The price of an AI chatbot for a website has four parts. These are the setup work, the monthly price of the platform, the use of the AI model and, when the chatbot also answers in WhatsApp messages, the fee Meta charges. The model itself is the cheapest part, since an ordinary conversation from a website costs from a fraction of a cent to a few cents. Ready platforms still charge about half a dollar to a dollar per conversation, because most of the price is the tool, the knowledge base and support. The best models understand Serbian almost as well as English, while smaller and cheaper models make noticeably more mistakes in Serbian.
In 2025, roughly one in ten enterprises in Serbia with at least ten employees used some kind of artificial intelligence, half as many as in the European Union. Many businesses are therefore deciding on a chatbot for the first time, and offers differ both in price and in what is actually being charged.
What is the price of an AI chatbot made of?
The first part is the setup work. A chatbot answers from a knowledge base, made up of the price list, opening hours, service descriptions, delivery terms and the questions customers actually ask. All of that needs to be put in order. Rules of behavior are also written, for example when the bot offers a conversation with a staff member and what it never promises. Then everything is checked on real questions. This work is measured in days. Taking the average pay in computer programming published by the Statistical Office of the Republic of Serbia, one day of such work is worth about €163.
The phrase “trained on the data of the business” in offers almost always means exactly such a knowledge base. The model itself is not trained further for the chatbot of a small business, and OpenAI states that it is shutting down its platform for further training of models. A separate item for “training in Serbian” therefore deserves the extra question of what exactly it includes.
The second part is the platform, the program the chatbot runs in. It provides the chat window on the website, a place for the knowledge base, a review of conversations and handover to a staff member. It is paid monthly, per conversation or per resolved question. The third part is the AI model that writes the answers. The fourth part exists only on some channels, and the best known example is WhatsApp messages, for which Meta charges a separate fee.
How much does the AI model itself cost per conversation?
Models are paid per token. A token is a piece of text, most often part of a word, and makers such as OpenAI and Anthropic state the price per million tokens. For a business owner that price says little on its own, so it is more useful to look at what a whole conversation costs.
One thing is worth knowing first. A model does not remember a conversation by itself, so with every new question it again receives the instructions, passages from the knowledge base and everything said so far. OpenAI explicitly states that those earlier parts of the conversation are charged again. That is why each next answer in the same conversation costs a little more than the one before.
For this text the price of a typical conversation was calculated, with six questions and as many answers in Serbian Latin script. A thousand such conversations cost about €3.4 with the cheapest models. With a midrange model, which the makers describe as a balance of capability and price, they cost about €69. Models that “think” before answering use more.
The platforms themselves see the same order of magnitude. Botpress counts about ten cents of model use into each conversation, and the rest of the price goes to the tool around the model.
“All previous input tokens for responses in the chain are billed as input tokens in the API.”
How do ready chatbot platforms charge?
Platforms charge in three ways, which makes offers hard to compare. The first is paying per resolved question. Fin, the AI agent from Intercom, charges about a dollar per outcome. An outcome counts when the customer confirms the problem is solved or no longer asks for help after the answer. If Fin does not know the answer or the customer asks for a person, nothing is charged. In a similar way, Zendesk also charges only for requests the AI resolves without a staff member.
The second way is paying per conversation. Botpress asks $150 a month for 250 conversations in the Plus plan, and each further conversation is paid separately. What matters here is what counts as a conversation. Every exchange in which the visitor sends at least two messages counts, and the same person who comes back several times can be counted several times. When the monthly quota is almost used up, new conversations are bought automatically.
The third way is message credits. Chatbase works that way, and its smallest paid plan costs $40 a month.
When everything is reduced to one conversation, ready platforms cost between half a dollar and a dollar. The model itself costs from a fraction of a cent to a few cents for the same conversation. The difference is the chat window, the knowledge base, the conversation review, handover to a staff member and support. Before choosing, it is worth asking what exactly counts as a conversation, what happens when the monthly quota runs out and how a resolved question is counted.
How much does WhatsApp add to the price?
A chatbot in WhatsApp messages has one more cost. Meta charges for messages the business sends, while messages from customers are free. Answers the business or its chatbot sends within 24 hours of the last message from the customer are called service messages. They were free for a while, but Meta charges for them again from October 1, 2026. Each business phone number gets 1,000 free service messages a month, and for Serbia each one after that costs less than two cents. After that period the business can write only with messages approved in advance, which are also charged.
At first glance these are small amounts, but they add up quickly. A chatbot that handles a thousand conversations a month, with five answers in each, pays Meta about €70 a month. That is more than the same conversations would cost with the cheapest model. Meta changes prices at the start of a quarter, so the WhatsApp bill is worth checking then.
How well do AI models understand Serbian?
Model makers do not publish results for Serbian. Anthropic, for example, publishes results for several languages, but Serbian is not among them. It advises checking every language on its own examples and setting the answer language explicitly in the instructions. Independent tests therefore say the most.
On the Global PIQA test, with everyday questions written by native speakers, the best models solved almost all Serbian tasks, in both Latin and Cyrillic. The test is still small and only measures whether a model understands the question, not how well it answers in Serbian.
Differences show on harder questions. On a test with exam questions translated into several languages, the strongest models were almost as good in Serbian as in English, and smaller models noticeably weaker. The models in that test are older than today’s, but the direction is clear. In Serbian, smaller and cheaper models fall further behind than large ones, and they are exactly the ones most attractive for a chatbot because of their price.
No study with a published method was found on mistakes in grammatical case and gender or on mixing Serbian and Croatian words. That is why this is checked by trying, on real customer questions.
“Test with any languages relevant to your specific use cases.”
Why does the same conversation cost more in Serbian than in English?
The tokenizer, the part of the model that splits text into tokens, was built mostly on English text. One study showed that the same content in some languages has up to 15 times more tokens than in English. For speakers of those languages that means a bigger bill, slower answers and less room for the knowledge base in a single question.
Serbian is not among the worst cases, but the gap exists. For this text the same set of sentences was measured in English and in Serbian, with the tokenizer used by OpenAI models from 2024 and 2025. Serbian Latin had about one and a half times as many tokens as English, and Cyrillic even more. The sentence “Dobar dan, koliko košta dostava do Beograda?”, which means “Hello, how much is delivery to Belgrade?”, has 14 tokens in Latin and 16 in Cyrillic.
Newer models were not measured, and the number of tokens changes even between versions from the same maker. Anthropic, for example, states that its new tokenizer produces about a third more tokens for the same text. For a rough estimate it is enough to count Serbian as using about one and a half times as many tokens as English, as was done in the calculation above.
“The same text translated into different languages can have drastically different tokenization lengths, with differences up to 15 times in some cases.”
How can a chatbot be checked before paying, and what if it does not know the answer?
The most useful check takes about ten minutes, on a trial version with the real data of the business. The same question is worth asking in Cyrillic, in Latin without diacritics, for example “koliko kosta dostava do Cacka”, and in another grammatical case. Then the answers are compared. Next comes a question the knowledge base has no answer to, where a good chatbot says it does not know and offers contact instead of making up an answer. Finally a question about price, checking whether the number matches the price list, and a request to talk to a person.
Handover to a person is no small detail. In a Gartner survey almost two thirds of customers said they would prefer companies not to use AI in customer service. The biggest concern was that it would be harder to reach a person. Gartner therefore advises that a chatbot should clearly say it will pass the conversation to a staff member when it does not know the solution.
The business is responsible for what its chatbot says. In 2024, in the dispute Moffatt v. Air Canada, a tribunal in Canada rejected the defense that the chatbot was a separate entity. The reasoning was simple. The company is responsible for all information on its website, whether it comes from a regular page or from a chatbot.
There are rules too. From August 2026 the EU Artificial Intelligence Act requires that users know they are talking to AI, at the latest at the start of the conversation. That can also apply to a business outside the European Union whose chatbot answers customers in the Union. In Serbia the Law on Personal Data Protection requires that conversations in which a customer leaves a name, phone number or email be kept only as long as necessary. The customer must also know how long the data is kept and whether it goes to another country. This matters because a chatbot conversation usually passes through the servers of the model maker.
All in all, the AI model itself is the smallest part of the bill. Most of the price is the setup, the platform and, for WhatsApp messages, the fee to Meta, and in Serbian stronger models make noticeably fewer mistakes than cheaper ones. How a chatbot with a knowledge base in Serbian is set up is described on the AI chatbot and virtual assistant page.