Less than a cent. That is what this website’s AI assistant spends on each conversation. The hard part is building it the right way and adapting it to the legal requirements that apply.
The last three conversations with the AI assistant on this website cost €0.0034. They were test conversations, with simulated clients that behave like real ones: a dental clinic asking about a website with online booking, someone wanting to know what maintenance includes and a company asking for a price on a management system. The assistant answered all three with the published rates, took their contact details and prepared a summary of each enquiry.
That figure contradicts a widespread idea, that artificial intelligence is expensive. A business that puts an assistant on its website pays only for the text the model reads and writes, and a sales enquiry, in those terms, usually consumes very little text. That said, the real cost of a chatbot lies in building it so that it does not get things wrong, and in keeping it within the limits the law sets when AI talks to your customers.
Why it costs so little
The language models the AI assistant uses are billed by tokens. Tokens are fragments of words, and they have one price for reading and another for writing. Besides the messages users type into the assistant, there is also a set of rules and context information attached to the conversations so that the AI model knows how to answer properly, and that context is usually very long, which makes the number of tokens sent to the model grow considerably.
In this website’s assistant, that context is the business’s instructions and rates, and it takes up about 6,300 tokens. Each call reads between 6,300 and 7,500 tokens to write answers of around 200 tokens on average. And there are more calls than messages, because before answering the assistant looks up the rates: the three test conversations added up to twelve messages from the visitor and, to answer them correctly, needed seventeen calls in all.
With that usage, the total cost depends on two decisions taken when building it.
The first is the cache. Providers charge far less for text they already read in the previous call, and since the instructions and rates are identical in every one, almost the whole conversation benefits from it. In the test, more than 93% of what the model read in the second and third conversations came out of the cache, at a price thirty times lower. Without it, those two conversations would have cost about six times more.
The second is the tier of the model. The one that answers on this website costs about €0.12 per million tokens read, and a high-end one, around €3.56. To answer a sales enquiry with the rates at hand there is no need to turn to a high-end model. The system works perfectly well with a budget model, since no deep reasoning is needed to hold a useful, fluent conversation with the visitor.
Below is an interactive calculator so you can see live how the cost of the calls to an AI model varies.
Conversations per month
Messages per conversation
Model
Cache
Estimated monthly cost
€0.55
€0.0011 per conversation
Input at €0.12 and output at €0.37 per million tokens
Usage measured on 5 October 2026 on this website's assistant: 6,300 fixed tokens per call, about 200 in the answer and 1.4 calls per visitor message. The budget tier uses the price its provider billed that day; the other two, list prices on 14 October 2026. Providers bill in dollars, converted here at €0.89 per dollar.
With the budget model and the cache switched on, five hundred conversations a month cost €0.55. With a high-end model and no cache, €100. The difference is enormous, and worth knowing.
Where the real cost of the system lies
An assistant like this one is built in one or two one-week sprints, from €1,190 to €2,380. Connecting the model to the system is solved quickly. The time goes mostly into making sure it always answers the customer well and correctly, even when the customer asks about things out of context or that have nothing to do with the purpose it was built for.
First come the instructions, which set how the assistant introduces itself, what it may talk about and what it must leave to a person. Second, the data. The rates and conditions reach it from the same source that feeds the website, so it cannot give a figure the page does not show. Third come the limits, which cap how many messages a conversation allows and how long each answer may run. That way the maximum spend is a known figure. As an extra safeguard, the assistant is given a limited budget that acts as a circuit breaker, so that if it runs out, the assistant stops and the page offers the form instead.
And finally there is what takes the longest, the testing. A language model answers just as fluently when it is right as when it is wrong, and the only way to know how often it fails is to measure it. Before going live, this assistant went through a test bench of three hundred conversations with simulated clients, which checks every figure against the rates, that no discounts are offered, that it does not invent services and that the summary arrives complete. With every change of model or instructions, the test bench is run again until the final result meets the agreed acceptance criteria. Reaching the optimal performance of this website’s AI agent took some 18,000 simulated conversations and about 19 uninterrupted hours of testing and improving the instructions.
The legal obligations and how they shape the system
The AI assistant exchanges information with customers and therefore receives personal data in every conversation, such as the name, email and telephone number of the enquirer and often the situation of their business. All that information is confidential and has to be handled in keeping with the law, so there are five important matters to bear in mind:
Information and transparency. The page introduces the assistant as an AI agent, and the privacy policy explains what it does, which providers are involved and on what legal basis.
Using European models. The provider that runs the model processes data on behalf of the business, and the GDPR requires a data processing agreement with it. There are European providers that include it in their general terms and process the data within the European Union, so they settle both matters at once. If providers that handle data outside the European Union are used, it is worth looking into it to line up with the legal requirements of the client’s country.
Knowing where the data travels. Most models run outside the European Union, and sending data there requires a safeguard. Here only providers that commit to not retaining the content of the conversations, nor training their models on it, are involved.
Data deletion. Conversations delete themselves after 90 days.
Leaving the decision to a person. The assistant summarises the enquiry. The assessment and the reply are sent by me.
To those obligations, since 2 August 2026, the European Artificial Intelligence Act adds one more: whoever talks to an AI system has to know it, unless it is obvious. It is enough for the assistant to introduce itself as what it is.
That said, it is essential to take all of these aspects into account before building an AI assistant, since, depending on the region, the legislation and the client’s requirements, the models and providers will have to be chosen in advance from among those that comply with everything.
What is set out in this article is general information and is no substitute for the advice of a lawyer, which is worth seeking when the assistant is going to handle health or financial data.
What to ask whoever builds it for you
With the real order of the costs in front of you, the questions for whoever is going to build you an AI assistant change. The monthly fee usually has little bearing on the price. What really matters is knowing where the assistant gets its answers from, what limits it has, how it has been tested, which providers see the conversations and when they are deleted. These are questions that should be answered in writing and before starting.
The infrastructure is the cheap part. The truly costly part is making sure it does not get things wrong.
Frequently asked questions
How much does an AI chatbot cost for a business?
Building a customer service assistant costs from €1,190 to €2,380, excluding VAT, which corresponds to one or two one-week sprints. That price is paid once. After that comes the model's usage, which depends on the number of conversations.
How much does an AI chatbot cost per month?
The one on this website spends €0.0011 per conversation, measured on 5 October 2026. With 500 conversations a month, the usage comes to around €0.55 using a budget model with the cache switched on.
Does an AI chatbot comply with the GDPR?
It can, and it depends on how it is built. You have to inform whoever writes to it, look at how the chosen provider handles the data, cover any transfers of data outside the European Union, set a deletion period and work with providers that do not train their models on the conversations.
Do you have to disclose that the chatbot is an AI?
Yes. Since 2 August 2026 the European Artificial Intelligence Act requires people to know they are talking to an AI system, unless it is obvious from the context.
Can the model provider use the conversations for training?
It depends on the provider’s terms, so it is worth having it in writing. I work only with providers that commit to not training on the data and to not retaining the content of the conversations.
Can the same chatbot run on the website and on WhatsApp?
Yes. The assistant is the same; only the channel through which messages are sent and received changes. Adding WhatsApp to a website assistant usually takes up the second sprint.
