context, or a shift to more expensive models. A flat cost per call with a rising bill means you are simply making more calls. This one metric splits every cost investigation in half before you start.
Calls per task
A single business task often triggers many model calls behind it. Tracking calls per task against the baseline a task should take reveals retry amplification, which is one of the most common causes of runaway spend. When observed calls run far above the expected baseline, retries or a broken loop are usually the cause, and you can catch it before it reaches the invoice.
Input-to-output token ratio
You pay for tokens you send as well as tokens you receive, and on many features input dominates. Tracking the ratio of input to output tokens per feature shows where context is quietly driving cost. A feature where input is many times output is a candidate for prompt compression, context trimming, or history truncation.
Spend by team and use case
Attribution is a metric, not just a report. Tracking spend broken down by team and use case is what lets you tell legitimate growth from noise. A team whose spend rises for a clear reason is fine. A jump you cannot attribute to anything is the signal worth chasing.
Cost per outcome
The most important metric is the one that connects spend to the business: cost per completed task, per resolved ticket, per served customer. Token costs are inputs. Outcomes are outputs. Tracking the relationship between them is what tells you whether spend is generating proportionate value, and it is the number a board actually wants to see.
Spend versus usage growth
Finally, track whether spend is growing faster than the thing your AI produces. Spend rising in line with usage is healthy. Spend rising faster than usage is where efficiency has slipped, and it is the earliest sign that a feature needs attention. That single comparison is worth more than any absolute dollar figure on the bill.