How much does an AI agent cost? The 7 factors that drive the price

What an AI agent costs is not decided by the AI itself but by seven factors: how many process steps it takes over, how many programs it connects to, how good your data is, how independently it may act, which model it uses, where it is hosted and who looks after it afterwards. On top of that comes the split between one-off development costs and running costs for the model, hosting and maintenance. The biggest lever is in your own hands: start with one clearly defined process and you will pay considerably less than someone who wants to automate the whole office in one go.

Let me be blunt: the most expensive line in an AI agent project is rarely the language model. It is almost always the connection to your existing software and cleaning up the data the agent needs. If you know that up front, you can save in the right places.

Which 7 factors drive the price of an AI agent?

An AI agent is a program that completes a task in several steps on its own: it reads a customer enquiry, for example, pulls data from your software, decides what to do and prepares a result. Each of these building blocks takes development time. If the difference from a simple chatbot is still a bit fuzzy, read Chatbot or AI agent? first.

1. Number and complexity of process steps

An agent that only sorts incoming emails into “enquiry”, “invoice” and “other” is quick to build. One that reads an enquiry, asks the customer for missing details, calculates materials, writes a quote and follows up a few days later has five steps – and each of them has to be built and tested separately. Rule of thumb: the more decisions the agent has to make along the way, the more test cases it takes until it runs reliably.

2. Number of system integrations

This is where trade businesses usually see the biggest effort. Typical candidates are your trade software, CRM, email inbox, calendar and accounting. If a program has a well-documented interface (API), connecting it is manageable. If not, it gets laborious: you need workarounds such as file exports or an inbox acting as a relay. So ask your software vendor up front whether there is an API – it is one of the cheapest bits of groundwork you can do.

3. State and quality of your data

An agent can only calculate as well as your price list is up to date. If material and labour rates sit in a maintained list, it can use them directly. If they are spread across old Excel files, PDFs and the head of the master craftsman, someone has to sort them out first. That is not AI work, but it is part of the project and often underestimated.

4. Level of autonomy and approval steps

May the agent send quotes on its own, or does it only prepare a draft that you approve with one click? This approval step is called “human in the loop”. It sounds like extra work, but at the start it is usually the cheaper option: an agent that only delivers drafts does not have to handle every edge case perfectly, because a person checks the result. Full autonomy, by contrast, requires plenty of safety rules, exception handling and a clean log of every action.

5. Choice of model and running API costs

The AI model is billed by the amount of text processed. A large model such as Anthropic’s Claude API is strong at understanding long, unstructured enquiries and at writing clean text. For simple tasks like sorting emails, a smaller and cheaper model, for example from Mistral, is often enough. Mistral is a French provider, which some businesses value for data protection reasons. A good agent does not use the most expensive model for every step but the right one.

6. Hosting and data protection requirements

Customer data, addresses and quote totals are personal or business-critical data. If you want them processed in the EU, you need suitable hosting. We often use Supabase with servers in the EU as the database and n8n for the workflows. n8n can run on your own server, so the data does not pass through third-party platforms. The stricter your requirements – your own server instead of a cloud service, for instance – the higher the setup and operating effort.

7. Maintenance and further development

An agent is never entirely “finished”. Your trade software gets an update, a model provider replaces an old model with a new one, your prices change. So plan from the start who will look after the agent. Whether that is a maintenance contract or work on demand has a big influence on running costs.

Which costs are one-off and which are ongoing?

Many quotes only mention development. For your own sums you need both sides:

One-off costs	Running costs
Process analysis: which workflow, which steps, which exceptions?	Model costs per processed case
Developing the agent logic and prompts	Hosting for servers, database and n8n
Connecting software, inbox, calendar	Maintenance when connected software is updated
Preparing data, such as price list and templates	Adjustments when your process changes
Testing with real example cases and onboarding	Monitoring: is everything running, where does it get stuck?

Running model costs depend directly on volume. A business with a handful of enquiries a week will barely notice them; with several hundred cases a day they become a real item. That is why it pays to roughly count how often the process actually happens before the project starts. How to do that properly is covered in the guide on process analysis before automating.

What does an AI agent look like in a trade business?

Take the process that eats up the most office time in many businesses: writing quotes. That is exactly what our AI-powered quote creation for trade businesses was built for. The system analyses enquiries arriving by email, WhatsApp or web form, identifies the type of job, its scope and materials, calculates based on current price lists, generates a PDF and automatically follows up after 48 hours.

What drives the price in this example maps directly onto the seven factors:

• Process steps: read the enquiry, calculate, create the PDF, follow up. A manageable, clearly defined workflow.
• Integrations: inbox and web form are easy. Connecting the trade software for customer records and prices depends heavily on the vendor.
• Data: a maintained price list is a must. Without one, clean-up work gets added.
• Autonomy: the master craftsman approves before anything is sent. That keeps the first version lean.
• Maintenance: prices change regularly, so you need a simple way to keep the list current.

The trick is to start small. Begin with quote creation only, let the agent deliver drafts instead of sending them itself, and add follow-ups or appointment booking only once the first part works in everyday use. That way you only pay for what has already proven itself. You will find more use cases on our page on AI for trade businesses and under AI agents for trade businesses.

When is an AI agent not worth it?

An AI agent is not a cure-all. In these cases I advise against it or suggest something simpler:

• The process rarely happens. If you write two quotes a month, you save too little time for development and maintenance to pay off.
• Fixed rules are enough. If a workflow is always the same (“form comes in, customer is created in the CRM, confirmation email goes out”), a classic workflow automation with n8n is cheaper and more stable. I explain the difference in workflow vs. AI automation.
• Every job is one of a kind. For highly individual renovations where nothing works without a site visit, an agent can at most pre-sort the enquiry.
• The data is missing. No price list, no calculation. Then step one is getting organised, not buying AI.
• The task is sensitive. Where a wrong answer is costly or legally risky, a person has to decide. The agent can prepare, but not take responsibility.

One more point that belongs in your planning: if your agent talks to customers directly, for example via chat or by sending emails on its own, the transparency obligation under Art. 50 of the EU AI Act has applied since 2 August 2026. Customers must be able to tell that they are dealing with an AI. What that means in practice is covered in the guide on chatbot labelling obligations. And if you are mainly looking for a chatbot for your website, have a look at what an AI chatbot costs.

My take

Don’t start by asking “How much does an AI agent cost?” Ask “Which single process is costing me the most time right now?” Write down how often it happens each week, how long it takes and which programs hold the data it needs. With those three answers the effort can be estimated realistically, and you will quickly see whether an agent makes sense or a simple automation will do.

That is exactly what we do in the free AI potential check (30 min): we look at your workflow together and tell you honestly whether an agent makes sense and at what scope. What we build in the area of AI agents is on our service page. Or get in touch directly for an initial call.

Note: The remarks on data protection and the EU AI Act are a technical assessment, not legal advice.