How much does an AI agent cost?
A custom AI agent built by Ideasforge typically costs between €2,500 and €10,000 to build, plus €150 to €500 a month for us to operate and measure it. The model and cloud bills are separate, and they are yours. This page explains what moves the number, with real cost data from our systems in production.
The short answer
A single-job agent sits at the lower end of the range. One channel, one system to connect to and a clearly scoped task, like reading the invoices that arrive in a chat and turning each one into a row in your spreadsheet. Building it starts around €2,500, and running it around €150 a month.
The upper end belongs to agents that touch several systems and need deeper validation before going live, like an assistant that answers from your documentation and also queries live data. Those builds approach €10,000 and their operating fee sits at the top of the monthly range.
Larger multi-agent systems are quoted per project.
And there is a third figure worth being clear about from the start. The model the agent uses and the cloud it runs on sit in accounts under your company’s name, so those bills are yours and they are not inside the monthly fee.
Subscription or custom agent
Search for this price and you’ll see figures running from a few euros a month to tens of thousands. They don’t contradict each other. They describe two different purchases.
An AI subscription for business costs about €20 per person per month. That’s what Microsoft 365 Copilot, ChatGPT Business, Claude Team and Google Workspace with Gemini charge on their official pricing as of October 2026, between 14 and 25 euros or dollars depending on the plan and how you pay. It gives you an assistant that helps each person while they work. Whoever asked for the output reviews it.
For that kind of work it’s usually enough, and we won’t sell you anything else. Where your process is standard, a proven product beats a custom build on time and on price.
A custom agent is a different purchase. It’s a system that produces the result on its own, with your data and your permissions. It turns every incoming invoice into a row of data ready to use. Or it handles a customer over chat and books the appointment in your calendar. That’s where the subscription that looked cheap fills up with workarounds, because nobody is reviewing each output.
So when is a subscription not enough? One question decides it, whether someone will review every result or the system has to produce it alone. If it’s the second, this is how we approach it in our custom AI agent development.
What moves the price
Four things explain almost every quote we send.
- How many systems it connects to. An agent that only answers questions is cheaper than one that also writes to your calendar, your CRM or your database, because every connected system needs its own permissions and its own tests.
- The state of your data. If the knowledge the agent needs lives in clean, readable sources, the model performs better and the build gets shorter. We usually gain more by cleaning up data and tools than by rewriting prompts.
- How much proof you need before going live. Our appointments assistant Wazzy does not ship a change until a battery of conversations annotated one by one passes. Not every project needs that depth, and choosing it is part of the price conversation.
- Who operates it afterwards. The monthly fee covers watching the system in production, and the next section shows where that money actually goes.
The math on whether it pays off
Before asking for a quote, run a calculation that only needs two of your own numbers: how many minutes the task you want off your plate takes today and how many times it happens a month.
Here’s how it looks in two of our systems. At a property management company, each utility bill took a minute of typing, and today 98% go through without anyone touching them. At a rental agency, each WhatsApp inquiry took five to ten minutes of manual checking, and the team gets back more than three hours a day.
Multiply your two numbers and you have the hours that work eats up every month. Put them next to the price range above and the math either works or it doesn’t. If it doesn’t, that’s an answer too, because some processes don’t pay for their automation, and finding out before you build costs you one multiplication. To add what you lose each time the task goes wrong, our free tool does the math with one of your tasks.
The two costs of a running agent
A running agent has two costs, and they are worth keeping apart. One is what the system consumes to work, the model and the infrastructure, which sits in accounts under your company’s name. The other is our fee, which pays for the work of watching it.
Every message that comes in triggers calls to the model provider. That is the raw running cost of having the agent live, your company pays it, and it goes up and down with usage, so we do not bury it inside our fee. Infrastructure works the same way. The cloud account is in your name, so its bill is too.
Keeping it separate has an upside for you. You can see it, so you can bring it down. Knowing how the spend splits is what lets you cut it by measuring instead of guessing.
With that map in front of you, the decision comes from data. In one of our systems we tried a cheaper model and the test battery rejected it, because overall quality fell ten points. The saving was real. It just was not worth what it cost in quality.
Our fee pays for something else, which is the work of watching it. Every week we run a batch of anonymized test conversations against the live system from end to end. And before any change ships, the test battery has to pass. Two different things, kept separate on purpose. The battery gates changes, and the weekly batch watches what is already running.
What you own at the end
The repository is in your name from day one and the infrastructure runs in a cloud account that belongs to you, not to us. If we part ways, the system stays yours, with its documentation and its record of changes.
That also explains what the fee does not include. You are not renting the agent, so the monthly cost is operation, not a license that stops working when you stop paying.
We cover this in detail
- What happens after an AI chatbot or agent is deliveredHow we maintain an AI chatbot after delivery: a golden dataset of real conversations, monitoring, alerts and results every week.
- Measure AI by profit, not by productivityProductivity is AI's comfortable metric, it always looks good and commits to nothing. We prefer the uncomfortable one, which concrete gain the system delivered.
Cost questions we hear most
How much does an AI chatbot cost?
A support chatbot sits at the lower end of the range, starting at €2,500 to build, because it usually lives on one channel and draws on one knowledge source. The price climbs when it stops only answering and starts acting, booking appointments or updating records, because every action needs its own permissions and tests.
Why is there a monthly fee at all?
Because the model your agent runs on changes underneath it. Providers update models without changing their name, and a system that answered well yesterday can start failing quietly. The fee pays for the metering and the weekly test runs that catch it before your users do.
And the arithmetic works because the expensive part, building the battery of cases with their annotated answers, was already paid for during the build. Running it before every change and watching every week is machine work.
Can we run it without you afterwards?
Yes. Everything is yours, so you can take over whenever you want, and we run handover sessions when a client asks for them. Keep one thing in mind, though. Operating an agent means measuring it, and if nobody keeps measuring, you find out from a customer.
Want a number for your case?
Tell us what you’re up against and we’ll reply within one business day. If we don’t think it will pay for itself, we’ll tell you.