People expect the model bill to be the big number. It rarely is. Here is what a typical production agent costs us to run each month, for a client processing around 10,000 documents.

The lines

Model calls: usually the smallest line, often a few hundred dollars. Document reading and storage: similar. Hosting, queues and logs: a little more. Human review of sampled and flagged cases: by far the largest, and the one that makes the rest trustworthy.

Cost per task

We report one number to clients: all-in cost per task. It includes the review time. For document-heavy work it usually lands between twenty cents and a dollar, against several dollars for the manual process.

Where it goes wrong

Costs creep when agents retry blindly, read whole documents when they need one page, or call the largest model for every step. We design for the cheap path first and escalate only when the eval set says we must.