AI Chatbots for Business Websites: Costs, Options and What to Expect

Website chatbots earned a poor reputation in the era of rigid decision trees that could not understand a simple question. Modern AI chatbots can answer in natural language from your own content, in English and French, at any hour. They also cost money to set up and run, and if they are built carelessly they can state wrong information with confidence. This guide explains the options, the real cost drivers and how to launch one you can trust.

In short: Off-the-shelf AI chatbots are usually priced per seat, per conversation or per resolved conversation, and suit businesses whose content already lives in a supported help desk. A custom chatbot built with retrieval-augmented generation (RAG) costs more up front but gives you control over data, integrations, channels such as WhatsApp and running costs. We quote custom AI chatbots individually after a free call, and they typically launch in two to three weeks; ongoing costs depend mainly on conversation volume and the model you choose.

Off-the-shelf or custom?

Off-the-shelf chatbot platforms

Many help desk and live-chat platforms now include an AI agent that reads your knowledge base and answers customers. Setup can take hours rather than weeks. Pricing models vary: a monthly subscription, a fee per agent seat, a fee per conversation, or a fee per "resolution" (a conversation the bot closes without a person). Some well-known help desk vendors charge around a dollar per resolution on top of their seat plans. Check each vendor's current pricing page and exactly how it defines a resolution, because that definition decides your bill.

Custom RAG chatbots

A custom chatbot is built on a model API from a provider such as OpenAI, Anthropic or Google Gemini, with your content indexed in a search store you control. When a visitor asks a question, the system retrieves the most relevant passages from your content and instructs the model to answer from them. You choose the model, the hosting, the integrations and the interface, and you pay the model provider directly for usage.

If your support already runs on a help desk with a good built-in AI agent, try that first. A custom build makes sense when you need actions inside your own systems, tighter control over data, specific channels or predictable costs at higher volume.

What drives the cost of a custom chatbot

Content preparation

This is the most underestimated part of any chatbot project. The bot is only as good as the content it retrieves. Expect to collect help articles, policies, product information and PDFs, remove outdated or contradictory material, and fill the gaps where common questions have no written answer. You also need a way to keep the index in sync when content changes.

Integrations and actions

Answering questions is one level. Doing things is another: looking up an order status in Shopify, booking an appointment, creating a support ticket or updating a contact in HubSpot. Each action needs API work, authentication, permission checks and testing, and anything involving a customer's account needs identity verification first.

Channels

A website widget is the usual starting point. WhatsApp is a natural fit when your customers already message you there. It runs through Meta's WhatsApp Business Platform, which involves business verification, approved message templates and its own fees. Since July 1, 2025, Meta has charged per message: free-form replies inside the 24-hour customer service window that opens when a customer messages you are free, while template messages sent outside that window are charged by category (marketing, utility, authentication) and by the recipient's country. Internal channels such as Slack or Microsoft Teams are another common option for staff-facing assistants.

Languages

If you serve customers in Québec, your chatbot should work properly in French, which means retrieving from French content, answering in the visitor's language and testing both languages separately. Our guide to bilingual websites for Québec explains the French-language requirements that apply to websites.

Guardrails and compliance

A business chatbot needs limits: topics it will not discuss, no legal, medical or financial advice unless that is your business and it is reviewed, careful handling of personal information, protection against prompt injection, rate limits and logging. This matters legally as well. In Moffatt v. Air Canada (2024), the British Columbia Civil Resolution Tribunal held the airline responsible for incorrect information its website chatbot gave a customer about bereavement fares. Treat everything your bot says as something your company said.

Analytics and administration

You will want conversation logs, reports of unanswered questions, a thumbs-up or thumbs-down on answers, and a simple way for your team to update content. These are what turn a launch into a chatbot that improves every week.

Running costs: how model usage is billed

Model providers bill by the token. A token is a small chunk of text, roughly three to four characters of English, or about three-quarters of a word. Prices are quoted per million tokens, with input (what you send the model) and output (what it writes back) priced separately. Output usually costs more per token than input.

In a RAG chatbot, the input for each reply includes much more than the visitor's question:

• The system instructions that define the bot's role and rules.
• The passages retrieved from your content.
• The recent conversation history.
• The visitor's new message.

That is why a ten-word question can still mean a few thousand input tokens. To estimate your monthly usage, multiply it out:

1. Messages: conversations per month × messages per conversation.
2. Input tokens: messages × average input tokens per message.
3. Output tokens: messages × average output tokens per message.
4. Cost: input tokens × the input rate, plus output tokens × the output rate, using your chosen model's current prices.

For example, 2,000 conversations a month with four messages each, about 3,000 input tokens and 300 output tokens per message, comes to roughly 24 million input tokens and 2.4 million output tokens. Prices per million tokens vary by orders of magnitude between small, fast models and the most capable ones, and they change often, so we price your estimate against current rates when we scope the project.

Ways to keep usage costs under control:

• Use the smallest model that meets your accuracy target, and route only difficult questions to a larger one.
• Retrieve fewer, better passages rather than many long ones.
• Trim long conversation histories.
• Use prompt caching where the provider supports it, so repeated instructions cost less.
• Set rate limits per visitor and spending alerts in the provider dashboard.

Other running costs to budget for: hosting for the app and search index, WhatsApp message fees if you use that channel, and maintenance. We offer monthly maintenance plans and can quote one alongside the build.

Accuracy and hallucination controls

A model "hallucinates" when it produces a confident answer that is not supported by facts. You cannot remove the risk entirely, but you can make it small and visible.

1. Ground every answer. Instruct the model to answer only from retrieved content, and show source links so visitors can check.
2. Allow "I don't know". When retrieval finds nothing relevant, the bot should say so and offer a person.
3. Use live data for exact facts. Prices, stock levels, opening hours and order status should come from an API call, not from text the model remembers.
4. Keep one source of truth. Outdated documents in the index are a common cause of wrong answers.
5. Build an evaluation set. Collect 100 or more real questions with correct answers and run them before every change to prompts, content or model.
6. Review conversations weekly. Flag wrong answers, fix the content or instructions and add the case to your evaluation set.
7. Limit permissions. Give the bot only the access it needs, and confirm with the visitor before any action that changes data.
8. Be transparent. Tell visitors they are talking to an AI assistant.

Human hand-off

A good chatbot knows when to step aside. Hand the conversation to a person when the visitor asks for one, when the bot's confidence is low, when the topic is sensitive (complaints, refunds, cancellations, account security), after repeated misunderstandings, and when a high-value lead is ready to talk.

The hand-off should pass the full transcript and context to your help desk, CRM or shared inbox, so the customer never has to repeat themselves. Be honest about when a person will reply, and collect contact details outside business hours. If you serve both English and French customers, route by language. Then measure the hand-off rate and what happens after it, because a sudden rise usually points to a content gap.

Chatbot launch checklist

1. Define the bot's job, audience and the questions it must handle well.
2. Collect, clean and update the content it will answer from.
3. Decide which actions it may take, and which require confirmation.
4. Write guardrails: off-limits topics, tone, languages and escalation rules.
5. Build an evaluation set of real questions with correct answers.
6. Test in English and French if you serve both.
7. Set up human hand-off with full transcripts.
8. Add an AI disclosure and update your privacy policy.
9. Configure consent for any tracking in the chat widget.
10. Set spending alerts and rate limits.
11. Launch to a small share of visitors first, then expand.
12. Review conversations weekly and keep improving the content.

Getting a quote

If you are weighing an off-the-shelf tool against a custom build, send us a few details through our contact page: your support volume, where your content lives, the channels you need and the systems the bot should connect to. We will recommend an approach, estimate running costs for your volume and send a fixed, itemized quote. You can read more about our AI chatbot development service, see how we quote projects on our pricing page, or explore our AI integration hub. If a chatbot is one of several ideas, our guide to adding AI to your business covers other good starting points, and our AI automation service handles the workflows behind the scenes.

Frequently asked questions

How much does an AI chatbot cost per month to run?

It depends on conversation volume, how much content each answer retrieves and which model you use. For a small business website, model usage is often a modest monthly amount, with hosting, any WhatsApp message fees and maintenance on top. We estimate usage for your actual volume before you commit and quote maintenance at the same time.

Can an AI chatbot give wrong answers?

Yes, which is why grounding, source links, an "I don't know" path and human hand-off matter. Testing against a set of real questions before every change and reviewing conversations each week keeps errors rare and visible. Your business is responsible for what the bot says, so treat its answers like any other customer communication.

Can the chatbot work on WhatsApp?

Yes. A custom chatbot can connect to WhatsApp through Meta's WhatsApp Business Platform, alongside your website. Meta charges per message, and free-form replies within the 24-hour window after a customer messages you are free, while template messages outside that window are charged. Business verification and template approval add some setup time.

How long does it take to build a custom AI chatbot?

Our custom AI chatbots typically take two to three weeks from kickoff to launch. Projects with several integrations, account lookups or multiple channels take longer. Content preparation is usually the biggest factor in the timeline, so gathering your documents early helps.