Are retries increasing your AI bill?
Retries can move your AI bill even when customer activity, completed workflows and model choice stay almost unchanged. Culpa keeps repeated calls attached to the workflow and conversation that caused them, so you can see when you’re paying several times for one piece of customer work.
🔁 The economics are simple. A retry doesn’t erase the first attempt. If two attempts fail and the third succeeds, your customer receives one result while you may have paid for three model calls.
That is why total calls per day isn’t enough. You need to compare model calls with the amount of customer work completed.
Here's an illustrative example. If 10,000 completed workflows normally require six model calls each, you make 60,000 calls. If the same 10,000 workflows suddenly require eight calls each, you make 80,000 calls for the same number of completed jobs.
That's 33% more model calls before customer volume changes.
🧠 “Retry” can also mean quite a few more things than your retry counter shows. Timeouts, queue restarts, fallbacks, reviewer rejections, regenerations and agent repair cycles can all repeat model work.
The expensive part also often gets repeated before the failure. An agent may plan, retrieve context, analyse the task and call a tool before the workflow breaks and then if the whole sequence restarts, you're buying that upstream work again.
Context can make the second pass even more expensive. Previous attempts, tool results and reviewer feedback may all come back into the next model call.
So I’d watch calls per completed workflow, repeated-call cost and the workflow steps that trigger the repetition. The goal isn’t zero retries but rather, understanding what you’re paying to repeat and whether that repetition is useful.
See this working on your own data with our cost leak scan:
🔓 Culpa stays on your infrastructure so your prompts and responses stay there, and It also fails open, so your application keeps running if the cost layer fails.
How many calls does one normal customer action create in your product today?
And how many of them happened because something ran twice?
Replies