AI chatbot
An AI chatbot for your business that answers from your own documents
An assistant on your website or WhatsApp that answers customers from your prices, opening hours and policies, says when it does not know, and passes the conversation to a person.
What a useful business chatbot does
We build AI chatbots for businesses that answer customer questions from your own documents: price lists, opening hours, delivery terms, returns policies, product sheets. A useful chatbot says when it does not know, passes the conversation to a person and records the details of interested customers, with their consent. We start with a feasibility study at €1,200–2,500, and a production chatbot starts at €4,000, plus the monthly cost of using the model.
What a chatbot that actually helps does
- It answers from your sources. Your documents are split into passages and indexed. For each question the system finds the relevant passages and the model writes the answer from them only, citing the source. The model is not trained on your data: when a price changes, you update the document.
- It calls a person when the question goes beyond the documents, when the customer is unhappy, or when they ask for one. That person gets the full history by email, in the CRM or in the support tool.
- It collects enquiries: what the customer is looking for and how to reach them, with a link to the privacy policy. For appointment-based services it sends them straight to the calendar, as described in our article on online booking with Google Calendar.
- It works on WhatsApp too. The WhatsApp Business Platform terms prohibit using the platform to offer a general-purpose AI assistant as the main product, and allow AI that is part of a business's own service, such as answers about orders or bookings. Meta may charge for messages separately.
What it should not do
A business chatbot does not promise discounts, approve refunds, give medical or legal advice, or make up answers the documents do not contain. Language models can state false things convincingly, so the bot is instructed to decline politely and bring in a person. We do not ask for health data, national ID numbers or card details in the chat. On a clinic's site, for example, the bot answers about hours and services and forwards the booking request without discussing symptoms.
The AI Act and the GDPR
Since 2 August 2026, Article 50 of the AI Act applies: an AI system that talks directly with people must tell them clearly that they are dealing with an AI, at the latest at the first interaction, unless that is obvious. The European Commission says the exception is to be read narrowly, so we put the notice in the bot's very first message. The grace period to 2 December 2026 covers only the marking of generated content. More in our AI Act guide for small businesses.
Chat conversations are personal data under the GDPR, so you need a clear privacy notice, a processing agreement with the model provider (Article 28) and a defined retention period. OpenAI and Anthropic say they do not train on API or business-plan data by default; other providers have other rules, compared in our guide to AI chatbot privacy settings.
Usage costs and evaluation
Besides the build, you pay the model provider for each conversation, based on the length of the questions, the passages sent and the answers. Simple questions can run on a small, cheap model and hard ones on a more capable model. During feasibility we estimate the cost per conversation using your real questions.
At the same stage we collect 100–300 real questions from emails, messages and calls, with the correct answer for each. The bot runs through all of them and we count correct answers, errors and justified refusals, including trick questions: discount requests, attempts to make it ignore its instructions, topics unrelated to your business. We launch when the result reaches the threshold we agreed together.
What the chatbot includes
A chatbot that knows its limits
Useful answers, a clear route to a person, and the legal notices built in.
Answers from your documents
Prices, hours, delivery, returns. Every answer comes from a passage you can check, and you keep the documents up to date yourself.
Hand-off to a person
When it does not know, when the customer is unhappy or asks for a human, the conversation reaches you with its full history.
Enquiries and contact details
The bot asks what the customer needs and how to reach them, with consent and a privacy link, then sends the enquiry to your CRM or inbox.
On your site and on WhatsApp
The same bot and the same sources, in a chat window on your site or on the WhatsApp Business Platform, within Meta's terms.
AI notice and GDPR
A clear AI notice from the first message, a privacy notice, limited retention of conversations and providers that do not train on your data.
Tested before launch
A set of real questions with correct answers, run before launch and again after every significant change to documents or model.
How it works
How a chatbot project runs
- 1
Feasibility
We read your documents, collect real customer questions and check whether a chatbot is the right tool. Sometimes a well-written FAQ page solves more.
- 2
Prototype and evaluation set
We build a prototype on your documents and measure it against the question set. Together we agree the quality threshold for launch.
- 3
Integration
We put the bot on your site or WhatsApp, connect the hand-off to your email, CRM or calendar, and set up the AI notice, logs and retention period.
- 4
Gradual launch and monitoring
We start with part of your visitors, read conversations weekly, fill gaps in the documents where the bot hesitated and run the evaluation again.
Budget
AI chatbot prices
Bands are indicative and negotiable, and the written quote follows the feasibility results. Model usage and any WhatsApp message charges are paid to those providers.
Feasibility and evaluation set
€1,200–2,500one-off
Find out whether a chatbot is worth it, how well it answers your real questions and what it would cost per month.
- Review of documents and real questions
- Evaluation set of 100–300 questions
- Prototype measured on your data
- Estimate of monthly usage cost
Production chatbot
from €4,000one-off
A chatbot live on your site, connected to your staff and your tools, with the legal notices in place.
- Answers from documents, with sources
- Hand-off to a person and enquiry capture
- Website, and WhatsApp on request
- AI notice, GDPR, logs and limited retention
- Admin for updating the documents
Follow-up workflow
€400–1,200per workflow
One automation that acts on what the chatbot collects.
- Enquiries sent to your CRM or a spreadsheet
- Weekly summary of questions by email
- Alerts for unanswered topics
Our AI work
We do not have a client chatbot in the portfolio yet, so we show the AI projects our experience comes from. My Claude Code is a local LLM gateway that routes requests across 57 model providers, with fallback chains and the cost of every request logged, the same kind of control a chatbot needs to stay within budget. AI Translator Pro and AI Transcription Studio are in development at Danube Labs and can run on local models, for cases where data must not leave the machine. Our other projects are listed in the portfolio.
FAQ
Questions, answered
How much does an AI chatbot for a website cost?
A feasibility study with an evaluation set costs €1,200–2,500 and tells you whether the chatbot is worth building. A production chatbot starts at €4,000, depending on the number of documents, channels (website, WhatsApp) and integrations. Follow-up automations, such as sending enquiries to your CRM, cost €400–1,200 each. Model usage is a separate monthly cost paid to the provider.
What does a chatbot cost to run each month?
It depends on the number of conversations, their length, how many document passages are sent with each question and the model chosen. We measure the cost per conversation during feasibility, with your real questions, and give you an estimate before launch. We also build in limits: maximum conversation length, messages per visitor and an alert when spending passes the budget you set.
Can an AI chatbot give wrong answers?
Yes, any language model can. We reduce the risk by making it answer only from your documents, telling it to decline and call a person when the documents are silent, testing it on a set of real questions before launch, and reading the logs in the first weeks. Anything with money or legal weight, such as a final price or a refund, is confirmed by a person.
Do I have to tell customers they are talking to an AI?
Yes. Article 50 of the AI Act, in force since 2 August 2026, requires people to be told clearly that they are interacting with an AI system, at the latest at the first interaction, unless it is obvious. The Commission reads that exception narrowly. Fines for transparency breaches reach €15 million or 3% of turnover, and for SMEs the lower of the two applies.
Is an AI chatbot GDPR compliant?
It can be, if it is built that way. You need a clear privacy notice in the chat, a legal basis for each use of the data, a processing agreement with the model provider, a fixed retention period and no sensitive data collected. We set up the technical side, including logs and deletion, and you have the final privacy text checked by your lawyer or data protection officer.
Can I have an AI chatbot on WhatsApp?
Yes, for your own business use. Meta's WhatsApp Business Platform terms bar general-purpose AI assistants offered as the main product, while a bot that answers your customers about orders, bookings or services is allowed. You need a WhatsApp Business Platform account, and Meta may charge for certain messages. We connect the same documents and hand-off rules as on your website.
Which documents do I need to prepare?
Price lists, existing FAQs, terms and conditions, delivery and returns policies, product or service sheets, and any internal notes your staff use to answer customers. Even more useful are 100 or more real questions from emails, chats and calls. Up-to-date documents matter more than many documents, because the bot is only as accurate as its sources.
What happens after the chatbot goes live?
In the first weeks we read conversations, fill gaps in the documents and rerun the evaluation. Later, you update documents from the admin panel. Providers retire older models over time, so before switching to a new one we run the evaluation set again to catch changes in behaviour. Further work is done on request, with an estimate first.
Find out if a chatbot is worth it for you
In a free call we look at the questions you receive and tell you honestly whether a chatbot can take them on.