How to read an AI spend bill.

Most teams see their AI spend as a single monthly total and a vague sense that it went up. That total is made of a small number of decisions, most of which are invisible unless you go looking for them. Here is what actually drives the number, in the order we usually find worth checking first.

Repeated context, not raw usage

The single biggest hidden driver is not how many requests a workflow makes — it's how much of the same context gets resent on every single one of them. A support workflow that reloads a full policy document, a long conversation history, or a big tool-definition list on every turn pays for that payload every single time, even when nothing in it changed since the last call. This is the difference between "useful context" and "repeated or low-value payload," and it is usually the fastest place to cut cost without touching output quality at all.

Model choice per task, not per workflow

A workflow that routes every step — a simple lookup and a genuinely hard judgment call alike — through the same large model is paying premium-model prices for lookup-tier work. The fix is not "use a cheaper model everywhere," which trades cost for quality where it matters; it's matching each task to the simplest reliable model that can actually do it, and reserving the expensive model for the steps that need it.

Retries and tool-call loops

A workflow that calls a tool, gets an ambiguous or malformed result, and retries automatically can silently multiply its own cost several times over for a single user request. These loops are easy to miss in a monthly total because they don't show up as a separate line item — they show up as a slightly higher average cost per request that nobody traces back to its cause.

Approval boundaries that got automated by accident

Cost and safety intersect here: a workflow that used to pause for human approval before a payment, a send, or a security decision, and now runs on model confidence alone, is usually running more total calls, not fewer — because removing the human checkpoint often means adding compensating verification calls instead. Tracking where automation quietly replaced an approval step is as much a cost question as a safety one.

Where to look first

If you can only check one thing this week: trace one real, representative run of your most expensive workflow end to end — every prompt, every piece of context, every model call, every retry — and ask honestly which pieces of context actually changed the output and which were just carried along out of habit. That single trace usually surfaces more savings than a month of staring at the aggregate bill.