Asking what a chatbot costs is like asking what a website costs: the word covers a two-page brochure and an online bank. The useful way to budget is to decide which kind of chatbot you need, list what it has to read and do, and count the running cost per conversation as well as the build.

Three kinds of chatbot, three budgets

  • Scripted bot: fixed menus and answers. Cheap to build, predictable, and frustrating once a question falls outside the script.
  • Answers from your content: a language model that answers from your documents, help centre or product data through retrieval (RAG). Most of the work is in the content and in proving the answers are right.
  • Agent that takes actions: the chatbot also checks orders, books appointments, raises tickets or updates records. Each action is an integration with permissions, approval rules and tests.

Where the build effort goes

  • Content and data preparation: collecting, cleaning and structuring what the chatbot answers from, and deciding who keeps it current.
  • Retrieval: splitting content, choosing how it is searched, and showing sources so an answer can be checked.
  • Integrations: your CRM, helpdesk, order system or database, each with its own authentication and error cases.
  • Conversation design: tone, hand-off to a person, what the chatbot must refuse, and how it says it does not know.
  • Evaluation: a test set from real customer questions, scored on every change, so a fix for one answer does not break ten others.
  • Channels: website widget, WhatsApp, mobile app or internal tools such as Slack — each one is its own integration.
  • Security and privacy: what personal data the chatbot may see, where conversations are stored, and protection against prompt injection.

What it costs to run

A language-model chatbot has a cost per conversation: the model provider charges for the text sent and received, and longer conversations or larger retrieved documents cost more. Add hosting, the search index, monitoring and the time someone spends reviewing conversations and updating content. We measure the cost per conversation during the build and set a ceiling, so the monthly figure is known before launch rather than discovered on the first invoice.

How to keep the cost down

  • Start with the questions that make up most of your volume, not every question anyone has ever asked.
  • Answer from content before taking actions; add actions one at a time once answers are reliable.
  • Use a smaller model where it scores as well on your evaluation set, and a larger one only where it is needed.
  • Keep a person in the loop for anything that changes money, access or a customer record.

How to get a number you can trust

Collect a few hundred real questions from your inbox, chat logs or support tickets, list the systems the chatbot would need, and decide which channels it must run on. A short discovery phase turns that into a scope, an evaluation plan, an estimate of the running cost and a fixed quote. A price given before anyone has seen your questions is the price of a demo.

Questions people ask

Is a ChatGPT-style chatbot for our website expensive?

A chatbot that answers from a small, well-kept set of content is one of the smaller AI projects. The cost grows with the amount of content, the number of systems it must act in, and how sure you need to be that every answer is right.

Do we need to train our own model?

Almost never for a chatbot. Retrieval over your own content is cheaper, easier to update and easier to check. Fine-tuning is worth it only when a model must learn a style or format that prompting cannot produce.

How long does a chatbot take to build?

It depends on the content and integrations, and the discovery phase gives you a plan with dates. A chatbot that answers from existing content is usually far quicker than one that takes actions in several systems.

How is a chatbot project priced at Vayro?

It is quoted as a fixed-scope build after discovery. Starting prices for each way of working with us are on our pricing page.
← All insights