How to read the result
An AI chatbot has two prices, and most quotes only show one. The first is the initial build: the pipeline that splits and indexes your documents, the prompt and its guardrails, the connections to your tools, the widget or chat channel. The second is the running cost, which arrives every month: model usage, billed per token, hosting for the application and the vector database, and maintenance, because a chatbot whose knowledge base is never updated starts giving wrong answers within a few months.
The lower part of the result weighs these costs against the time the chatbot saves your team. The formula is deliberately simple: number of conversations, times the share solved without a human, times the time a team member would have spent on each request and their loaded hourly cost. The payback period then divides the build budget by that gain, net of monthly costs. If the gain does not cover running costs, the calculator says so and recommends an off-the-shelf product instead of a custom build: at low volume, that is the right answer.
The assumptions to adjust first
The share of requests solved without a human moves the result more than anything else, and it is where sales promises are least reliable. On first-line support with clean documentation, 30 to 50 percent is a realistic target after a few weeks of tuning. Going higher requires very repetitive use cases. The default of 35 percent is cautious on purpose: a project that pays back on modest assumptions is better than one that only works if everything goes right.
Time per request is easy to measure in your ticketing tool: it is the average handling time of a first-line ticket, often between 4 and 10 minutes. The loaded hourly cost of a support agent in France is around 25 to 35 EUR once employer contributions are included; adjust it to your market. For a lead qualification chatbot the real gain lies elsewhere, in prospects qualified at night and at weekends, so treat the time-saved figure as a floor.
The grid behind the calculator
Here are the durations per building block, in senior developer days. These are the figures I use for my own chatbot quotes, with proven components (TypeScript, Postgres with pgvector, Cloudflare) rather than a proprietary platform that would lock you in.
| Building block | Days |
|---|---|
| Customer support foundation (RAG over your documents, web widget) | 8 |
| Lead qualification foundation (flow, scoring, push to CRM) | 10 |
| Internal assistant foundation (RAG, per-team access rights) | 12 |
| Acting agent foundation (tool calls, guardrails, approvals) | 20 |
| Indexing 100 to 1,000 pages | 3 |
| Indexing over 1,000 pages or several sources | 8 |
| Per extra channel (WhatsApp, Teams, Slack, email) | 3 |
| Per tool connection (CRM, ticketing, ERP, calendar) | 3 |
| Handoff to a human with history | 3 |
| Conversation dashboard | 4 |
| Multilingual answers | 2 |
| Quality test suite (automated evals) | 3 |
Sensitive data or mandatory EU hosting adds 15 percent to the build (logging, encryption, access control, choosing a model provider hosted in Europe) and raises monthly hosting. The UK ICO guidance on AI and data protection and the French CNIL practical sheets are both worth reading before plugging a chatbot into customer data. The final range runs from minus 10 to plus 15 percent on days, because software projects overrun more often than they finish early.
The cost of AI, conversation by conversation
A support conversation averages five to eight exchanges. With each question, the chatbot sends the model the question, the history and the document excerpts judged relevant: a few thousand input tokens and a few hundred output tokens. With a budget model, that is roughly one to three cents per conversation; with a high-end model, four to ten cents. Current rates are listed on the pricing pages of Anthropic, OpenAI and Mistral AI, and they keep going down.
Two levers cut the bill without hurting answer quality. Prompt caching bills the repeated part of the context, such as instructions and reference documents, at a fraction of the normal price. Routing sends simple questions to a small model and keeps the high-end model for complex requests. In a well-built chatbot, AI usage stays a modest line compared with development and maintenance. Our guide to AI integration costs walks through the same calculation for other use cases.
What the calculator does not cover
It does not price the work on your side, which nonetheless drives the quality of the result: gathering and cleaning documentation, reviewing answers during testing, naming someone to update the knowledge base when your offers change. A RAG chatbot answers no better than the documents it is given, and that is the first thing I check during scoping. Our guide to RAG for businesses explains the mechanism without jargon.
Nor does it cover compliance with the EU AI Act, whose Article 50 has required since 2 August 2026 that users be told they are talking to an AI. For a support chatbot, that is a welcome sentence and a line in your privacy policy, not a project; for an agent that makes decisions about people, the analysis is heavier.
Custom build or off-the-shelf?
Off-the-shelf SaaS chatbots cost 50 to 500 EUR per month and plug in within a day. They fit when the need is standard (answering a public FAQ on a website) and volume is modest. A custom build makes sense when the chatbot must read internal documents, act in your tools, meet hosting constraints, or when a per-conversation subscription at your volume costs more than owning the tool. The full comparison is in our AI chatbot pricing guide.
What next?
If the range fits your budget, the calculator button opens the contact form with your configuration already filled in. I reply within one business day with the questions needed to turn the estimate into a firm quote, and I will tell you frankly if an off-the-shelf product would do the job. To see how we work, have a look at our custom AI chatbot service.
Frequently asked questions
How much does a custom AI chatbot cost for a small business?
A support chatbot that answers from your documents, with one tool connection and a human handoff, usually costs between 8,000 and 15,000 EUR excl. VAT to build. Then plan for 50 to 400 EUR per month of AI usage and hosting for a few thousand conversations. The calculator above refines these numbers for your case.
Is the calculator result a quote?
No. It is an indicative range based on a senior day rate of 550 to 750 EUR excl. VAT and typical durations per building block. A firm quote requires a 30-minute scoping call, mainly to assess the quality of your documents, which weighs more on the outcome than the number of features.
How much does one AI chatbot conversation cost?
Between 0.01 and 0.10 EUR in most cases, depending on the model, the length of the exchange and how much document context is sent with each question. Prompt caching, offered by Anthropic and OpenAI, cuts this cost sharply when the same context comes back often.
When is a custom chatbot not worth it?
When volume is low. Below a few hundred requests per month, the time saved does not pay back the build within two years, and an off-the-shelf SaaS chatbot at 50 to 500 EUR per month is the sensible choice. The calculator says so explicitly when that is the case.
Do users have to be told they are talking to an AI?
In the EU, yes. Article 50 of the AI Act requires informing people that they are interacting with an AI system unless it is obvious, and these transparency obligations apply since 2 August 2026. A clear message at the start of the conversation is enough in most cases.