Guide · AI & automation

How much does it cost to build an AI chatbot for your website? Build vs buy, with real model prices

The real cost to build an AI chatbot in 2026: RAG explained plainly, Chatbase and Intercom Fin pricing, per-conversation model costs, and WhatsApp chatbot fees.

By Akash Bansal, founding engineerUpdated 5 November 20267 min read

The cost to build an AI chatbot for your website splits into two numbers founders often mix up: what it costs to build or subscribe, and what each conversation costs to run. The second number has dropped so far that it rarely decides anything. The first one does.

I build AI systems for a living, mostly pipelines where a language model's answer has to be checked before anyone trusts it. On NextMantra, several models score each candidate against the role's rubric. On Brahmo, Gemini reads handwritten exam answers, a matching step finds the relevant parts of the mark scheme, Claude scores them on a weighted rubric, and a human expert can override. A website chatbot uses the same core move: find the right source text, then make the model answer from it.

How does an AI chatbot answer questions about my business?

A modern website chatbot uses a pattern called RAG, retrieval-augmented generation. In plain terms, it is an open-book exam. The bot doesn't memorise your business. Every time someone asks a question, it looks up the most relevant passages from your content, hands them to the language model, and tells it to answer using only those passages.

1. Prepare your content. Your help pages, pricing, policies and FAQs are split into short chunks.
2. Index it. Each chunk is turned into an embedding, a list of numbers that captures its meaning, and stored in a searchable index.
3. Retrieve. A visitor's question is turned into an embedding too, and the index returns the chunks closest in meaning.
4. Answer. The model gets the question plus those chunks and writes a reply, ideally with a link to the source.
5. Hand over. If nothing relevant is found, the bot says so and offers a person.

Two practical consequences. First, the bot is only as good as your written content: if your refund policy lives in one support agent's head, the bot can't know it. Second, updating the bot means updating the documents, not retraining anything.

How much does an AI chatbot cost to buy?

Off-the-shelf chatbots charge either a monthly fee with a message allowance or a fee per resolved conversation. As of November 2026, from the vendors' own pricing pages:

Per-resolution pricing is easy to reason about. At 1,000 resolved conversations a month, Fin costs about $990 on top of seats. Whether that is cheap depends on what a human-handled ticket costs you.

How much does it cost to build a custom AI chatbot?

A custom build has a one-time cost and small running costs. My starting price for an AI feature is $8,000, on a fixed quote after scoping. For that you'd typically get the retrieval pipeline over your content, the chat widget, handover to email or your helpdesk, an admin view of conversations, and an evaluation set: a list of real questions with the right answers, so every change can be tested before it goes live.

The price goes up with the integrations, not the chat window. A bot that answers from documents is the cheap end. A bot that looks up orders, changes bookings or issues refunds needs authentication, permissions and careful testing of each action. The app cost calculator prices those integrations one by one, so you can see the rough range for your own bot.

What does each conversation cost to run?

Model costs are now small enough to calculate on the back of an envelope. Providers charge per million tokens (a token is roughly three-quarters of a word). Prices from the providers' own pages, November 2026:

Assumptions for the estimate: six back-and-forth turns, each sending about 4,000 tokens (retrieved passages plus the conversation so far) and getting about 300 back, so roughly 24,000 tokens in and 1,800 out per conversation. Prompt caching can cut the input cost further. At 2,000 conversations a month, that is about $7 on the small model and about $130 on the larger one, plus hosting in the tens of dollars.

The lesson: model choice is a quality decision, not a budget one, at small-business volumes. Start with the smaller model, measure answer quality on your evaluation set, and only move up if it fails.

Build vs buy: which should you choose?

Not for anyone: a bot launched over thin or outdated content. It will answer confidently from whatever it has, and that is worse than no bot.

Is there a middle option between buying and building?

Yes, and it is often the right one. Start on a hosted product, and connect it to your systems through the custom actions or webhooks many of them offer, so the bot can look up an order status without a full custom build. You keep the vendor's widget, hosting and analytics, and pay an engineer only for the integration. The trade-off is that you depend on the vendor's limits: what an action can do, how answers are evaluated, and what happens to your data if you leave. If those limits start shaping your product, that is the signal to move to a build, and by then you'll have months of real questions to design it from.

How much does a WhatsApp chatbot cost?

A WhatsApp chatbot costs about the same to build as a website bot, plus Meta's messaging fees. Since July 2025, the WhatsApp Business Platform charges per delivered message, by category (marketing, utility, authentication, service) and by the recipient's country (Meta, WhatsApp pricing).

• When a customer messages you, a 24-hour customer service window opens, and your bot can reply freely inside it.
• Those replies are service messages. The first 1,000 a month per business phone number are free.
• From 1 October 2026, service messages beyond that free tier are charged, and so are utility messages sent inside the window.
• Messages you start outside the window must use pre-approved templates, priced by category.

For a support bot that only replies to customers, the fees stay small. For one that sends reminders and promotions, model the template costs by country before you commit.

What goes wrong with website chatbots

• Invented answers. In 2024, a Canadian tribunal held Air Canada responsible after its chatbot told a passenger he could claim a bereavement fare after travelling, which the airline's policy didn't allow (CBS News). Restrict the bot to retrieved content, and make "I don't know, here's a person" an acceptable answer.
• Stale content. Prices change on the site but not in the index. Re-index automatically whenever pages change.
• No evaluation set. Someone tweaks the prompt and quietly breaks refund answers. Keep 50 to 100 real questions with correct answers and test every change against them.
• Handover loses context. The customer explains everything to the bot, then again to a human. Pass the transcript along.
• Nobody reads the logs. The unanswered questions are your best content roadmap. Review them weekly.

A checklist before you spend anything

1. Pull last month's support questions. Count how many your existing docs already answer.
2. Fill the gaps in the docs first. That helps human agents too.
3. Trial a hosted bot on those docs for two weeks and read every transcript.
4. Write down what it couldn't do. Those are the build requirements, if any.
5. If you build, own the content index, the prompts, the evaluation set and the model API account.

If you've hit the limit of the hosted products (the bot needs to act in your systems, or it is the product), that is the kind of AI build I do. Here's how I work and what it costs. For the related question of whether you need an agent or a simpler workflow, see hiring an AI agent developer; for phone calls, see AI receptionists, build vs buy.