What an AI agent actually costs to build and run

An agent is not a chatbot with extra steps

An agent plans, calls a tool, reads the result, plans again, calls another tool and checks its own work. Every loop resends the context, and it is the input tokens you keep resending that dominate the bill. An unoptimised agent can cost ten to a hundred dollars a session. A properly engineered one doing the same job comes in under a dollar. Same model. The gap is engineering.

• Prompting and orchestration: about 15% of the build. The part everyone talks about.

• Integration with existing systems: closer to 35%. ERP APIs were never meant to be called four thousand times a month.

• Data cleaning and the evaluation set: another 25%. Tedious, and the asset that keeps the agent alive.

What it costs to run

One well-scoped agent at hundreds to low thousands of tasks a day typically costs between 500 and 3,000 dollars a month including model tokens. Context size, model tiering, prompt caching and a hard cap on loops are what move that number.

If a vendor quotes a build where prompting is 80% of the effort, they have built a demo and are hoping you will not notice.

If a vendor quotes a build where prompting is 80% of the effort, they have built a demo and are hoping you will not notice.

If a vendor quotes a build where prompting is 80% of the effort, they have built a demo and are hoping you will not notice.

The cost that never appears on an invoice

Somebody has to own the exception queue. Budget a fifth to a third of one operations person per production agent for the first six months. It drops after that. It never reaches zero, and pretending it will is one of the quieter ways projects get cancelled.

Three phases, not one

An expensive build. A messy first quarter where the agent costs more than it saves. Then a long tail where it is cheap and nobody thinks about it. Companies that budget for phase three alone tend to cancel in phase two.

Tell us the process, the monthly volume, how many systems it touches and what a mistake costs. From those four we can give a build estimate, a run rate and a straight answer on whether it is worth doing.