How much does one AI agent actually cost?

One AI agent doing one job costs three things: the tools, the running, and the build. Standard plans run from free to about a hundred euros a month. Running it adds cents to a few euros per job. The build is the real cost: your time, or a professional's fee.

That is the short version, and the ranges are real. If you have searched this question you have met the whiplash: one page says an agent costs twenty euros a month, the next says fifty thousand. Neither is lying. They are pricing different things.

This page splits the cost into its three layers, puts a genuine range on each, and shows where every number comes from. It also says plainly when the right spend is nothing at all. Everything is scoped to the thing a small business actually buys first: one agent, doing one job. If the word agent is still fuzzy, start with the plain-English guide to AI agents and come back. This page is about the money.

Two notes before the numbers. Every price below was checked against the vendor's live pricing page in July 2026; prices move, so the linked pages are the source of truth. And most AI vendors bill in US dollars, so exact prices appear the way the vendor shows them, with euros kept for the rough shape.

On this page: What you pay for · The build · Cost drivers · When not to spend · In short · FAQ

What do you actually pay for with an AI agent?

An AI agent has three costs: the monthly plans for the tools it runs on, a running cost based on how much text it reads and writes, and the one-off cost of building it. That is the whole model. Every quote you will ever see is some mix of these three, and most quotes only price one of them.

Think of hiring a capable temp. There is the job ad: the build, written once and paid once. There is the day rate: the tools, a predictable monthly cost. And there is the tea they drink while working: the running, small and steady, scaling with how much work gets done.

In one line each: the tools are a monthly plan, the running is charged by how much the agent reads and writes, and the build is paid once, in time or in fees.

The monthly cost of an AI agent is the price of the tools it runs on, from free to about a hundred euros a month for one agent on standard small-business plans. The tools layer is the easy one. When I checked in July 2026, Claude had a free plan. Its Pro plan cost 17 US dollars a month if you pay for the year, or 20 month by month, and its Max plans started at 100 dollars. ChatGPT follows the same shape: a free plan, then paid tiers. Zapier, the workflow tool many agents lean on, also has a free plan, and its Professional plan started from 19.99 dollars a month. One agent doing one job typically needs one of these plans, sometimes two. That is the whole layer: free to about a hundred euros a month, with most first agents at the bottom of that range.

The running cost of an AI agent is what you pay for the reading and writing it does, usually cents to a few euros per completed job. The running layer is the one nobody explains. AI tools charge by text volume, so the bill follows the work. Here is the worked shape, assumptions in the open. A customer email is a few hundred words, and reading it plus drafting a reply passes perhaps a thousand words through the AI. On Anthropic's published per-use prices for its mid-range model, checked in July 2026, that costs about a cent. Thirty emails a day is around thirty cents. A heavy job that drags whole documents along at every step can climb to a few euros per run. So the honest range is cents to a few euros per completed job, and a paid monthly plan usually includes a decent amount of use before you pay anything extra.

How much does it cost to build an AI agent?

Building an AI agent costs time more than money: someone has to write down how the job is done, connect the tools, and test the result against that standard. The money side depends entirely on who does that work.

The cheapest build is yours. The first rung is pasting a written standard into a free chat and asking for a draft: zero euros, plus an afternoon of writing. When the job earns a proper build, the method is the same one I use for clients. It is all written up in how I build AI agents with Claude, step by step, and the money cost on this route is still mostly the plans above.

The middle rung is a no-code builder. Workflow tools such as Zapier or Make let you wire an agent into your other software on a standard plan, the sort already in the table. No developer needed, though the time cost grows with every system you connect.

The top rung is paying someone, and here the market gets loud. The cost guides ranking for this search quote wide ranges, almost none of them sourced, and every one written by someone selling the thing it prices. So, one honest paragraph on the choice. Doing it yourself costs time and teaches you your own process. Done-for-you costs a fee and should hand you a working, documented asset your team owns without the builder. Which is right depends on the job, and that comparison deserves its own page. Whatever route you take, ask any quote which of the three layers it covers. My own prices are published openly rather than quoted here, because prices belong on pricing pages; the link is at the end.

What the build actually is

Three things, none of them magic. The standard: the job written down the way you would brief a capable temp. The connections: giving the agent access to the systems the job touches, such as your inbox or your forms. The testing: running real examples through it until the output passes the standard. Knowing this is what kills the mystery that lets anyone quote anything.

Why are AI agents so expensive (and when they are not)?

Most of an agent's bill is not fixed. It follows a handful of choices, and you control nearly all of them.

What makes it dearer

Judgement mid-task is the big one. An agent re-reads its notes at every step, so a job with many steps reads far more text than a chat about the same job would. Long documents cost more, because they get dragged along each time. High volume costs more, for the obvious reason. Connecting real systems adds build time. And anything customer-facing needs human checking, which costs your hours even when the software is cheap.

What makes it cheaper

Fixed steps: if the job never needs judgement, a plain workflow does it for a fraction of the price. Small jobs, taken one at a time. Free tiers first, until the job proves it deserves a budget. Drafts only, with you on the send button. And writing the standard yourself, which moves the dearest layer from fee to time. One warning while we are here. Any page promising a precise return percentage on an agent it has never seen is not doing maths, it is doing theatre with a spreadsheet. Admire the costumes and move on.

This is also why drafting jobs sit at the cheap end. A cafe drafting review replies pays nothing for the platform, because a Google Business Profile is free. The whole bill is a plan and cents of running.

When is an AI agent not worth the money?

If a job runs the same fixed steps every time, plain automation does it cheaper than an AI agent, often on a plan you already pay for. That rule alone saves more money than any discount ever will. Do not spend when:

• The job runs on fixed steps. Invoice reminders are the classic: Xero and QuickBooks both include them, sent in your own wording on a schedule you set. Switching that setting on costs nothing new.
• The job happens once. Agents pay off through repetition. A one-off task belongs in a chat window, on the free tier.
• You cannot describe the job. If you cannot write down what a good result looks like, no build fee is worth paying yet. Write the standard first.
• The checking would cost more than the doing. Count the hours of human review before you count the savings.
• You have not tried the free tier. Pasting the job into a free chat is the cheapest market research there is.

One more plumbing note. A website enquiry that needs to become a contact and a notification is not an agent job at all: HubSpot's free tier does that with no AI and no spend.

Halfway note: I build one-job agents like this for a living, and unlike most of this market my prices are published openly on the work-with-me page. Or keep reading; the free route above is real, and it is where I would start too.

In short

• One agent, one job, three costs: the tools, the running, and the build.
• The tools run from free to about a hundred euros a month; the running adds cents to a few euros per completed job.
• The build is the real cost, paid once, in your time or in a professional's fee.
• If the job runs on fixed steps, buy automation instead, and check what you already pay for first.
• Start on the free tier and let the job prove it deserves a budget.

Common questions

References and further reading

This guide was written in July 2026, and the honest way to date it is to say so here. Every page below was checked and loading on 24 July 2026, and every price in the article traces to one of them. OpenAI's pricing page blocks automated checking, which is why ChatGPT appears above without exact numbers; its pricing page is easy to find and worth reading directly.

• Claude pricing, Anthropic. The free, Pro and Max plan prices quoted above.
• Anthropic per-use pricing, Anthropic. The per-use prices behind the running-cost example.
• Zapier pricing, Zapier. The free plan and the Professional plan price quoted above.
• Turn invoice reminders on or off, Xero. The built-in reminder setting named in the when-not-to-spend section. QuickBooks documents its own version in its help centre.

If you want to build it yourself, start with how I build AI agents with Claude, step by step and the free tiers above; that path costs you an afternoon. If you would rather have it built properly, one agent doing one job, on your process is exactly what I sell. A Scoping Sprint is the smaller first step if you are not sure which job to pick. Email me at ameera@asurekha.com and tell me the job. Plain words are fine.