What an AI agent actually costs to build, and to run

Search for what an AI agent costs and you get ranges too wide to plan with: low five figures for a "basic" agent, six figures for an "autonomous" one, mostly from firms that sell the build. The range is wide because "AI agent" covers everything from a chatbot with one tool to a system that reads, decides and acts across five of your apps. Price the job, not the label.

Five things that move the build price

• How many systems it touches. Each integration (your CRM, WhatsApp, Stripe, a database) needs sign-in, error handling and tests. Integrations are usually the biggest line, not the model.
• Whether it reads or writes. An agent that answers questions from your docs costs less than one that changes orders, issues refunds or books calls, because every write needs permissions, confirmations and a way to undo it.
• How much judgement the steps need. If a decision fits in rules, that part should be ordinary code, which is cheaper to build and cheaper to run.
• What your data looks like. Clean, structured records connect quickly. PDFs, scanned forms and three spreadsheets that disagree need extraction and cleaning first.
• How right it has to be. An internal helper can be wrong now and then. An agent that talks to customers or touches money needs a test set of real cases, guardrails and a hand-over to a person.

The costs that arrive after launch

Most quotes stop at the build. An agent then costs money every month, and the bill grows with use:

• Model usage, billed per token. It rises with every conversation and every long document the agent reads.
• Hosting, logs and monitoring.
• Third-party fees: phone minutes, WhatsApp conversations, search and data APIs.
• Someone reading transcripts and fixing what the agent got wrong, most of all in the first month.

Ask every quote for two numbers: the build price, and the monthly running cost at your expected volume. If the second one is missing, so is half the budget.

How to keep the number small

• Start with one workflow, not a platform. One agent that does one job well beats a framework for ten jobs nobody uses.
• Put the rules in code and the judgement in the model. It is cheaper per run and far easier to test.
• Pilot on a slice of real traffic before turning it on for everything.
• Keep the accounts in your name: code, model keys and logs. Then changing vendors later is a handover, not a rebuild.

How we price it

We scope agent work after a free audit of the workflow: where the work comes in, which decisions get made and which systems the agent has to touch. You get a fixed price in writing for the build, an estimate of the monthly running cost at your volume, and the list of what we cut. If the workflow is not ready for an agent, the audit says so, and it is yours to keep either way.