← Back to Feed

Post from 9 November 2026

When an AI agent costs too much to run, the obvious fix is a cheaper model. The model’s usage charges are the cost you can see. But the people checking the agent’s work can cost much more than the model does.

Say you run an agent (an AI system that does a task start to finish on its own) that processes supplier invoices. The model provider charges you by usage, roughly by how much text the model reads and writes. So a cheaper model looks like an easy saving.

But the agent can finish an invoice and still get it wrong. It reads the wrong amount, picks the wrong supplier, or tries to post the same payment twice. Someone has to catch that, redo the invoice and fix whatever already went through. That work goes up with how often the agent is wrong and how many invoices it handles, and comparing usage charges between models doesn’t see any of it.

McKinsey put numbers on this in 2026, for one bank’s customer-service workflow. Of the running costs that grow with each request handled, the model’s usage charges were about 20-25%. The people checking and correcting its work were about 70-75%. That’s one bank’s workflow, not a rule for every system.

In your system the usage charges may really be the bigger cost. That can happen when the agent rarely gets things wrong and little of its work needs checking. To find out, I’d count how many invoices a person had to check, redo or fix, and what that time cost, then put it next to the usage charges.

A pricier model that gets it right the first time can come out cheaper than a bargain one that keeps needing fixes. You’ll only see that if your system records whether each invoice was actually processed correctly, not just whether the agent finished running.

So before you switch to a cheaper model, work out what it costs to get one invoice done correctly, retries and human checking included.

Also on Bluesky · Mastodon

Loading comments…