What does an AI agent actually cost to build?
Nobody publishes real numbers, so here is what actually drives the price of an AI agent — where budgets get spent, where they get wasted, and how to scope one so it does not run away from you.
Every agency answers this with “it depends”, which is true and useless. The honest version is that the price is driven by a handful of specific things, and once you know what they are you can estimate your own project reasonably well — and spot a quote that has been padded.
There are two costs, and people usually only think about the first one.
Cost one: building it
This is a software project, and it prices like a software project. What moves the number:
How many systems it has to touch. An agent that reads your inbox and drafts replies is one integration. An agent that reads the inbox, checks your CRM, looks up stock, and raises an invoice is four — each with its own authentication, its own rate limits, and its own way of being unavailable on a Tuesday. This is usually the single biggest cost driver, and it is the one people underestimate most.
How messy the inputs are. Structured data from an API is cheap. PDFs from forty different suppliers, each with their own layout, is not. If a human currently has to interpret something rather than just move it, expect that to be the expensive part.
What happens when it is wrong. An agent that drafts a message for someone to approve needs far less engineering than one that sends money. Consequence drives the amount of validation, logging, approval gating and testing required, and those can easily double the build.
How much it has to decide. “Extract these five fields” is a task. “Work out what this customer wants and handle it” is a system. The second needs planning loops, retries, and a way to give up gracefully — meaningfully more work.
Cost two: running it
This is the one that surprises people, because traditional software does not work this way. Every time your agent thinks, you pay for tokens.
The rough shape: cost scales with how much text goes in and out, multiplied by how many steps the agent takes, multiplied by how often it runs. An agent that reads a long document on every request is dramatically more expensive than one that retrieves only the relevant paragraphs first.
Two things reduce this a lot, and any competent build should be doing both. Retrieval means fetching the three relevant paragraphs instead of stuffing an entire handbook into every request. Model routing means using a small, cheap model for the easy majority of steps and a large one only where the reasoning genuinely needs it.
Ask for a projected monthly running cost before you sign anything. If nobody can give you one, they have not thought about it.
Where budgets get wasted
Building an agent for a job that is not agentic. If the process is the same five steps every time, you want a script, not an agent. It will be cheaper, faster, and it will not surprise you. We turn work down on this basis regularly.
Automating a broken process. Automation makes an existing process faster, including a bad one. Fix the process on paper first.
Chasing the last 5%. Getting an agent to handle 80% of cases is often straightforward. The final stretch of rare edge cases can cost more than everything before it. Frequently the right answer is to handle the common cases well and route the rest to a person.
How to scope one sensibly
Pick one process. Write down how many times a week it happens and roughly how long it takes a person. That number is your budget ceiling — if the automation cannot pay for itself against it in a sensible period, build something else first.
Then insist on a phased build. A narrow version that handles the most common case end to end, in production, beats a comprehensive version that is still in development. You will learn things in week one that change what you want, and it is much cheaper to learn them against something small.
What we will tell you
We do not publish price lists, because a number written before we understand your process would be wrong in both directions. What we will do on a first call is tell you which of the drivers above apply to your project, roughly where the effort sits, and whether the thing you are describing needs an agent at all.
Sometimes the honest answer is that a well-placed script and a tidy database will solve it for a fraction of the cost. We would rather say that early than build you something impressive that you did not need.