Guides What an AI agent costs to build and to run

Budgets for AI agents tend to start with the model price per token, because it is the number that is easiest to find. It is usually not the number that decides the project. The build cost sits mostly in integration and measurement, and the running cost is something you can measure precisely before launch instead of guessing.

Where the build effort goes

• Integration: connecting the agent to the systems it acts in — authentication, data access, error handling and the tools themselves. This is often the largest share.
• Evaluation: building a test set from real inputs and a way to score it, so the team can tell whether each change helps.
• Permissions and approvals: deciding what the agent may do alone, what needs a person, and enforcing it.
• Observability: traces, cost tracking and alerts for the live system.
• The prompts and orchestration logic, which are usually smaller than teams expect.

A single, well-defined workflow — one job, a handful of tools, inside an existing product — is typically a matter of weeks to a first production release. Broad assistants that try to do everything take longer and are harder to evaluate, which is why starting narrow is usually cheaper overall.

What drives the running cost

Running cost is roughly the number of runs, multiplied by the model calls per run, multiplied by the tokens per call, at the chosen model's price — plus any retrieval, tool and hosting costs. The variables that move it most are:

• How much context each call carries. Long documents and long conversation histories are the usual cause of runaway cost.
• How many steps a run takes. Agents that loop or retry without limits are expensive and slow.
• Which model handles which step. Simple classification or routing steps rarely need the most capable model.
• How often the same work is repeated. Caching stable context reduces cost where providers support it.

Measure it before launch

Because the evaluation suite runs real inputs through the agent, it also measures cost per run. Before launch you can know the median and worst-case cost of a run, multiply by expected volume, and set a ceiling per run and per day that stops the agent rather than surprising you on the invoice. Track cost per feature and per customer afterwards, so that a change that doubles cost is noticed in days, not at the end of the month.

How to keep the first project affordable

Pick one workflow with clear value and a clear definition of success. Build the evaluation set first. Connect to real systems early. Put a person in the loop for anything with real consequences. A narrow agent that measurably works is a better investment than a broad one nobody can vouch for — and it is the foundation the next workflow builds on.