An AI agent is not priced by the model behind it - it is priced by how many of your rules it has to know before you can stop watching it. When people search for what an AI agent costs, they come back empty handed, because the first five results are agencies with a contact form and a line about individual pricing. I understand why they do that. What I do not understand is why nobody breaks the number down, when it can be explained in plain language in a few paragraphs.
So here is what I say in the first conversation, before any figure is named: what the price is made of, where the cost sits after launch, when I talk clients out of the project, and how to cut the scope down to a single month so the decision rests on your own data rather than on somebody’s slide deck.
One distinction first, because it decides the order of magnitude. A chatbot answers questions: text in, text out, and when it does not know something it either invents an answer or hands you over to a person. An agent goes further - it reads a request, checks the data in your system, creates a record, changes a status, sends an email, and hands the case to a human when it is unsure. A chatbot talks. An agent acts, and that is why it is priced differently: every action it is allowed to take is a separate thing to build, fence off and test.
What an AI agent costs for a company - what the price is made of
A quote breaks into five parts, and only one of them is about artificial intelligence itself.
The language model. The smallest item, and usually the only one people ask about. You are not buying a program - you pay the provider for usage, billed by the amount of text processed. For one process in one company this is the item that disappears next to the human work around it. It only grows when the agent has to chew through an entire catalogue or years of correspondence, which is why I always ask up front how much material has to pass through it.
Access to your data. This is where the real work starts. An agent that is supposed to answer a question about an order has to be able to find that order. That means a connection to what you already run: the store, the warehouse system, the mailbox, the spreadsheet. The cost does not come from the connection itself, but from what happens when the other side does not answer, or answers with an error.
Rules and the limits of its authority. The most expensive part, and the one most often missing from quotes. What the agent may do on its own and what it has to hand to a person for approval. What it must never touch. How it behaves when it is not certain. Without that layer you get a tool that sometimes does something brilliantly and sometimes writes nonsense into your system in a confident tone, and somebody has to unpick it afterwards.
Testing on your real cases. An agent checked against three invented examples has not been checked. You take archived requests, run them through the agent and compare the result with what the human did at the time. That takes hours, and it is the one part of the quote where cutting corners never pays.
A panel for the human. A place where you can see what the agent did, undo it, and look at what it based the decision on. Without that, nobody in your company can be made responsible for the agent’s work.
In practice a narrow scope - one process, one data source, a human approving the result - is a fixed-bid project for me, with a written quote after a conversation and billing in stages. A wider scope, with several integrations and the agent acting without approval, is quoted from the feature list, item by item, and I always put that list in writing before anything starts. The same logic applies to any custom system, not only to agents - I laid it out in more detail in the piece on what a custom web application costs.
The costs that show up after launch
Rolling out AI in a company does not end on the day the agent starts working. There are three recurring items.
The first is usage at the model provider, which grows with the number of cases handled. It is the only cost that scales directly with the benefit, so it is the least troublesome one.
The second is keeping the connections alive. A provider changes an interface version, something stops talking to something else, and it has to be fixed in hours rather than weeks. On my side that sits inside ongoing care, and plans and prices are on the support page; I match the plan to whether the system only has to keep running or also has to grow.
The third is the work on your side that nobody can buy for you: reviewing what the agent did in the first weeks and writing down rules for the exceptions that only surface then. Skip it and the agent stays frozen at launch day while your processes move on.
When it does not pay off
There are four situations where I talk clients out of an agent, and I would rather say it before the quote than after handover.
When the process is not written down anywhere. If everyone in the company does the task slightly differently, the agent has nothing to learn from. Agree the rules between people first, automate second - otherwise you are paying to freeze the mess into code.
When the task has exactly one correct answer. Assigning an order to a courier by weight and postcode is a plain condition, not a job for a language model. Ordinary automation is cheaper to build, cheaper to maintain, and never makes a creative mistake. I drew that line in the piece on WooCommerce automation with Zapier and n8n.
When a mistake is expensive and nobody will be supervising. Issuing documents, changing prices, writing to a public authority. The agent can prepare a version for approval, but somebody has to look at it. If there is nobody to do that, this is not the project for now.
When the volume is small. A handful of cases a week will not pay back an implementation. A simple test I use: work out how many hours a month the task eats today and multiply that by your own rate. If the result sits far below the implementation bracket, you already have your answer, and you are better off looking for something off the shelf. Often a plugin is enough, and which of them actually save time is something I went through in the piece on AI plugins for WordPress.
A one month pilot instead of a one year project
The worst way to buy an AI agent is to pick everything at once off the back of a demo. The best way is to carve out a single process and check it against your own data before any larger money is committed.
This is how I stage it:
- One task, the most repetitive one. Not the hardest, not the most impressive - the most repetitive. Usually that means sorting incoming requests, drafting the first reply to a standard question, or copying data out of emails into a system.
- Read only at the start, no writing. The agent prepares a proposal and a person approves it with one click. Write access comes later, once the numbers show the proposals are good.
- One data source. The system holding most of the answers. The rest can wait.
- A measure agreed before the start. How many cases the agent handled without correction, and how much time that took off a person. Without that number you will be arguing about impressions a month later.
- A decision after a month. Extend it, point it at a different process, or shut it down. Shutting it down is not a failure - it is a cheap answer to a question that would otherwise have cost you a year.
A pilot like that is quoted the way any narrow scope is, and on my side it goes fixed-bid, with the item list in writing after a conversation. If you have a process that looks like a candidate, describe it through the contact form - write out what a person does today, step by step, and I will send back a pilot scope with a price and a date. If I conclude that plain automation at a fraction of that cost would do the job, I will say so.
One sentence about data, because the question comes up in every conversation: before an agent gets access to orders or correspondence, you have to settle what exactly leaves for the model provider and what stays with you. That is decided while the scope is being written, not after the rollout.
Frequently asked questions
What does an AI agent cost for a small company? For one process, one data source and a human approving the output, I quote that scope fixed-bid, in writing after a conversation, billed in stages; on top come provider usage and technical care. The figure grows with the number of systems and permissions the agent has, not with how clever it sounds.
Is a chatbot cheaper than an agent? Usually yes, because a chatbot only answers and changes nothing inside your systems, so there is no layer of permissions or undo to build. The thing worth checking is whether it solves your problem at all. If people still have to retype the data into a system after talking to the bot, the saving is imaginary.
Would an off the shelf subscription tool not be better? Wherever your process is typical, off the shelf wins, and I tell clients that openly. A custom build starts paying off when the agent has to know rules no ready made tool lets you configure, or reach into a system nothing off the shelf connects to.
How do I check that the agent is not making things up? On archived cases, before launch, by comparing its output with the decision a person made. After launch, through a panel with history and the right to undo every change. A model with no access to your data will always invent answers, because that is its nature - supervision means giving it something real to answer with and making its answer verifiable.
Will this improve how visible my company is in AI answers? That is a separate matter. An agent handles your processes internally, while whether ChatGPT knows about your company depends on the content on your site. How to check that is something I wrote up in the piece on the AI visibility audit.
Building an application, a panel or an integration rather than a website? I write B2B systems to order - backends, APIs, client panels, billing modules. Let’s talk about your project.