In 30 seconds

The headline price of an AI agent is rarely the real cost. The usage fee is often the smallest line. The bigger costs are the human time to set it up, the review that keeps it safe, and the work created when it gets something wrong. Add those together before you decide whether an agent is worth it.

The four cost drivers

1. Usage fees

Most AI runs on a pay-per-use model, priced by the amount of text going in and out (measured in tokens). For a single short task this is often a fraction of a cent. It only becomes significant at high volume, or when each task feeds the model a lot of context.

2. Setup and integration

Someone has to connect the agent to your systems, write and test its instructions, and prepare the knowledge it reads. For a small workflow this might be a few days. For anything touching several systems it is the largest cost in year one, whether you build it or pay a vendor.

3. Human review and oversight

A useful agent still needs a person checking its work, at least early on. That review time is a running cost. It falls as trust grows, but it never reaches zero for anything consequential.

4. The cost of being wrong

When an agent makes a mistake, someone spends time finding it, fixing it, and reassuring whoever was affected. A cheap agent that creates expensive errors is not cheap. This is why the cost of errors belongs in the sum from the start.

A simple way to hold it in your head

Why the usage fee is usually the wrong thing to worry about

New teams tend to fixate on the per-use price because it is the number the vendor shows. But for most SME workflows the volume is modest, so usage stays small. The setup and the review are where the money and the risk actually sit. Judge those first.

What this does not solve

Costing does not tell you whether the agent will work, only what it will take to run. Pair this with an honest look at whether the workflow is even a good fit, and estimate the numbers with how to estimate the cost of an AI workflow.

Good enough to move on when

  • You can name a rough number for each of the four drivers, not just the usage fee.
  • You know the expected volume, because cost scales with it.
  • You have compared that total against the value of the time the agent saves.