What does an AI agent cost to build?
There is no single number, and the reason is that three different things get sold under one name. My published starting point for AI agents and automation is $499, which covers a workflow built, tested and handed over. Small jobs start at $49. A full agent with its own tools, permissions and limits sits above that, because the guardrails are the build.
What moves the number is not how clever the model is. It is how many systems the work touches, how many cases it has to handle, who has to approve what it does, and whether anyone but you will still be able to run it in a year.
- How many tools it has to reach: forms, sheets, a shop platform, an accounts package, a database.
- Whether a person approves anything a customer reads, and how that approval happens.
- What gets logged, so a bad run is something you can trace afterwards.
- How many real cases exist to test against before it goes live.
- Whether it runs on your server or someone else's, and who pays for that.
How do the three levels compare on cost?
Agentic automation pricing follows the level, not the buzzword. The same three levels I use on the automation page, priced: they are not a ladder of ambition, and the top one is not the best, it is just the most expensive to get right.
| The build | Build effort | What the running bill is made of | What makes it expensive | When it is worth it |
|---|---|---|---|---|
| Fixed workflow, no model | Lowest. The steps are known, so most of the work is connecting and testing. | The automation tool's plan, and the server it runs on. Nothing per run. | Every extra system it touches, and the edge cases nobody wrote down. | Any process you could already describe to a new starter. This is most of what a small business needs. |
| Workflow with one AI step | Low to mid. The route is simple; the work is the prompt, the failure path and the check. | The model call for each run, plus the tool plan and hosting. | Volume, because the bill scales with runs, and long documents, because they are expensive to send. | When the stages are settled but one of them needs a reading of the content: an enquiry, a contract, a photo of a form. |
| Agent that picks its own next step | Highest. Tools, permissions, retries, logs and a ceiling on cost all get built first. | Every model call it makes on the way, including the calls a fixed path would never have made. | Loops and long chains, where one job quietly spends the budget of fifty. | When nobody can draw the flowchart in advance, and the damage of a wasted minute is a wasted minute. |
How do you work out the running cost?
Count three things and multiply them: what one run sends to the model, how many runs happen in a month, and the rate the provider publishes for that direction. Nothing else belongs in the figure, and any estimate that skips your volume is a guess dressed as a number.
An example with invented inputs, not a client result. A run that sends 4,000 tokens in and gets 600 tokens back, done 2,000 times a month, is 8 million tokens in and 1.2 million tokens out. Multiply each column by the current published rate and add the two, and that is the month, before anyone's time. Change the volume and the number moves with it, which is the point: the model bill is a usage bill.
- Tokens in per run: the document, the prompt and whatever history travels with it.
- Tokens out per run: the answer, which is usually far smaller than what went in.
- Runs per month, including the retries and the runs nobody asked for.
- The published rate for each of those two columns, from the provider's own pricing page, which is why I link it at the bottom of this page rather than quote a figure that ages.
How do you know whether it pays back?
There is one comparison to make. The AI automation cost, meaning the build plus a year of running, against the loaded hourly cost of the person who does it by hand. Loaded means the rate with taxes, tools and overhead on top, because the cheap hour is never the real one.
- Example inputs, not a client result: a task of 20 minutes, done 20 days a month, is roughly seven hours a month.
- Seven hours times that person's loaded hourly rate, times twelve, is what the task costs you in a year of doing it by hand.
- Add a year of running cost to the build price. If that total is clearly below the number above, the payback is real and the decision is easy.
- If the two are close, buy the simpler level: a fixed workflow with one AI step usually captures most of the saving at a fraction of the risk.
- If the job happens three times a month, do not build anything. Two hours of someone's time is cheaper than any software, and pretending otherwise wastes your money.
When should you not build one?
When the volume is small. An automation that runs four times a month has to be almost free to justify itself, and nothing I build is that cheap.
When the process is not agreed yet. Automating a fight just gives the argument a server, and the rebuild is what you will pay for.
When the step is irreversible and unsupervised: refunds, payroll, contract sending, stock movements. Those need a person, and the honest design is a draft plus an approval, not an agent.
When a ready-made tool already does it. Most small businesses can buy this in a subscription they already have, and I will tell you which one.
And when the reason is the demo. An agent looks impressive in a meeting and costs money every day after, so if the driver is being seen to use AI, that is a different conversation and a cheaper one.
Is this the same question as an AI app build?
No, and the two guides are deliberately separate. This one is about software that runs your processes in the background, on a schedule or a trigger. If the AI is inside a product your customers use, with screens, accounts and a store listing, that is what AI app development costs, the starting price is much higher, and the running bill follows your users rather than your jobs.
And if the real decision is who to hire rather than what to build, the four-way comparison is in agentic developer vs agency.
What to remember
- Two bills: a fixed build price and a running cost that follows your volume.
- Pick the cheapest level that finishes the job, which is usually a workflow with one AI step.
- Running cost is tokens per run, times runs per month, times the published rate, counted in both directions.
- Logs, past cases, hosting, the tool's plan and the person approving drafts are real costs and belong on the quote.
- Compare hours saved against build plus a year of running, at the loaded hourly rate.
- Low volume, an unsettled process or an irreversible step each mean: do not build an agent.
Sources
Primary documentation, for anyone who wants to check the details.