AI Chatbot Implementation in Your Business: Cost, Stages and Limitations
What makes up the cost of an AI chatbot and what a single conversation costs, the 7 stages of implementation, the limitations of language models and your obligations under GDPR and the AI Act. Based on the chat that runs on our website and books calls.
In short
- An AI chatbot makes sense when customers ask the same questions several times a week, including after hours, and the answers are in your own materials: your offer, price list, terms and conditions. When every question is different, a form or a booking calendar will serve you better.
- The cost has 4 parts: the build (one-time), integrations, the AI model fee based on usage, and the time it takes to keep the knowledge up to date. At UNSEEN, an AI module with a chatbot costs from 1,650 EUR net, integrations from 1,400 EUR net, and the first module is up and running after 2 weeks.
- The model itself costs less than you might think: with a small model, a typical conversation costs about 0.005 USD (calculation below). What costs more is a chatbot that makes up a price or a date.
- Rules that must not fail (booking only a free slot, no invitations where you do not want them) belong in the code, not just in the instructions for the model. The model usually follows its instructions, but not always.
- From August 2, 2026, a chatbot must tell people that they are talking to AI (Art. 50(1) of the AI Act). On top of that comes GDPR: a privacy policy, a data processing agreement with the model provider and a retention period for conversations that someone actually enforces.
- A few inquiries a week, each one different
- An FAQ page and a booking calendar, no chatbot
- Recurring questions about your offer and prices, also after hours
- A chatbot using company knowledge
- The customer should book a slot or leave a request right away
- A chatbot connected to a calendar and CRM
All the examples in this article come from the code of the chat that runs on our website and of Ania, our AI voice assistant. We describe the problem a chatbot solves most often on a separate page: customer questions after hours go unanswered.
How much does AI chatbot implementation cost?
Companies sell very different things under the name “chatbot”: from a chat window that answers based on a few documents to an assistant that checks available slots and saves requests in a CRM. That is why you should compare offers by their cost components, not by the name of the package.
| Component | What it covers | How you pay | At UNSEEN |
|---|---|---|---|
| Build | Instructions for the model, knowledge base, chat window on the website, rules and tests | One-time | AI module (UN·AI) from 1,650 EUR net |
| Integrations | Calendar, CRM, email, SMS, ordering system | One-time, per connection | Integrations (UN·Connect) from 1,400 EUR net |
| AI model | Requests to the language model, counted in tokens | By usage, to the model provider | By usage (calculation in the next section) |
| Maintenance | Updating the knowledge, reviewing conversations, fixes, server | Your team’s time or ongoing support from the vendor | Optional support (UN·Care) from 700 EUR net per month |
The cheapest packages are usually a chat that answers based on a few documents. The price goes up when the chatbot has to do something: check an available slot, save a request, hand a case over to a person. Each such action is an integration and a set of rules that have to be tested. We describe the scope and modules on our AI implementation for businesses page.
How much does a single chatbot conversation cost?
The model provider charges by tokens, that is, pieces of text. You pay for what goes into the model (instructions, knowledge about the company, the conversation history) and for what the model writes back. With every customer question, the model receives the content it needs all over again, so the cost grows with the length of the instructions and of the conversation.
Here is the calculation for gpt-4o-mini, the model our chat uses. According to OpenAI’s price list, it costs 0.15 USD per million input tokens and 0.60 USD per million output tokens (as of November 13, 2026). Let’s assume a conversation with 6 questions, in which each request sends about 4,000 tokens to the model and each answer is about 300 tokens long:
- input: 6 × 4,000 = 24,000 tokens, or 0.0036 USD;
- output: 6 × 300 = 1,800 tokens, or about 0.0011 USD;
- in total about 0.005 USD per conversation, or about 5 USD per 1,000 conversations.
These are assumptions, not measurements. A larger model or a long knowledge base sent with every question can multiply the cost several times over, or even more than tenfold. Ask the vendor which model they will use and how much text goes to the model with a single question. Voice conversations add speech recognition and speech synthesis, and phone calls also add a phone number and call minutes from the telecom operator.
How does a chatbot implementation work? 7 stages
We deliver the first working module after 2 weeks. Stages 2, 3 and 6 take the most work, and stages 3 and 7 are the ones most often skipped.
- Goal and boundaries. Write down the 10-20 questions customers ask most often and what a conversation should end with: a booked slot, contact details left behind, a referral to a form. Also write down what the chatbot does not do: it does not quote prices outside the price list, does not give discounts, does not promise delivery dates, does not handle complaints and does not ask for data it does not need.
- Knowledge base. Your offer, price list, terms and conditions, business hours, service area and the most common questions from emails. The model answers well only on the basis of what it is given, and an outdated price list in the knowledge base means outdated prices in the answers. Appoint a person who updates the knowledge whenever the offer changes. The technique in which the model searches documents for relevant passages before answering is described in the entry RAG in our IT glossary.
- Rules in the instructions and in the code. The instructions (the prompt) tell the model how to behave. Anything that must not happen should also be safeguarded in the program, because the model sometimes strays from its instructions. Examples from our chat are in the next section.
- Actions and integrations. If the chatbot is to check an available slot or save a request, it gets tools: the model asks for data, and the program looks it up in the calendar or CRM and sends back the result. Dates, statuses and prices then come from the system, not from the model’s memory.
- GDPR and AI disclosure. A notice that it is AI in the first message, a description of the chat in the privacy policy, a data processing agreement and a retention period for conversations (details below).
- Testing before launch. Prepare a set of 30-50 questions: your customers’ most common questions and the hard cases, such as a price outside the price list, a request for a discount, a request to book a slot for yesterday, a question in a foreign language, a request to ignore the instructions. For each one, write down what the correct answer is. Check that wherever the chatbot does not know the answer, it says so plainly and points the way to a person. Run the same set after every change to the knowledge base or the instructions.
- Reviewing conversations after launch. For the first few weeks, read the conversations. They will show you questions that were not on the list and answers that need fixing. Count not only the conversations but also how many of them end with a booked slot or contact details.
What does the chatbot on our website do, and what does it deliberately not do?
Our chat runs on the homepage and on the AI systems pages (the button in the bottom right corner). Below are its rules, taken from the code rather than from a marketing description, together with where each one is set. “Instructions” means the model enforces the rule; “code” means the program does, regardless of what the model writes.
| Rule | Where it is set | How it works |
|---|---|---|
| Says it is AI | Code | The first message: “I’m UNSEEN’s AI assistant”, and the chat window shows the label “Powered by AI”. |
| Books a free 30-minute call at a specific time | Instructions and code | The model fetches available times from the same calendar as the contact page. Just before booking, the program checks whether the time is still free, so a slot made up by the model cannot be booked. |
| Books only after confirmation | Instructions and code | The model repeats the day, the time, the type of call and the contact details, and books only after a clear “yes”. The program allows one booking per conversation and at most 3 attempts per hour from one IP address. |
| Does not lose the customer when the calendar does not respond | Code | The request is saved, and the team calls back to confirm the time. |
| Books calls only in Polish or English | Instructions | The chat replies in the customer’s language, but when someone writes in another language, it tells them that the call will be held in Polish or English. |
| Does not invite people to our office | Instructions and code | In-person meetings only in Warsaw; by default, the call is online or by phone. If the model still invites someone to the office, the program removes that sentence from the answer (patterns in 7 languages). |
| Does not convert currencies | Instructions | On the versions of the site with prices in euros, the model receives a ready-made table of prices in EUR, because when it converts on its own, it gets the rounding wrong. |
| Knows today’s date | Code | Every request starts with the current date and time in Warsaw time. The model does not know it on its own, and without it, it cannot book “tomorrow at 10”. |
| Passes contact details on to the team | Code | When the chat asks for contact details, a form appears below the conversation (name, email, what you need). The request reaches us by email and SMS, and after the window is closed, we receive a transcript of the whole conversation by email, so we know what the customer asked about. |
| Keeps answers short | Code | An answer is at most 500 tokens long, and the model sees the last 20 messages of the conversation. |
| Deletes old conversations | Code | A daily task on the server deletes conversations with no new message for 12 months, as the privacy policy promises. |
Ania, our AI voice assistant, books calls in the same calendar. Before you turn on the microphone, you see a notice that you are talking to an AI assistant. Clicking means you agree to talk to AI, and without that consent the server will not open the conversation. Ania does not record audio; only a text transcript is kept. When the assistant is unavailable, or the daily conversation limit for one address has been used up, you are back to the text chat. A description with screenshots: voicebot and AI receptionist.
What are the limitations of an AI chatbot?
A language model writes fluently and confidently, even when it is wrong. Most of its limitations follow from that.
- It makes things up. When it does not know the answer, it can give a plausible-sounding price, date or feature that does not exist (these are called hallucinations). What helps: a knowledge base, instructions to say “I don’t know” and refer the customer to a person, and tools that provide facts from the system. You will not eliminate this entirely, so do not let the chatbot make binding offers.
- It only knows what you give it. It does not know your price list, stock or calendar until it is given them, and the knowledge in the base ages along with your offer.
- It does not always follow instructions. A long conversation, an unusual question or a request like “ignore your previous instructions” can make the model do something against its instructions. That is why we keep critical rules in the code.
- It is bad at math. Do currency conversions, discounts and totals in the program, and give the model the result.
- It has a short memory. It sees only part of the conversation (in our case, the last 20 messages) and does not remember earlier conversations with the same customer unless the system provides them.
- It depends on the provider. When the model’s service does not respond or limits the number of requests, the chatbot goes silent. Our chat then shows a message in the language of the page, and the link to talk to a person stays below it.
- It cannot replace a person in disputes. The chatbot should quickly hand complaints, negotiations and conversations with an upset customer over to the team.
AI chatbots, GDPR and the AI Act: what do you need to ensure?
- Disclosure that it is AI. From August 2, 2026, an AI system intended to interact with people must inform them that they are talking to AI, unless this is obvious from the circumstances (Art. 50(1) of the AI Act), in a clear way and at the latest at the first interaction (Art. 50(5)). The obligation lies with the provider of the system, which also includes a company that commissions a chatbot and runs it under its own name (Art. 3(3)). The July 2026 amendment to the AI Act did not postpone this deadline. The simplest way: in the first message and in the chat window.
- Privacy policy. What data the chat collects, for what purpose, who processes it (the model provider, hosting, the database) and how long you keep it. Our policy states 12 months for chatbot conversations.
- Data processing agreement (Art. 28 GDPR) with the vendor, listing the model providers as sub-processors. If a provider transfers data outside the European Economic Area (e.g. to the US), you need a legal basis for such a transfer, e.g. standard contractual clauses.
- A retention period enforced by the program. A clause in the policy achieves nothing if nobody deletes the data. In our case, a daily task on the server does it.
- Minimum data. The chat does not need to ask for a national ID number (PESEL in Poland) or about health. Ask for contact details only when the customer wants to book a slot or leave a request, and say why they are needed. Our chat then adds a link to the privacy policy.
- Data at the model provider. Your retention period and the provider’s retention period are two different things. Check in the provider’s terms whether data sent through the API is used to train models and how long it is kept. OpenAI states that it does not use API data for training (unless the customer opts in), and that it keeps abuse monitoring logs for up to 30 days.
Before implementation, it is worth going through the list of questions from the Polish data protection authority (UODO), “Before you implement an AI tool, check whether it complies with GDPR rules” (August 2026). One of its versions is intended for small and medium-sized businesses that use off-the-shelf AI tools.
When does a chatbot not pay off?
Not every business needs a chatbot. Skip it when:
- you get a few inquiries a week and answer them within an hour;
- the questions are different every time and require a quote or an on-site inspection;
- nobody at your company will update the knowledge and read the conversations;
- few people visit your website: in that case, improve the website and the description of your offer first.
In these cases, a page with answers to the most common questions and a calendar in which customers pick a slot themselves will work better. You can work out how much time your team loses on repetitive answers today in the manual work cost calculator.
Not sure whether a chatbot makes sense for your business? On a free call, we will go through the questions your customers ask and tell you whether an FAQ page and a calendar are enough, or whether it is worth implementing AI. If a chatbot will not pay off, we will tell you so plainly.
Frequently asked questions about chatbot implementation
Update history
- First version. UNSEEN prices, OpenAI’s price list, Art. 50 of the AI Act and the chat rules (from the code) checked on November 13, 2026.
Sources
- Regulation (EU) 2024/1689 (AI Act), consolidated text: Art. 3(3), Art. 50(1) and (5), Art. 113 (checked on November 13, 2026)
- Regulation (EU) 2026/1744: amendment to the AI Act (checked on November 13, 2026)
- Regulation (EU) 2016/679 (GDPR), Art. 28: processor (checked on November 13, 2026)
- UODO: Before you implement an AI tool, check whether it complies with GDPR rules (checked on November 13, 2026)
- OpenAI: API pricing (checked on November 13, 2026)
- OpenAI: data sent through the API (checked on November 13, 2026)
Product names belong to their respective owners. This article is not legal advice.
Need personalized advice?
Book a free consultation. We’ll analyze your business and propose specific solutions.
Book a free 30-min call